Page 1
Transformer‐Based Digital Soil Mapping Reveals Land‐Use
Legacy Effects on (Sub)Tropical Soil Carbon
Hayley Peter‐Contesse
1,2
, Aron Boettcher
3
, Susan E. Crow
1,4
, and Sabine Grunwald
4
1
Department of Natural Resources and Environmental Management, University of Hawaiʻi at Mānoa, Honolulu, HI, USA,
2
Now at Department of Forest Ecosystems and Society, Oregon State University, Corvallis, OR, USA,
3
Department of
Geography and Environment, University of Hawaiʻi at Mānoa, Honolulu, HI, USA,
4
Soil, Water, and Ecosystem Sciences
Department, University of Florida, Gainesville, FL, USA
Abstract
Tropical and subtropical soils store approximately one‐third of global soil organic carbon (SOC)
despite covering less than one‐quarter of terrestrial surfaces, yet accurate spatiotemporal prediction methods
remain inadequate. We present the first application of transformer architectures to multi‐temporal Landsat
observations for SOC prediction, demonstrating that temporal attention mechanisms can extract management
and phenological signals that static modeling approaches cannot capture. Using 556 soil observations across the
Hawaiian Islands and 36‐month sequences of Landsat imagery integrated with climate and terrain data, the
model predicted SOC with
R
2
=
0.57 (RPIQ
=
1.88) in leave‐one‐island‐out cross‐validation and
R
2
=
0.60
(RPIQ
=
2.00) in 10‐fold cross‐validation. We used a space‐for‐time substitution, applying the learned SOC‐
environment relationships to time‐varying covariates to infer temporal SOC patterns across
∼
360 km
2
of
agricultural landscapes on Maui. The model inferred widespread SOC decline from 2017 to 2024, with former
sugarcane lands declining the most (
−
0.96
±
0.49 percentage points in 0–20 cm soil depth). Climate
significantly modulated inferred SOC decline in abandoned pineapple lands: drier zones lost twice the carbon of
wetter zones, an effect absent where pineapple was converted to managed rangeland. Our transformer‐based
approach provides a scalable, transferable framework for inferring carbon dynamics in data‐sparse agricultural
landscapes with complex environmental gradients and land use histories.
Plain Language Summary
Tropical soils store enormous amounts of carbon, but scientists lack
effective tools to track how this carbon changes over time. We developed a new artificial intelligence approach
that analyzes 10 years of satellite images to predict soil carbon across agricultural land on Maui, Hawaiʻi. Unlike
previous methods that use single snapshots in time, our model learns from sequences of images, allowing it to
detect patterns related to seasonal crop growth and land management. We tested the model using 556 soil
samples from across the Hawaiian Islands and found it accurately predicted soil carbon. Because repeated
measurements over time are unavailable, we used the model to infer how carbon likely changed by tracking its
response to shifting environmental conditions. Applied to former plantation lands on Maui, it inferred
widespread carbon losses between 2017 and 2024, with the steepest declines on former sugarcane fields.
Climate mattered on abandoned pineapple lands, where drier sites lost roughly twice as much carbon as wetter
ones, but this effect disappeared where those lands were converted to managed pasture, suggesting active
management can buffer carbon against climate stress. This satellite‐based approach offers a practical tool for
tracking carbon across large agricultural regions where traditional sampling is impractical.
1. Introduction
Tropical and subtropical soils represent a significant global terrestrial carbon reservoir, storing up to 71% of total
ecosystem carbon in Hawaiian systems (Selmants et al.,
2017
) and accounting for approximately one‐third of
global soil organic carbon (SOC) despite covering less than one quarter of terrestrial surfaces (Hiederer &
Köchy,
2011
; Peel et al.,
2007
; Scharlemann et al.,
2014
). These stocks are dynamic: organic carbon (C) fractions
in surface soils tend to have much shorter turnover times than SOC deeper in the profile, responding more quickly
to changes in management practices and playing a crucial role in active C cycling (Jat et al.,
2020
; Trum-
bore,
2000
). This sensitivity to land management makes accurate spatiotemporal SOC characterization essential
for stewardship and policy implementation. In Hawaiʻi, where ambitious climate targets mandate net‐negative
emissions by 2045 (Act 15, 2018) and 50% reductions below 2005 levels by 2030 (Act 238, 2022), such
RESEARCH ARTICLE
10.1029/2026EF008563
Special Collection:
Advances in Understanding Land
System Change
Key Points:
•
A transformer using satellite time‐
series predicted Hawaiian soil carbon
(
R
2
=
0.60), outperforming six con-
ventional ML models
•
Model‐inferred soil carbon trajectories
show former sugarcane lands declined
by
∼
0.96 percentage points (0–20 cm)
from 2017 to 2024
•
Inferred soil carbon loss on abandoned
pineapple lands doubled at drier sites;
conversion to rangeland removed this
climate effect
Supporting Information:
Supporting Information may be found in
the online version of this article.
Correspondence to:
H. Peter‐Contesse,
hayley.peter-contesse@oregonstate.edu
Citation:
Peter‐Contesse, H., Boettcher, A., Crow, S.
E., & Grunwald, S. (2026).
Transformer‐based digital soil mapping
reveals land‐use legacy effects on
(sub)tropical soil carbon.
Earth's Future
,
14
, e2026EF008563.
https://doi.org/10.
1029/2026EF008563
Received 17 MAR 2026
Accepted 1 AUG 2026
Author Contributions:
Conceptualization:
Hayley Peter‐
Contesse, Aron Boettcher
Funding acquisition:
Susan E. Crow
Investigation:
Hayley Peter‐Contesse,
Aron Boettcher
Methodology:
Hayley Peter‐Contesse,
Aron Boettcher, Susan E. Crow,
Sabine Grunwald
Project administration:
Hayley Peter‐
Contesse, Susan E. Crow
© 2026 The Author(s). Earth's Future
published by Wiley Periodicals LLC on
behalf of American Geophysical Union.
This is an open access article under the
terms of the
Creative Commons
Attribution
License, which permits use,
distribution and reproduction in any
medium, provided the original work is
properly cited.
PETER‐CONTESSE ET AL.
1 of 18
Page 2
monitoring is particularly urgent as the latest statewide C inventory (2022) reported increasing rather than
declining emissions (State of Hawaiʻi Department of Health, Clean Air Branch,
2018
,
2022
,
2025
).
However, current modeling and mapping tools remain inadequate for capturing Hawaiian soil carbon dynamics.
Existing gridded products like SoilGrids and OpenLandMap (Consoli et al.,
2025
; Poggio et al.,
2021
), while
recently improved to 30 m resolution, provide static snapshots poorly calibrated to Hawaiʻi's biophysical
complexity and land‐use legacies. Process‐based models like DayCent (Parton et al.,
1998
) and RothC (Coleman
& Jenkinson,
1996
) are parameterized for temperate mineral soils and fail to represent the dominant C stabili-
zation mechanisms in volcanic soils—particularly mineral‐organic associations mediated by short‐range‐order
minerals and the development of pseudosand aggregates that obscure true soil texture (Basile‐Doelsch
et al.,
2007
; Kramer & Chadwick,
2018
). More fundamentally, treating SOC as temporally static ignores the
dynamic nature of soil C cycling in response to land management and climate variability. Models capable of
capturing spatiotemporal complexity across Hawaiʻi's steep environmental and land use gradients are thus
essential for accurate C monitoring and informed land management.
Transformer architectures have revolutionized sequence modeling across multiple domains, from natural lan-
guage processing to computer vision, by capturing long‐range dependencies and complex temporal patterns
(Vaswani et al.,
2017
). Despite their demonstrated power in handling sequential data, transformers remain largely
unexplored for soil property prediction using multi‐temporal satellite observations. This gap represents both a
methodological opportunity and a practical limitation. Traditional machine learning approaches to soil mapping
typically rely on static environmental covariates or long‐term climate averages (Grunwald,
2022
), effectively
discarding the rich temporal dynamics captured in satellite time series. This limitation is particularly constraining
for SOC prediction, where vegetation phenology, management practices, and seasonal soil exposure create
temporal signatures that contain crucial information about belowground C stocks. However, emerging multi‐
sensor soil profiling technologies can now collect high‐resolution (mm‐scale) vertical soil data in situ at
depths up to 120 cm in approximately 60 s (Grunwald et al.,
2024
), demonstrating the technical feasibility of
capturing temporal soil dynamics at scales that match satellite‐based environmental monitoring.
While recent studies have applied transformers to soil spectral data (Cao et al.,
2024
; Jin et al.,
2023
) or combined
static remote sensing with temporal climate variables (Kakhani et al.,
2024
), these approaches do not system-
atically exploit the temporal dynamics within satellite imagery itself. Here we advance transformer applications to
soil C prediction by leveraging multi‐temporal Landsat observations to capture vegetation phenology, manage-
ment practices, and seasonal soil exposure—temporal above‐ground signatures that contain crucial information
about belowground C stocks. Our approach represents a fundamental shift from treating satellite data as static
spatial layers to exploiting the full temporal dimensionality of earth observation data for soil inference.
Hawaiian agricultural landscapes present both opportunities and challenges for developing and validating
transformer‐based SOC prediction, including extreme environmental gradients, diverse land use transitions, and
highly variable soil C. The shift from industrial sugarcane and pineapple cultivation to diversified agriculture,
rangeland, and abandonment has created a mosaic of management histories. Soils across the main islands of the
archipelago range from young, weakly‐developed volcanic materials to highly weathered Oxisols and volcanic
ash‐derived Andisols (Soil Survey Staff,
2022
), with the latter exhibiting exceptional C stabilization capacity
through short‐range‐order (SRO) mineral‐organic associations (Basile‐Doelsch et al.,
2007
; Kramer & Chad-
wick,
2018
). This soil diversity contributes to SOC concentrations that vary substantially across short distances
(Hodges et al.,
2024
). Agricultural disturbance can disrupt SOC stabilization mechanisms with effects that persist
for centuries: on the Kohala Peninsula (Hawaiʻi Island), soils cultivated by early Hawaiians showed 26% lower C
stocks than adjacent reference sites more than 200 years after agricultural cessation, with the largest differences
occurring in high‐SRO mineral soils where C retention capacity appeared to remain compromised (Cusack
et al.,
2013
). These persistent legacy effects interacting with complex climate‐soil‐management gradients pose
challenges for SOC modeling, as traditional modeling frameworks often rely on equilibrium assumptions
(Rumpel et al.,
2020
; Smith et al.,
2020
) that may not capture centuries‐long persistence of disturbance impacts.
To address these challenges, we developed a transformer‐based machine learning approach that integrates multi‐
temporal satellite observations with environmental covariates for spatiotemporal SOC prediction. Because
repeated SOC measurements through time are largely unavailable in Hawaiʻi, as in most regions, and funda-
mentally, it is not possible to destructively sample the same soil twice, we adopt space‐for‐time substitution (Blois
et al.,
2013
; Pickett,
1989
). Spatial variation in SOC across sites sampled at different times is used to infer how
Supervision:
Susan E. Crow,
Sabine Grunwald
Visualization:
Hayley Peter‐Contesse,
Aron Boettcher
Writing – original draft:
Hayley Peter‐
Contesse, Aron Boettcher
Writing – review & editing:
Hayley Peter‐Contesse, Aron Boettcher,
Susan E. Crow, Sabine Grunwald
Earth's Future
10.1029/2026EF008563
PETER‐CONTESSE ET AL.
2 of 18
23284277, 2026, 9, Downloaded from https://agupubs.onlinelibrary.wiley.com/doi/10.1029/2026EF008563 by Oregon State University, Wiley Online Library on [19/10/2026]. See the Terms and Conditions (https://onlinelibrary.wiley.com/terms-and-conditions) on Wiley Online Library for rules of use; OA articles are governed by the applicable Creative Commons License
Page 3
SOC responds to changing environmental conditions. The model learns SOC‐environment relationships from
one‐time observations distributed across space and applies them to time‐varying covariate sequences at each
location, yielding model‐inferred SOC trajectories rather than directly observed change. This framework has been
used previously for machine‐learning‐based SOC change mapping (Heuvelink et al.,
2021
), and we extend it here
with a transformer architecture capable of extracting phenological and management signals from satellite time
series.
Following cross‐validation across the Hawaiian Islands, we apply our model to a study region on the N‐NW
slopes of Haleakalā Volcano on Maui, Hawaiʻi. The region encompasses diverse agroecosystems with com-
plex management histories, including transitions from sugarcane and pineapple monocultures to rangeland,
diversified agriculture, and abandonment. These land use transitions create distinct temporal signatures in satellite
observations that the model leverages for SOC inference. Our objectives were to (a) develop and validate a
transformer‐based model leveraging multi‐temporal satellite imagery for SOC prediction across complex agri-
cultural landscapes, and (b) infer spatial and temporal surface SOC patterns across land use transitions and
environmental gradients.
2. Methods
2.1. Study Region
We selected a
∼
360 km
2
region on the north‐ and northwest‐facing windward slopes of Haleakalā Volcano on
Maui, Hawaiʻi (Figure
1
). This area was chosen for two primary reasons: (a) strong gradients in climate and land
use history providing diverse soil‐forming conditions, and (b) high spatial density and temporal coverage of soil
observations relative to other areas across the Hawaiian Islands (spatial density:
∼
1 sample per 0.94 km
2
;
temporal coverage: 1963–2025).
The study area extends from near sea level to approximately 1,800 m elevation on Haleakalā, a 3,055 m volcano
that creates steep orographic climate gradients. Mean annual precipitation increases sharply from west to east and
with elevation, ranging from
∼
370 mm at the coast to 5,900 mm at higher elevations within 25 km (Longman
et al.,
2024
). Mean annual temperature decreases from 23.7°C at low elevations (5 m) to 10.3°C at high elevations
(2,100 m), with corresponding daily temperature ranges of 19.9–28.3°C and 5.4–15.7°C, respectively (Longman
et al.,
2024
).
Parent material is predominantly 1.1‐million‐year‐old Kula volcanics formed from Honomanu Basalt flows
deposited near the end of the shield‐building stage, with volcanic ash deposition influencing soil development at
higher elevations (Soil Survey Staff,
2025
). The precipitation gradient and volcanic parent materials produce
diverse soil types ranging from drier Ustands and Torrox near the coast to wetter andic soils at high elevations.
The study area has undergone substantial land use change over the past two centuries. Historical land uses include
indigenous agricultural systems, ranching and pasture development, and extensive sugarcane and pineapple
monocropping during the plantation era. Current land uses include native and introduced forest, actively managed
rangeland, diversified agriculture, and abandoned former plantation lands undergoing secondary succession
(Bremer et al.,
2021
; Perroy et al.,
2016
).
2.2. Data Sources and Data Preparation
Data sources used in transformer model development and training are included in Table
1
.
Ground Reference Data
: SOC (%) measurements were aggregated and harmonized from four sources: Forest
Inventory and Analysis (FIA; U.S. Forest Service,
2024
), Hawaiʻi Soil Carbon Database (Crow et al.,
2021
), The
Hawaiʻi Soil Health Database (
2025
), and National Cooperative Soil Survey (NCSS,
n.d.
) (total
n
=
556 samples)
and included the 0–20 cm SOC concentration and sampling location coordinates. For a small subset of the data
(
<
5%), SOC concentrations measured at the 0–15 cm depth were standardized to the 0–20 cm depth using a depth‐
weighted approach. We assumed organic C declined by 15% between 15 and 20 cm depth (C
15‐20
=
C
0‐15
×
0.85),
and calculated the 0–20 cm concentration as: C
0‐20
=
(15
×
C
0‐15
+
5
×
C
15‐20
)/20. This decay rate is conservative
relative to observed soil C depth profiles in surface horizons across the study area in our SOC stock data sets.
Reference SOC concentrations across the main Hawaiian Islands (
n
=
556, 0–20 cm depth) ranged from 0.12% to
39.50% (mean
=
8.39
±
6.38%, median
=
7.51%, IQR
=
3.45%–11.5%).
Earth's Future
10.1029/2026EF008563
PETER‐CONTESSE ET AL.
3 of 18
23284277, 2026, 9, Downloaded from https://agupubs.onlinelibrary.wiley.com/doi/10.1029/2026EF008563 by Oregon State University, Wiley Online Library on [19/10/2026]. See the Terms and Conditions (https://onlinelibrary.wiley.com/terms-and-conditions) on Wiley Online Library for rules of use; OA articles are governed by the applicable Creative Commons License
Page 4
Figure 1.
Map of study area on Haleakalā’s N‐NW flank on the island of Maui, including the diversity of mapped soil orders (Soil Survey Staff,
2025
) across the region.
Inset map shows the main chain of Hawaiian islands.
Table 1
Data Sources Used in Transformer Model Development and Training
Data source
Resolution
Variables (n)
Variables
a
Landsat 7/8/9
30 m
6
Bands: Blue, Green, Red, NIR, SWIR‐1, SWIR‐2
TerraClimate
4 km, monthly
14
Reference evapotranspiration (mm), Climate water deficit (mm), PDSI, Actual
evapotranspiration (mm), Precipitation accumulation (mm), Runoff (mm), Soil moisture
(mm), Downward surface shortwave radiation (W m
−
2
), Snow water equivalent (mm),
Minimum temperature (°C), Maximum temperature (°C), Vapor pressure (kPA), Vapor
pressure deficit (kPA), Windspeed at 10 m (m s
−
1
)
Terrain (SRTM and 3DEP)
b
10–30 m
5
Elevation, Aspect (sin/cos), Slope, Topographic Position Index
Temporal
–
2
Day of year (sin and cos cyclical encoding)
Prediction target
c
–
1
Soil organic carbon (0–20 cm depth, %)
a
NIR is near infrared, SWIR is short‐wave infrared, and PDSI is Palmer Drought Severity Index.
b
SRTM is NASA's Shuttle Radar Topography Mission and 3DEP is
U.S. Geological Survey 3D Elevation Program (NASA Jet Propulsion Laboratory (JPL),
2013
; U.S. Geological Survey, 3D Elevation Program (3DEP),
n.d.
).
c
SOC is
the predicted variable and is not provided as a model input at any stage.
Earth's Future
10.1029/2026EF008563
PETER‐CONTESSE ET AL.
4 of 18
23284277, 2026, 9, Downloaded from https://agupubs.onlinelibrary.wiley.com/doi/10.1029/2026EF008563 by Oregon State University, Wiley Online Library on [19/10/2026]. See the Terms and Conditions (https://onlinelibrary.wiley.com/terms-and-conditions) on Wiley Online Library for rules of use; OA articles are governed by the applicable Creative Commons License
Page 5
Landsat Time Series
: Landsat 7–9 surface reflectance imagery (2014–2024; Earth Resources Observation and
Science (EROS),
2020
) provided the spectral foundation for temporal feature extraction. Bands red, green, blue,
near infrared, and SWIR‐1 and SWIR‐2 were used. Because Landsat observations are irregular due to cloud cover
and orbital patterns, time series were smoothed using LOESS (locally estimated scatterplot smoothing) and
resampled to 30‐day intervals to align with TerraClimate data and create consistent temporal sequences.
Climate Variables
: TerraClimate monthly data (Abatzoglou et al.,
2018
) provide comprehensive climate coverage
at 4 km resolution. Primary variables include maximum and minimum temperature, vapor pressure, precipitation
accumulation, downward surface shortwave radiation, and wind speed. Derived variables include reference
evapotranspiration, actual evapotranspiration, climate water deficit, soil moisture, Palmer Drought Severity Index
(PDSI), and vapor pressure.
Topographic Variables
: Elevation, slope, and aspect were derived from the USGS 3D Elevation Program (3DEP)
10 m product. Multi‐scale topographic position index (mTPI) from the Shuttle Radar Topography Mission
(SRTM) characterized landscape position and topographic context at 30 m resolution (U.S. Geological Survey, n.
d.; NASA Jet Propulsion Laboratory (JPL),
2013
).
For each soil sampling location, temporal sequences of spectral data and climate variables were extracted along
with static topographic attributes. The transformer model processed these aligned temporal sequences to predict
surface SOC concentration. Temporal sequences were extracted for the span of 3 years prior to the sampling date.
Landsat observations were cloud‐filtered, then smoothed using LOESS and resampled to 30‐day intervals to
create regular time series aligned with TerraClimate monthly data. The 6‐month LOESS span was chosen to
balance suppression of high‐frequency noise from residual cloud contamination, sensor differences across
Landsat 7/8/9, and irregular revisit gaps against retention of the seasonal and management signals the model relies
on. Features were normalized using mean and standard deviation to handle outliers common in environmental
data. Cloud‐contaminated observations in Landsat time series were removed, and the resulting gaps were
interpolated using LOESS smoothing to produce evenly‐spaced 30‐day observations.
2.3. Model Architecture
The model architecture follows Vaswani et al. (
2017
), with layer count and dimensions reduced to suit our data set
size (Table S1 in Supporting Information
S1
The model employs a sliding window approach in which a sequence
of 36 consecutive monthly timesteps (t
0
‐t
35
) serves as input to predict SOC concentration at the following
timestep (t
36
). Each timestep comprises 27 features including 22 temporal variables and 5 static terrain features
(Table
1
)). Day‐of‐year is encoded as the sin and cos of day position within the year to provide a cyclical temporal
reference within each sequence. At the predicted timestep, the model outputs the full set of covariates including
SOC as the prediction target. However, SOC is never provided as a model input at any timestep. A stride of one
timestep generates overlapping windows from each location's timeseries (e.g., t
1
to t
36
predicting t
37
), and suc-
cessive windows are processed independently, so the model does not operate autoregressively. All input features
were normalized using Z‐score standardization based on training‐set statistics (Figure
2
).
2.4. Model Training and Application
The model was trained using a location‐based random split to ensure spatial independence between training and
testing data sets. Of the 556 soil sampling locations, 500 (90%) were allocated to the training set and 56 (10%) to
the test set. The sliding window method applied to these locations generated thousands of training sequences,
providing the model with diverse examples of the relationship between temporal environmental sequences and
SOC across Hawaiʻi's varied land uses, climate gradients, and soil types.
Training was conducted using mini‐batch gradient descent with a batch size of 1,024 sequences. We employed
mean squared error as the loss function and optimized model parameters using the Adam optimizer with a learning
rate of 1
×
10
−
4
. The model was trained for a fixed 40 epochs. The relatively small learning rate and extended
training period allow the model to learn subtle relationships between environmental sequences and SOC without
overshooting optimal parameter values. Training and test loss curves were inspected to confirm convergence
without overfitting. Additional hyperparameters are listed in Table S1 in Supporting Information
S1
.
Following training, the model was applied to generate monthly SOC predictions for each location across the
2017–2024 study period. Generating a prediction for the first timestep (2017) required a preceding 36‐month
Earth's Future
10.1029/2026EF008563
PETER‐CONTESSE ET AL.
5 of 18
23284277, 2026, 9, Downloaded from https://agupubs.onlinelibrary.wiley.com/doi/10.1029/2026EF008563 by Oregon State University, Wiley Online Library on [19/10/2026]. See the Terms and Conditions (https://onlinelibrary.wiley.com/terms-and-conditions) on Wiley Online Library for rules of use; OA articles are governed by the applicable Creative Commons License
Page 6
input window (2014–2016). For each location, the model processed overlapping 36‐month input windows and
generated sequential predictions, producing a continuous, model‐inferred SOC series across the study period.
2.5. Model Validation
We used two complementary cross‐validation strategies, both partitioned by location, to assess spatial predictive
performance and generalization. Leave‐one‐island‐out cross‐validation (LOIOCV) evaluated the model's ability
to generalize to geographically distinct regions. In this approach, the model was iteratively trained on data from all
islands except one, then tested on the held‐out island. Each held‐out island represents pedogenic and land‐use
conditions partially absent from training, providing a robust test of spatial transfer.
We also performed 10‐fold cross‐validation to assess overall predictive performance across the full range of
observed SOC conditions. Locations were randomly partitioned into 10 subsets, and the model was trained and
tested iteratively with each subset serving once as the test set (approximately 500 train and 56 test locations per
fold). This validation strategy complements the spatial LOIOCV approach by ensuring the model performs
consistently across diverse climate conditions, land uses, and SOC concentration ranges present in the training
data. Because both schemes partition by location rather than by time, they evaluate spatial prediction and
generalization. We do not validate temporal predictions directly, as repeated soil measurements through time are
unavailable (see Section
4
).
Model performance was quantified using multiple complementary metrics. The coefficient of determination (
R
2
)
assessed overall prediction fit and linear agreement between observed and predicted values. Root mean squared
error (RMSE) quantified absolute prediction error magnitude in the original SOC units. The concordance cor-
relation coefficient (CCC) evaluated both accuracy and precision by incorporating deviation from the 1:1 line,
providing a more stringent assessment than
R
2
alone. Bias was calculated as the mean difference between pre-
dicted and observed values to identify systematic over‐ or under‐prediction tendencies. The ratio of performance
Figure 2.
Transformer sliding‐window prediction framework. For each prediction, a 36‐month input window (timesteps t
0
to t
35
) of 27 features is passed to the
transformer encoder/decoder, which predicts soil organic carbon (SOC) at the following timestep (t
36
). Each input timestep comprises six Landsat spectral bands, 14
TerraClimate variables, cyclical day‐of‐year encoding, and five static terrain covariates (see feature set listed in Table
1
). SOC, as the prediction target, is predicted
alongside the full set of covariates at each timestep, but is never included among the model inputs. To generate a continuous monthly SOC series, the window advances
by 1 month (e.g., t
1
to 1
36
predicting t
37
) and the prediction repeats. Successive windows are processed independently such that no predicted SOC value is returned to any
input window.
Earth's Future
10.1029/2026EF008563
PETER‐CONTESSE ET AL.
6 of 18
23284277, 2026, 9, Downloaded from https://agupubs.onlinelibrary.wiley.com/doi/10.1029/2026EF008563 by Oregon State University, Wiley Online Library on [19/10/2026]. See the Terms and Conditions (https://onlinelibrary.wiley.com/terms-and-conditions) on Wiley Online Library for rules of use; OA articles are governed by the applicable Creative Commons License
Page 7
to interquartile range (RPIQ) normalized RMSE by the interquartile range of observed values, facilitating
comparison of model performance across data sets with different SOC concentration ranges.
2.6. Analysis of SOC Change by Land Use and Climate
We classified land use transitions by comparing 1978–1980 agricultural land use (Hawaii Statewide GIS Pro-
gram,
2021
) with 2020 land use (Perroy & Collier,
2020
) for each polygon (Table S3 in Supporting Informa-
tion
S1
). For each transition type, we extracted monthly SOC predictions for 2017–2024, calculated seasonally
decomposed mean annual values for 2017 and 2024, and computed the change in SOC concentration. Point‐level
predictions were aggregated to 337 agricultural management polygons, and we retained nine land use transitions
represented by at least two polygons, resulting in 297 polygons used for analysis. Land uses included rangeland,
pineapple plantation, sugarcane plantation, abandoned lands, and diversified agriculture. Diversified agriculture
broadly includes leafy vegetables and herbs, dryland taro, potatoes and sweet potatoes, melons, cucumbers, to-
matoes, ginger, beans, asparagus, and small plantings of banana, tropical fruits, and coffee (Melrose et al.,
2015
).
Abandoned agricultural lands were identified through temporal comparison of agricultural land maps. Polygons
classified as agriculture in the baseline year (1980) but reclassified to non‐agricultural uses in the subsequent time
period (2015 or 2020, respectively) were designated as abandoned agricultural lands.
2.7. Statistical Analysis
We compared model‐predicted polygon‐level SOC for 2017 and 2024. For each transition, we performed paired
statistical tests to determine whether predicted SOC concentrations differed significantly between the 2 years.
Statistical significance was determined at
α
=
0.05. All statistical analyses were performed in Python using the
SciPy library (Virtanen et al.,
2020
). These paired comparisons assess whether model‐predicted SOC differs
consistently across management polygons between 2017 and 2024. They test the spatial consistency of the
predicted change and do not account for the model's prediction uncertainty (RMSE
≈
4 percentage points; see
Section
2.5
).
Prior to hypothesis testing, we assessed the normality of paired differences (2024–2017) using the Shapiro‐Wilk
test. When differences were normally distributed (
p
>
0.05), we used paired t‐tests to compare means. For non‐
normally distributed data (
p
≤
0.05), we used the non‐parametric Wilcoxon signed‐rank test. Effect sizes were
calculated using Cohen's d, defined as the mean difference divided by the standard deviation of differences. We
interpreted effect sizes as small (|
d
|
<
0.2), medium (0.2
≤
|
d
|
<
0.8), or large (|
d
|
≥
0.8) following Cohen (
1988
).
We then tested the effects of land use and climate in two complementary analyses. In both, polygons were
assigned to climate bins defined by maximum annual temperature and precipitation [thresholds in Figure S1 and
Table S2 in Supporting Information
S1
; bins: cool/moderate‐high rainfall, moderate temperature/low‐moderate
rainfall, moderate temperature/high rainfall, warm/low rainfall, and warm/moderate‐high rainfall]. For all
Kruskal‐Wallis tests we applied a minimum inclusion threshold of
n
≥
3 polygons per group and used Dunn's
post‐hoc test with FDR correction. First, to test whether current land use affects predicted SOC change while
holding climate approximately constant (Table
3
), we grouped polygons into a warmer/drier and cooler/wetter
regime, reflecting that sugarcane occupied warmer areas while rangeland, pineapple, and diversified land uses
occupied cooler areas. We ran the Kruskal‐Wallis test across transitions within each regime to address the
questions: (a) among formerly diversified or pineapple lands, does current land use affect SOC change? and (b)
among formerly sugarcane lands, does current land use affect SOC change? Second, to test whether climate
effects predicted SOC change while holding land use constant (Table
4
), we ran Kruskal‐Wallis across the full set
of temperature x rainfall climate bins within individual transitions, restricted to trainsitions represented across
multiple bins (continuous rangeland, pineapple to abandoned, and pineapple to rangeland). Former sugarcane
transitions could not be included in this within‐transition climate analysis, since sugarcane occupied the warmest,
driest portion of the gradient almost exclusively. Land‐use legacy and climate therefore cannot be statistically
separated at the dry end of the gradient, and the climate effect could be tested only within transitions occupying
the moderate‐temperature portion of the gradient.
Earth's Future
10.1029/2026EF008563
PETER‐CONTESSE ET AL.
7 of 18
23284277, 2026, 9, Downloaded from https://agupubs.onlinelibrary.wiley.com/doi/10.1029/2026EF008563 by Oregon State University, Wiley Online Library on [19/10/2026]. See the Terms and Conditions (https://onlinelibrary.wiley.com/terms-and-conditions) on Wiley Online Library for rules of use; OA articles are governed by the applicable Creative Commons License
Page 8
2.8. Prediction Uncertainty
To characterize how prediction error varied across the range of predicted SOC, we binned cross‐validation
predictions into five percent intervals and summarized error within each interval. For both the leave‐one‐
island‐out and 10‐fold cross‐validation schemes, predictions were assigned to 5‐percentage‐point bins based
on predicted SOC value. Within each bin we calculated the bias (mean of predicted minus observed values) and
Table 2
SOC Change by Land Use Transition Based on Historical (1978–1980) and Recent (2020) Agricultural Land Use Polygons (Data Sources Specified in Table S3 in
Supporting Information
S1
) With Paired Statistical Tests for 2017 and 2024 SOC Predictions
Land use transition
Management unit polygons; (n) (SOC increasing,
SOC decreasing)
2017
SOC (%)
±
SD
2024
SOC (%)
±
SD
Absolute SOC change,
2017–2024 (%)
±
SD
Cohen's
d
p
‐value
pineapple →
abandoned
33 (0, 33)
4.72
±
0.42
4.22
±
0.31
−
0.50
±
0.30
−
1.68
<
0.0001
a
pineapple →
rangeland
19 (1, 18)
4.29
±
0.7
3.76
±
0.58
−
0.52
±
0.29
−
1.82
<
0.0001
b
sugarcane →
diversified
5 (0, 5)
3.09
±
0.38
1.96
±
0.31
−
1.13
±
0.42
−
2.67
0.004
b
sugarcane →
abandoned
75 (1, 74)
3.13
±
0.51
2.03
±
0.66
−
1.10
±
0.47
−
2.37
<
0.0001
b
sugarcane →
rangeland
47 (3, 44)
3.56
±
0.45
2.85
±
0.66
−
0.71
±
0.44
−
1.62
<
0.0001
b
diversified →
diversified
34 (7, 27)
4.55
±
1.05
4.24
±
1.08
−
0.30
±
0.40
−
0.77
0.0001
b
diversified →
pineapple
44 (10, 34)
4.68
±
0.92
4.45
±
1.03
−
0.24
±
0.53
−
0.45
0.005
a
diversified →
rangeland
4 (0, 4)
5.01
±
0.50
4.41
±
0.55
−
0.61
±
0.33
−
1.82
0.0358
b
rangeland →
rangeland
36 (2, 34)
5.65
±
1.46
5.13
±
1.71
−
0.52
±
0.40
−
1.31
<
0.0001
a
a
Wilcoxon signed‐rank test (non‐normal distribution).
b
Paired t‐test; SD is Standard Deviation.
Table 3
Land Use Transition Effects Within Climate Regimes
Land use transition
Management unit polygons (n)
Mean ΔSOC (%)
a
SD
Dunn's Group
b
Warmer, drier regime (Kruskal‐Wallis:
H
=
57.1,
p
<
0.0001)
sugarcane → abandoned
75
−
1.10
0.47
A
sugarcane → rangeland
47
−
0.72
0.44
B
pineapple → abandoned
21
−
0.50
0.30
C
pineapple → rangeland
10
−
0.52
0.29
C
rangeland → rangeland
10
−
0.52
0.40
C
diversified → pineapple
8
−
0.24
0.53
C
Cooler, wetter regime (Kruskal‐Wallis:
H
=
20.3,
p
=
0.0004)
pineapple → abandoned
12
−
0.50
0.30
A
pineapple → rangeland
9
−
0.52
0.29
A
rangeland → rangeland
26
−
0.52
0.40
A
diversified → diversified
33
−
0.30
0.40
B
diversified → pineapple
36
−
0.24
0.53
B
a
Mean ΔSOC represents absolute change in SOC percentage points (2024–2017).
b
Dunn's post‐hoc groupings with FDR
correction. Transitions sharing letters are not significantly different (
p
≥
0.05).
Earth's Future
10.1029/2026EF008563
PETER‐CONTESSE ET AL.
8 of 18
23284277, 2026, 9, Downloaded from https://agupubs.onlinelibrary.wiley.com/doi/10.1029/2026EF008563 by Oregon State University, Wiley Online Library on [19/10/2026]. See the Terms and Conditions (https://onlinelibrary.wiley.com/terms-and-conditions) on Wiley Online Library for rules of use; OA articles are governed by the applicable Creative Commons License
Page 9
the standard deviation of residuals, yielding a prediction‐indexed character-
ization of systematic error and variability across the SOC range (Tables S4
and S5 in Supporting Information
S1
). This approach quantifies the expected
error associated with a given predicted SOC value, directly informing inter-
pretation of the spatial predictions and predicted change on Maui. Because
repeated soil measurements through time are unavailable, it characterizes
uncertainty in SOC prediction across the concentration range rather than
spatially explicit or temporal prediction uncertainty.
2.9. Model Comparisons
To evaluate whether the transformer's processing of full temporal sequences
provides predictive advantage over conventional approaches, we compared it
against six machine learning models: Cubist, extremely randomized trees
(extra‐trees; ERT), random forest (RF), extreme gradient boosting
(XGBoost), gradient‐boosted decision trees (GBDT), and support vector
regression (SVR). The comparison was designed to isolate the contribution of
temporal information: all models received the same covariates, but whereas
the transformer ingests the full 36‐month sequence at each location, the
comparison models cannot accept sequential input and were instead provided
each temporal covariate reduced to its mean over the 36‐month time window.
Static terrain covariates were identical across the models. The transformer's
performance relative to models given only temporal means therefore reflects
the predictive value of the temporal structure itself, beyond the information
contained in the covariate means. All comparison models were evaluated
under the same leave‐one‐island‐out cross‐validation scheme as the transformer and default hyperparameters
were used.
3. Results
3.1. Cross‐Validation
We performed two separate cross‐validations for the transformer model: 10‐fold and leave‐one‐island‐out cross‐
validation. Ten‐fold cross‐validation across all Hawaiian Islands (
n
=
556) yielded an
R
2
of 0.60, Lin's
concordance correlation coefficient (CCC) of 0.79, and RPIQ of 2.00 (Figure
3
), indicating acceptable to
excellent model performance. The RMSE was 4.0 percentage points with minimal bias (0.1 percentage points),
suggesting consistent predictions across the range of observed SOC concentrations (0%–40%).
Leave‐one‐island‐out cross‐validation (LOIOCV) assessed the model's ability to predict SOC on islands excluded
from training (Figure
4
). Across the three main islands with sufficient samples (
>
40 samples; Hawaiʻi Island,
Maui, and Oʻahu;
n
=
493 combined), the model achieved an
R
2
of 0.57, CCC of 0.77, and RPIQ of 1.88
(acceptable to excellent performance), with RMSE of 4.3 percentage points and minimal bias (
−
0.2 percentage
points). Island‐specific performance was consistent, with
R
2
ranging from 0.50 (Hawaiʻi Island) to 0.61 (Oʻahu)
and RPIQ values between 1.76 and 1.99 (all acceptable to excellent). These results demonstrate the model's ability
to capture generalizable SOC‐environment relationships that transfer across islands with distinct pedogenic and
land use histories.
3.2. Soil Carbon Spatial Predictions
Mean 0–20 cm SOC predictions for the 2017–2024 study period exhibited strong spatial patterning, following
elevation/rainfall, land management, and topographic gradients across the Haleakalā subregion (Figure
5
). The
highest predicted SOC (
∼
5%–16%) was in higher rainfall regions and at greater altitudes, while low SOC
(0.02%–3%) tended to follow individual management units in lower lying areas.
Change in predicted SOC between 2017 and 2024 was overwhelmingly negative (decreasing) across the 297
management units, with mean change
−
0.65%
±
0.53, maximum decreases of 2.12%, but some management units
displaying increases of up to 1.97% (Figure
6
).
Figure 3.
10‐fold cross‐validation results for observed versus predicted SOC
percentage across the main Hawaiian Islands.
Earth's Future
10.1029/2026EF008563
PETER‐CONTESSE ET AL.
9 of 18
23284277, 2026, 9, Downloaded from https://agupubs.onlinelibrary.wiley.com/doi/10.1029/2026EF008563 by Oregon State University, Wiley Online Library on [19/10/2026]. See the Terms and Conditions (https://onlinelibrary.wiley.com/terms-and-conditions) on Wiley Online Library for rules of use; OA articles are governed by the applicable Creative Commons License
Page 10
3.3. Climate Effects on Soil Carbon Temporal Predictions
All climate zones experienced decline in mean SOC from 2017 to 2024, with the largest losses in warm/low
rainfall zones (
−
1.11%
±
0.43 loss) and the smallest losses in cool/moderate‐high rainfall zones (
−
0.14%
±
0.31
loss; Figure
7
).
3.4. Land Use Effects on Soil Carbon
Management units that transitioned from sugarcane to either abandoned or diversified experienced the greatest
SOC losses (
−
1.10
±
0.47 and
−
1.13
±
0.42, respectively; mean loss in sugarcane across all transi-
tions
= −
0.96
±
0.49), while regions that maintained diversified crops, pineapple, or rangeland lost less SOC
(
−
0.24% to
−
0.61%) overall, even gaining C in some management units (Figures
6
and
8
; Table
2
).
Figure 4.
Leave‐one‐island‐out cross‐validation for Hawaiian Islands with
>
40 samples.
Earth's Future
10.1029/2026EF008563
PETER‐CONTESSE ET AL.
10 of 18
23284277, 2026, 9, Downloaded from https://agupubs.onlinelibrary.wiley.com/doi/10.1029/2026EF008563 by Oregon State University, Wiley Online Library on [19/10/2026]. See the Terms and Conditions (https://onlinelibrary.wiley.com/terms-and-conditions) on Wiley Online Library for rules of use; OA articles are governed by the applicable Creative Commons License
Page 11
Figure 5.
Mean 2017–2024 SOC (%) predictions for 0–20 cm mineral soil depth on Haleakalā.
Figure 6.
(a) Predicted SOC % change in 0–20 cm depth from 2017 to 2024 aggregated to management units. (b) Point estimates of predicted SOC % change in 0–20 cm
depth from 2017 to 2024.
Earth's Future
10.1029/2026EF008563
PETER‐CONTESSE ET AL.
11 of 18
23284277, 2026, 9, Downloaded from https://agupubs.onlinelibrary.wiley.com/doi/10.1029/2026EF008563 by Oregon State University, Wiley Online Library on [19/10/2026]. See the Terms and Conditions (https://onlinelibrary.wiley.com/terms-and-conditions) on Wiley Online Library for rules of use; OA articles are governed by the applicable Creative Commons License
Page 12
3.5. Effects of Climate and Land Use on SOC Concentrations
We partitioned the study area into (a) cooler, wetter and (b) warmer, drier climate regimes to account for
geographic clustering of historical land use. Within each regime, land use transitions significantly affected SOC
change (Kruskal‐Wallis: warmer
H
=
57.1,
p
<
0.0001; cooler
H
=
20.3,
p
=
0.0004; Table
3
).
In the warmer, drier regime, former sugarcane lands showed the greatest losses, with the sugarcane transition to
abandoned declining significantly more than all other transitions (Dunn's post‐hoc
p
<
0.003). In the cooler,
wetter regime, maintaining diversified agriculture and diversified transition to pineapple lost less SOC than
former pineapple lands and continuous rangeland (Dunn's post‐hoc
p
<
0.02; Table
3
).
Climate modulated SOC change only in transitions lacking active management (Table
4
). Abandoned pineapple
lands in the moderate zone (moderate temperature, low‐moderate rainfall) lost twice the SOC (
−
0.85
±
0.36%)
compared to wetter zones (
−
0.34%–0.40%; Kruskal‐Wallis
H
=
9.53,
p
=
0.0085). Similarly, continuous ran-
geland showed greater loss in the moderate zone (
−
0.76%) than in the coolest zone (
−
0.08%,
p
=
0.033). In
contrast, active conversion of abandoned pineapple lands to managed rangeland showed no climate effect
(
H
=
0.03,
p
=
0.98).
3.6. Prediction Uncertainty
We characterized prediction uncertainty by binning predictions by SOC concentration and summarizing bias and
residual variability within each bin for both cross‐validation schemes (Tables S4 and S5 in Supporting Infor-
mation
S1
). Overall bias was minimal (LOIOCV: 0.06; 10‐fold:
−
0.32 percentage points), but both bias and
spread were strongly concentration‐dependent. The model slightly underpredicted in the lowest‐SOC bin
(0%–5%: bias
−
1.02 and
−
0.74 percentage points in LOIOCV and 10‐fold, respectively) and overpredicted at
mid‐range concentrations (10%–20%), with residual standard deviation increasing markedly with SOC (from
∼
1.9–2.3 percentage points in the 0%–5% bin to
∼
9 percentage points above 15%). The sparsely populated
highest bins (
≥
20%,
n
=
3–15) showed large bias and high variability, reflecting few samples and regression of
extreme values toward the mean.
Figure 7.
Predicted mean SOC concentrations (
±
SD) across all islands in 0–20 cm depth by year and climate bin, aggregated by management unit (n).
Earth's Future
10.1029/2026EF008563
PETER‐CONTESSE ET AL.
12 of 18
23284277, 2026, 9, Downloaded from https://agupubs.onlinelibrary.wiley.com/doi/10.1029/2026EF008563 by Oregon State University, Wiley Online Library on [19/10/2026]. See the Terms and Conditions (https://onlinelibrary.wiley.com/terms-and-conditions) on Wiley Online Library for rules of use; OA articles are governed by the applicable Creative Commons License
Page 13
This error structure is relevant to interpreting predicted change on Maui, where the former sugarcane polygons
showing the largest predicted losses fall in the 2%–3% SOC range. In this range the model is most reliable: re-
sidual standard deviation is lowest (
∼
1.9–2.3 percentage points) and bias is small, indicating that predicted SOC
and predicted change in these polygons fall within the best‐constrained portion of the model's range. Predictions
in higher‐SOC transitions (e.g., continuous rangeland and diversified lands, with baselines of 4.5%–5.7%) carry
correspondingly greater uncertainty.
3.7. Model Comparisons
To assess whether the transformer's processing of full temporal sequences provides predictive advantage over
conventional approaches, we compared it against six machine learning models trained on the same covariates
reduced to their temporal means (Table
5
). Under leave‐one‐island‐out cross‐validation, the transformer out-
performed all six comparison models on every pooled metric, with lower RMSE (4.28 vs. 4.68–5.31 percentage
points), higher concordance (CCC 0.77 vs. 0.55–0.72), and higher RPIQ (1.88 vs. 1.56–1.77). Among the
comparison models, performance declined from tree‐based ensembles (Cubist, ERT, RF; RMSE 4.68–4.77, CCC
0.70–0.72) through gradient‐boosting approaches (XGBoost, GBDT; RMSE 4.88–5.08) to support vector ma-
chine, which performed worst (RMSE 5.31, CCC 0.55).
Figure 8.
(a) Historical land use by management unit (
∼
1980). (b) 2020 land use by management unit. (c) Sankey diagram of SOC changes by land use transition using
mean predictions from 2017 to 2024. The width of each flow path represents the mean relative SOC percentage in the starting and ending years. Land area (ha) is
reported by land use for each year.
Earth's Future
10.1029/2026EF008563
PETER‐CONTESSE ET AL.
13 of 18
23284277, 2026, 9, Downloaded from https://agupubs.onlinelibrary.wiley.com/doi/10.1029/2026EF008563 by Oregon State University, Wiley Online Library on [19/10/2026]. See the Terms and Conditions (https://onlinelibrary.wiley.com/terms-and-conditions) on Wiley Online Library for rules of use; OA articles are governed by the applicable Creative Commons License
Page 14
Island‐specific
R
2
revealed that the transformer's advantage was not uniform. The transformer generalized best to
Oʻahu (
R
2
=
0.61), substantially exceeding all comparison models (0.32–0.53) on the island where the con-
ventional models struggled most. On Maui, the transformer (
R
2
=
0.60) performed comparably to the strongest
tree‐based models (0.57–0.63). On Hawaiʻi Island, however, the transformer's
R
2
(0.50) fell in the middle of the
comparison‐model range (0.45–0.60), with XGBoost achieving the highest single‐island value (0.60). The
transformer's clearest gains thus occurred in cross‐island transfer to Oʻahu rather than uniformly, indicating that
temporal sequence information most aided generalization where static‐covariate models had the greatest
difficulty.
4. Discussion
Previous research on Haleakalā revealed that SOC patterns are shaped by climate gradients, parent material age,
volcanic ash inputs, and land‐use histories. Hodges et al. (
2024
) found that SOC increased with precipitation until
approximately 2,300 mm MAP (mean annual precipitation), after which SOC declined, with highest stocks
Table 4
Climate Effects Within Land Use Transitions
Climate bin
Management unit polygons (n)
Mean ΔSOC (%)
a
SD
Dunn's Group
b
rangeland → rangeland (Kruskal‐Wallis:
H
=
8.75,
p
=
0.033)
cool/moderate‐high rainfall
moderate temperature/low‐moderate rainfall
6
9
−
0.08
−
0.76
0.67
0.29
A
B
moderate temperature/high rainfall
11
−
0.57
0.13
AB
warm/moderate‐high rainfall
10
−
0.51
0.28
AB
pineapple → abandoned (Kruskal‐Wallis:
H
=
9.53,
p
=
0.0085)
moderate temperature/low‐moderate rainfall
8
−
0.85
0.36
A
moderate temperature/high rainfall
4
−
0.34
0.11
B
warm/moderate‐high rainfall
21
−
0.40
0.17
B
pineapple → rangeland (Kruskal‐Wallis:
H
=
0.03,
p
=
0.98)
moderate temperature/low‐moderate rainfall
6
−
0.56
0.27
–
moderate temperature/high rainfall
3
−
0.55
0.35
–
warm/moderate‐high rainfall
10
−
0.49
0.31
–
a
Mean ΔSOC represents absolute change in SOC percentage points (2024–2017).
b
Dunn's post‐hoc groupings with FDR correction. Transitions sharing letters are not
significantly different (
p
≥
0.05).
Table 5
Performance Metrics for Comparison Among Machine Learning Models
a
Model
b
RMSE
CCC
RPIQ
Hawaiʻi Island
R
2
Maui
R
2
Oʻahu
R
2
Cubist
4.69
0.71
1.76
0.55
0.60
0.53
ERT
4.68
0.72
1.77
0.56
0.63
0.42
RF
4.77
0.70
1.74
0.55
0.62
0.42
XGBoost
4.88
0.70
1.70
0.60
0.57
0.37
GBDT
5.08
0.69
1.63
0.50
0.54
0.32
SVR
5.31
0.55
1.56
0.45
0.46
0.37
Transformer (LOIOCV)
4.28
c
0.77
c
1.88
c
0.50
0.60
0.61
Transformer (10‐fold)
4.02
c
0.79
c
2.00
c
–
–
–
Note.
Bold indicates the transformer model's results, shown for ease of comparison against baseline models.
a
Transformer
ingests full temporal sequences, while all other models are fed temporally aggregated mean values.
b
All models were
evaluated under leave‐one‐island‐out cross‐validation except where noted. ERT is extra‐trees, RF is random forest, XGBoost
is extreme gradient boosting, GBDT is gradient‐boosted decision trees, and SVR is support vector regression.
c
Results were
pooled across the main Hawaiian islands.
Earth's Future
10.1029/2026EF008563
PETER‐CONTESSE ET AL.
14 of 18
23284277, 2026, 9, Downloaded from https://agupubs.onlinelibrary.wiley.com/doi/10.1029/2026EF008563 by Oregon State University, Wiley Online Library on [19/10/2026]. See the Terms and Conditions (https://onlinelibrary.wiley.com/terms-and-conditions) on Wiley Online Library for rules of use; OA articles are governed by the applicable Creative Commons License
Page 15
occurring at mid‐elevation mesic sites. Krueger (
2021
) surveyed actively grazed rangeland and found average
SOC stocks of 552 Mg C ha
−
1
to 1 m depth, with short‐range‐order minerals serving as the strongest predictors.
Crow et al. (
2023
) demonstrated that unmanaged previously intensive agricultural lands were the most degraded
across the Hawaiian Islands, while well‐managed pastures exhibited soil health similar to protected forests. Wells
et al. (
2017
) documented modest SOC sequestration (2.34
±
1.03 Mg C ha
−
1
yr
−
1
in 0–30 cm) following con-
version of sugarcane fields to zero‐tillage perennial grasses. Our findings of persistent SOC depletion in former
sugarcane lands align with these broader patterns of agricultural legacy effects in Hawaiian soils.
The magnitude and persistence of SOC losses in former sugarcane lands (
−
1.10
±
0.47 percentage points in
sugarcane to abandoned transitions) is consistent with disruption of mineral‐organic associations critical for C
stabilization in volcanic soils. Osher et al. (
2003
) demonstrated that sugarcane management practices, including
liming, fertilization, and silica amendments, alter soil solution chemistry in ways that dissociate iron and
aluminum oxide‐humus complexes in Andisols. Once released from mineral protection, organic matter becomes
vulnerable to both microbial decomposition and vertical transport as dissolved organic C. This mechanism ex-
plains why sugarcane cultivation created more severe and long‐lasting depletion than pre‐contact agriculture
documented by Cusack et al. (
2013
), who found 26% lower SOC (0–30 cm) on Kohala Peninsula more than
200 years after cultivation ceased, with radiocarbon dating revealing older, less dynamic C pools. The interaction
between SRO minerals and land use further compounds these effects. Cusack et al. (
2013
) observed that high‐
SRO mineral soils, despite superior C storage capacity, exhibited the largest agricultural‐induced losses and
slowest recovery on Kohala. Our findings extend this pattern to industrial‐era agriculture, where former sugarcane
lands in the leeward regime (warmer, drier, corresponding to lower SRO mineral content) showed losses of
−
1.10% while continuous rangeland in high‐SRO zones showed smaller but still significant losses (
−
0.52%).
Crow et al. (
2023
) demonstrated that unmanaged previously intensive agricultural lands remain the most
degraded across the Hawaiian Islands, underscoring that disrupted mineral‐organic associations may not readily
restore even under conditions favorable for C accumulation.
The critical role of active management in determining post‐agricultural C trajectories is evident across Hawaiian
systems. Well‐managed rangelands in Hawaiʻi can maintain high SOC stocks, particularly on volcanic Andisols,
with some of the highest stocks recorded under century‐long grazing (Bremer et al.,
2021
). Similarly, Crow
et al. (
2020
) measured rapid SOC accumulation (11.4 Mg C ha
−
1
year
−
1
to 100 cm depth) following conversion of
intensive sugarcane to no‐till energycane within our same study area on Maui. In contrast, our findings of C losses
in pineapple to abandoned transitions, particularly in drier climates (
−
0.85%), and climate‐mediated losses in
continuous rangeland (
−
0.76%) likely reflect degradation when agricultural lands lack active management. The
difference between pineapple to rangeland transitions showing minimal climate effects (
p
=
0.98) and abandoned
lands showing strong rainfall‐driven losses suggests that continuous rangeland management may buffer against
climate‐driven C losses.
The inferred climate effects across land use transitions suggest three patterns. First, among abandoned pineapple
lands, the moderate‐temperature, lower‐rainfall zone showed roughly twice the inferred SOC loss of higher‐
rainfall zones (
−
0.85% vs.
−
0.34%–0.40%). This pattern is consistent with water constraints on revegetation:
slow secondary succession in the drier zone may allow decomposition of legacy organic matter to exceed new
inputs from sparse vegetation regrowth. Second, continuous rangeland appeared responsive to both temperature
and rainfall, with the coolest zone showing minimal loss (
−
0.08%), consistent with slower decomposition, and the
moderate‐temperature, drier zone showed the greatest loss (
−
0.76%), consistent with water limitation on grass
productivity and C inputs. Third, conversion of abandoned pineapple lands to managed rangeland showed no
detectable climate effect, suggesting that active management may buffer SOC against both temperature and
precipitation constraints by maintaining C inputs across the gradient.
Our cross‐validation framework accounts for the complexities of Hawaiian landscapes, where island‐specific
combinations of climate, geology, and land use history create unique pedogenic conditions. LOIOCV pro-
vided insight into whether the model captures generalizable relationships between environmental drivers and
SOC dynamics or relies on island‐specific patterns that may not transfer across the archipelago. The model's
spatial prediction performance (LOIOCV
R
2
=
0.57) provides confidence in the spatial predictions underlying the
land use and climate effects.
Comparison against conventional models indicates that the transformer's advantage stems from its use of temporal
sequence structure. When the same covariates were provided to six comparison models as temporal means, the
Earth's Future
10.1029/2026EF008563
PETER‐CONTESSE ET AL.
15 of 18
23284277, 2026, 9, Downloaded from https://agupubs.onlinelibrary.wiley.com/doi/10.1029/2026EF008563 by Oregon State University, Wiley Online Library on [19/10/2026]. See the Terms and Conditions (https://onlinelibrary.wiley.com/terms-and-conditions) on Wiley Online Library for rules of use; OA articles are governed by the applicable Creative Commons License
Page 16
transformer achieved lower error and higher concordance under identical leave‐one‐island‐out cross‐validation
(RMSE 4.28 vs. 4.68–5.31; CCC 0.77 vs. 0.55–0.72). Because the comparison models received the same envi-
ronmental information reduced to temporal averages, this margin reflects the predictive value of the within‐
sequence temporal structure itself, rather than the covariates alone. The advantage was not uniform across
islands: the transformer generalized markedly better to Oʻahu (
R
2
=
0.61 vs. 0.32–0.53), performed comparably
to the strongest tree‐based models on Maui, and fell within the comparison‐model range on Hawaiʻi Island. The
clearest gains thus occurred in cross‐island transfer to the island where static‐covariate models struggled most,
indicating that temporal sequence information is most valuable precisely where generalization is hardest. This
supports our central methodological contribution: temporal attention over satellite time series extracts pheno-
logical and management signals that static‐covariate digital soil mapping cannot.
Prediction error was concentration‐dependent, with implications for which inferences the model best supports.
Residual variability was lowest at low predicted SOC (
∼
1.9–2.3 percentage points below 5%) and increased
substantially at higher concentrations (Tables S4 and S5 in Supporting Information
S1
). The former sugarcane
polygons central to our land‐use findings fall in the 2%–3% range, where the model is most reliable, lending
confidence to the inferred losses along those transitions. Predictions in higher‐SOC transitions, including
continuous rangeland and diversified lands, carry correspondingly greater uncertainty, and inferences there
should be treated as more provisional.
Our statistical tests of paired change (Tables
2–4
) assess whether predicted SOC differs consistently across
management polygons between 2017 and 2024; they evaluate the spatial consistency of the inferred change rather
than confidence that any single polygon's change exceeds model prediction error. The mean inferred changes
(0.24–1.13 percentage points) are smaller than the model's cross‐validation RMSE (
∼
4 percentage points), so
individual‐polygon changes should not be considered resolved at that precision. Our inferences instead rest on the
systematic patterning of these changes. Inferred losses decline in a consistent order across land use, from former
sugarcane to former pineapple to continuously managed lands, and across climate, from lower‐rainfall to wetter
zones. This ordering is difficult to attribute to prediction noise, which would not align systematically with land‐
use history and climate. The convergence of these inferred patterns with independent evidence from Hawaiian
agricultural legacy studies (Crow et al.,
2023
; Cusack et al.,
2013
) lends further support.
Our temporal inferences rest on a space‐for‐time substitution rather than direct observation of change. The model
learns SOC‐environment relationships from observations distributed across space and applies them to time‐
varying covariates, yielding model‐inferred trajectories whose validity depends on the assumption that spatial
covariate‐SOC relationships approximate the temporal response of SOC to changing conditions. Robust SOC data
sets included in our study, particularly repeat measurements such as those in the FIA program, provide critical
baseline data, though realizing the full potential of transformer‐based temporal prediction will require pairing
such systematic monitoring with the continuous sensor approaches.
We have demonstrated the first application of transformer architectures to soil carbon spatiotemporal prediction,
with the transformer architecture's capacity for utilizing satellite timeseries enabling spatiotemporal prediction
not possible with other modeling frameworks. However, the method is fundamentally limited by the paucity of
model training data, particularly the lack of timeseries soil C measurements—a challenge that pervades AI‐driven
SOC modeling more broadly (Grunwald,
2022
). Definitive validation of temporal prediction accuracy requires a
shift from episodic soil sampling to continuous monitoring approaches that can capture sub‐annual C dynamics.
Recent advances in multi‐sensor soil profiling technology demonstrate the technical feasibility of such moni-
toring, with systems now capable of collecting continuous, high‐resolution (mm‐scale) vertical soil data
(Grunwald et al.,
2024
). We thus argue for broader adoption of continuous sensor monitoring approaches to pair
the temporal resolution of satellite data on aboveground processes with equivalent temporal resolution in
belowground SOC dynamics, enabling the validation and refinement of spatiotemporal models that can inform
SOC and land management, particularly in regions like Hawaiʻi where C neutrality goals necessitate strategic
management of post‐agricultural landscapes.
Conflict of Interest
The authors declare no conflicts of interest relevant to this study.
Earth's Future
10.1029/2026EF008563
PETER‐CONTESSE ET AL.
16 of 18
23284277, 2026, 9, Downloaded from https://agupubs.onlinelibrary.wiley.com/doi/10.1029/2026EF008563 by Oregon State University, Wiley Online Library on [19/10/2026]. See the Terms and Conditions (https://onlinelibrary.wiley.com/terms-and-conditions) on Wiley Online Library for rules of use; OA articles are governed by the applicable Creative Commons License
Page 17
Availability Statement
The transformer model, data set of predictions for land management polygons, associated Python script for land
use and climate analyses, and geopackage containing soil carbon prediction spatial layers are available at
https://
doi.org/10.5281/zenodo.21876655
. Data sensitivity associated with original soil carbon data sets precludes
publication of coordinate locations; therefore, transformer training and testing data sets will not be published. The
publicly available soil and covariate data sets used in this paper are available at: TerraClimate:
https://developers.
google.com/earth‐engine/datasets/catalog/IDAHO_EPSCOR_TERRACLIMATE
(Abatzoglou et al.,
2018
);
Hawaii Soil Carbon database
https://doi.org/10.17605/OSF.IO/HMTV6
(Crow et al.,
2021
); Forest Inventory and
Analysis Database:
https://apps.fs.usda.gov/fia/datamart/datamart.html
(U.S. Forest Service,
2024
); Landsat 7:
https://doi.org/10.5066/P9C7I13B
(Earth Resources Observation and Science (EROS) Center,
2020a
); Landsat
8–9:
https://doi.org/10.5066/P9OGBGM6
(Earth Resources Observation and Science (EROS) Center,
2020b
);
Agricultural Land Use Maps:
https://geoportal.hawaii.gov/datasets/HiStateGIS::agricultural‐land‐use‐maps‐alu
m/about
(Hawaii Statewide GIS Program,
2021
); The Hawaiʻi Climate Data Portal (HCDP):
https://doi.org/10.
1175/BAMS‐D‐23‐0188.1
(Longman et al.,
2024
), State agricultural land use baseline 2015:
https://dab.hawaii.
gov/wp‐content/uploads/2016/02/StateAgLandUseBaseline2015.pdf
(Melrose et al.,
2015
); NASA Shuttle Ra-
dar Topography Mission:
https://developers.google.com/earth‐engine/datasets/catalog/USGS_SRTMGL1_003
(NASA JPL,
2013
); NCSS Soil Characterization Database:
https://ncsslabdatamart.sc.egov.usda.gov/
(National
Cooperative Soil Survey,
n.d.
); 2020 update to the Hawaiʻi statewide agricultural land use baseline:
https://dab.hawaii.
gov/wp‐content/uploads/2022/01/2020_Update_Ag_Baseline_all_Hawaiian_Islands_v5.pdf
(Perroy & Collier,
2020
);
Web Soil Survey:
https://websoilsurvey.nrcs.usda.gov/
(Soil Survey Staff,
2025
); USGS 3DEP:
https://www.usgs.
gov/3d‐elevation‐program
(U.S. Geological Survey, 3D Elevation Program (3DEP),
n.d.
).
References
Abatzoglou, J. T., Dobrowski, S. Z., Parks, S. A., & Hegewisch, K. C. (2018). TerraClimate, a high‐resolution global dataset of monthly climate
and climatic water balance from 1958–present [Dataset].
Scientific Data
,
5
, 170191.
https://doi.org/10.1038/sdata.2017.191
Basile‐Doelsch, R., Amundson, R., Stone, W. E. E., Borschneck, D., Bottero, J. Y., Moustier, S., et al. (2007). Mineral control of carbon pools in a
volcanic soil horizon.
Geoderma
,
137
(3–4), 477–489.
https://doi.org/10.1016/j.geoderma.2006.10.006
Blois, J. L., Williams, J. W., Fitzpatrick, M. C., Jackson, S. T., & Ferrier, S. (2013). Space can substitute for time in predicting climate‐change
effects on biodiversity.
PNAS
,
110
(23), 9374–9379.
https://doi.org/10.1073/pnas.1220228110
Bremer, L. L., Nathan, N., Trauernicht, C., Pascua, P., Krueger, N., Jokiel, J., et al. (2021). Maintaining the many societal benefits of rangelands:
The case of Hawaiʻi.
Land
,
10
(7), 764.
Cao, L., Sun, M., Yang, Z., Jiang, D., Yin, D., & Duan, Y. (2024). A novel transformer‐CNN approach for predicting soil properties from LUCAS
Vis‐NIR spectral data.
Agronomy
,
14
(9), 1998.
Cohen, J. (1988).
Statistical power analysis for the behavioral sciences
(2nd ed.). Lawrence Erlbaum Associates.
Coleman, K., & Jenkinson, D. S. (1996). RothC‐26.3‐A model for the turnover of carbon in soil. In
Evaluation of soil organic matter models:
Using existing long‐term datasets
(pp. 237–246). Springer Berlin Heidelberg.
Consoli, D., Tian, X., Isik, S., Simoes, R., & Hengl, T. (2025). OpenLandMap‐soildb: Soil organic carbon density (mg/cm3) 2000‐2005 0‐30cm
below ground (Version v20250204) [Dataset].
Zenodo
.
https://doi.org/10.5281/zenodo.15470432
Crow, S. E., Hubanks, H., Deenik, J. L., McClellan Maaz, T., Tallamy Glazer, C., Vizka, E., & Rivera‐Zayas, J. (2023). The legacy of intensive
agricultural history on the soil health of (sub)tropical landscapes.
Frontiers of Environmental Science
,
10
, 991262.
https://doi.org/10.3389/fenv
s.2022.991262
Crow, S. E., Rivera‐Zayas, J., & Vizka, E. (2021). Hawaii soil carbon database [Dataset].
https://doi.org/10.17605/OSF.IO/HMTV6
Crow, S. E., Wells, J. M., Sierra, C. A., Youkhana, A. H., Ogoshi, R. M., Richardson, D., et al. (2020). Carbon flow through energycane
agroecosystems established post‐intensive agriculture.
GCB Bioenergy
,
12
(10), 806–817.
Cusack, D. F., Chadwick, O. A., Ladefoged, T., & Vitousek, P. M. (2013). Long‐term effects of agriculture on soil carbon pools and carbon
chemistry along a Hawaiian environmental gradient.
Biogeochemistry
,
112
(1), 229–243.
https://doi.org/10.1007/s10533‐012‐9718‐z
Earth Resources Observation and Science (EROS) Center. (2020a). Landsat 7 enhanced thematic mapper plus Level‐2, collection 2 [Dataset].
U.S.
Geological Survey
.
https://doi.org/10.5066/P9C7I13B
Earth Resources Observation and Science (EROS) Center. (2020b). Landsat 8‐9 operational land imager/thermal infrared sensor Level‐2,
collection 2 [Dataset].
U.S. Geological Survey
.
https://doi.org/10.5066/P9OGBGM6
Grunwald, S. (2022). Artificial intelligence and soil carbon modeling demystified: Power, potentials, and perils.
Carbon Footprints
,
1
(1).
Grunwald, S., Murad, M. O. F., Farrington, S., Wallace, W., & Rooney, D. (2024). Multi‐sensor soil probe and machine learning modeling for
predicting soil properties.
Sensors
,
24
(21), 6855.
https://doi.org/10.3390/s24216855
Hawaii Statewide GIS Program. (2021). Agricultural land use maps (ALUM) for the islands of Kauai, Oahu, Maui, molokai, lanai and Hawaii,
1978‐1980 [GIS Dataset].
Office of Planning and Sustainable Development, State of Hawaii
. Retrieved from
https://geoportal.hawaii.gov/data
sets/HiStateGIS::agricultural‐land‐use‐maps‐alum/about
Heuvelink, G. B. M., Angelini, M. E., Poggio, L., Bai, Z., Batjes, N. H., van den Bosch, R., et al. (2021). Machine learning in space and time for
modelling soil organic carbon change.
European Journal of Soil Science
,
72
(4), 1607–1623.
Hiederer, R., & Köchy, M. (2011).
Global soil organic carbon estimates and the harmonized world soil database. EUR 25225 EN
. Publications
Office of the EU.
Hodges, R. C., Boettinger, J. L., & Deenik, J. L. (2024). Elemental trends and weathering indices to assess volcanic ash deposition on soil
weathering status along Maui climosequences.
Geoderma Regional
,
39
, e00872.
https://doi.org/10.1016/j.geodrs.2024.e00872
Acknowledgments
This material is based upon work
supported by the U.S. Department of
Agriculture Forest Service under U.S.
Forest Service Agreement 21‐JV‐
11261919‐089. Any opinions, findings,
conclusions, or recommendations
expressed in this publication are those of
the author(s) and do not necessarily reflect
the views of the U.S. Department of
Agriculture.
Earth's Future
10.1029/2026EF008563
PETER‐CONTESSE ET AL.
17 of 18
23284277, 2026, 9, Downloaded from https://agupubs.onlinelibrary.wiley.com/doi/10.1029/2026EF008563 by Oregon State University, Wiley Online Library on [19/10/2026]. See the Terms and Conditions (https://onlinelibrary.wiley.com/terms-and-conditions) on Wiley Online Library for rules of use; OA articles are governed by the applicable Creative Commons License
Page 18
Jat, H. S., Choudhary, M., Datta, A., Yadav, A. K., Meena, M. D., Devi, R., et al. (2020). Temporal changes in soil microbial properties and
nutrient dynamics under climate smart agriculture practices.
Soil and Tillage Research
,
199
, 104595.
https://doi.org/10.1016/j.still.2020.10
4595
Jin, X., Zhou, J., Rao, Y., Zhang, X., Zhang, W., Ba, W., et al. (2023). An innovative approach for integrating two‐dimensional conversion of Vis‐
NIR spectra with the Swin Transformer model to leverage deep learning for predicting soil properties.
Geoderma
,
436
, 116555.
https://doi.org/
10.1016/j.geoderma.2023.116555
Kakhani, N., Rangzan, M., Jamali, A., Attarchi, S., Alavipanah, S. K., Mommert, M., et al. (2024). SSL‐SoilNet: A hybrid transformer‐based
framework with self‐supervised learning for large‐scale soil organic carbon prediction.
IEEE Transactions on Geoscience and Remote
Sensing
,
62
, 1–15. Article 4509915.
https://doi.org/10.1109/TGRS.2024.3446042
Kramer, M. G., & Chadwick, O. A. (2018). Climate‐driven thresholds in reactive mineral retention of soil carbon at the global scale.
Nature
Climate Change
,
8
, 1104–1108.
https://doi.org/10.1038/s41558‐018‐0341‐4
Krueger, N. C. (2021).
Soil carbon in Hawaiian rangelands: Evaluation of extent, controls, and options for management
[Doctoral dissertation,
University of Hawaiʻi at Mānoa]. ScholarSpace. Retrieved from
http://hdl.handle.net/10125/75973
Longman, R. J., Lucas, M. P., Mclean, J., Cleveland, S. B., Kodama, K., Frazier, A. G., et al. (2024). The Hawaiʻi Climate Data Portal (HCDP)
[Dataset].
Bulletin of the American Meteorological Society
,
105
(7), E1225–E1249.
https://doi.org/10.1175/BAMS‐D‐23‐0188.1
Melrose, J., Perroy, R., & Cares, S. (2015). State agricultural land use baseline 2015 [GIS Dataset].
University of Hawai'i at Hilo, Spatial Data
Analysis & Visualization Research Lab
. Prepared for the State of Hawaii Department of Agriculture
https://dab.hawaii.gov/wp‐content/uploa
ds/2016/02/StateAgLandUseBaseline2015.pdf
NASA, J. P. L. (2013). NASA shuttle radar topography mission global 1 arc second [Dataset].
https://developers.google.com/earth‐engine/datase
ts/catalog/USGS_SRTMGL1_003
National Cooperative Soil Survey. (n.d.). NCSS soil characterization database [Dataset].
United States Department of Agriculture, Natural
Resources Conservation Service
. Retrieved from
https://ncsslabdatamart.sc.egov.usda.gov/
Osher, L. J., Matson, P. A., & Amundson, R. (2003). Effect of land use change on soil carbon in Hawaii.
Biogeochemistry
,
65
(2), 213–232.
Parton, W. J., Hartman, M., Ojima, D., & Schimel, D. (1998). DAYCENT and its land surface submodel: Description and testing.
Global and
Planetary Change
,
19
(1–4), 35–48.
Peel, M. C., Finlayson, B. L., & McMahon, T. A. (2007). Updated world map of the Köppen‐Geiger climate classification.
Hydrology and Earth
System Sciences
,
11
(5), 1633–1644.
https://doi.org/10.5194/hess‐11‐1633‐2007
Perroy, R., & Collier, E. (2020). 2020 update to the Hawaiʻi statewide agricultural land use baseline [GIS Dataset].
University of Hawaiʻi at Hilo,
Spatial Data Analysis & Visualization Research Laboratory
. Prepared for the Hawaiʻi Department of Agriculture
https://dab.hawaii.gov/wp‐co
ntent/uploads/2022/01/2020_Update_Ag_Baseline_all_Hawaiian_Islands_v5.pdf
Perroy, R. L., Melrose, J., & Cares, S. (2016). The evolving agricultural landscape of post‐plantation Hawaii.
Applied Geography
,
76
, 154–162.
Pickett, S. T. A. (1989). Space‐for‐Time substitution as an alternative to long‐term studies. In G. E. Likens (Ed.),
Long‐term studies in ecology:
Approaches and alternatives
(pp. 110–135). Springer.
Poggio, L., de Sousa, L. M., Batjes, N. H., Heuvelink, G. B. M., Kempen, B., Ribeiro, E., & Rossiter, D. (2021). SoilGrids 2.0: Producing soil
information for the globe with quantified spatial uncertainty.
SOIL
,
7
(1), 217–240.
https://doi.org/10.5194/soil‐7‐217‐2021
Rumpel, C., Amiraslani, F., Chenu, C., Garcia Cardenas, M., Kaonga, M., Koutika, L. S., et al. (2020). The 4p1000 initiative: Opportunities,
limitations and challenges for implementing soil organic carbon sequestration as a sustainable development strategy.
Ambio
,
49
(1), 350–360.
https://doi.org/10.1007/s13280‐019‐01165‐2
Scharlemann, J. P., Tanner, E. V., Hiederer, R., & Kapos, V. (2014). Global soil carbon: Understanding and managing the largest terrestrial carbon
pool.
Carbon Management
,
5
(1), 81–91.
https://doi.org/10.4155/cmt.13.77
Selmants, P. C., Giardina, C. P., Jacobi, J. D., & Zhu, Z. (Eds.) (2017).,
Baseline and projected future carbon storage and carbon fluxes in
ecosystems of Hawaiʻi (Professional Paper 1834)
. U.S. Department of the Interior, U.S. Geological Survey.
https://doi.org/10.3133/pp1834
Smith, P., Soussana, J.‐F., Angers, D., Schipper, L., Chenu, C., Rasse, D. P., et al. (2020). How to measure, report and verify soil carbon change to
realize the potential of soil carbon sequestration for atmospheric greenhouse gas removal.
Global Change Biology
,
26
(1), 219–241.
https://doi.
org/10.1111/gcb.14815
Soil Survey Staff. (2022).
Keys to soil taxonomy
(13th ed.). United States Department of Agriculture, Natural Resources Conservation Service.
Retrieved from
https://www.nrcs.usda.gov/resources/guides‐and‐instructions/keys‐to‐soil‐taxonomy
Soil Survey Staff. (2025). Natural resources conservation service, United States department of agriculture [GIS Dataset].
Web Soil Survey
.
Available online at
https://websoilsurvey.nrcs.usda.gov/
State of Hawaiʻi Department of Health. (2018). Act 15, Session Laws of Hawaii. Retrieved from
https://health.hawaii.gov/cab/files/2023/05/
SLH2018_Act15.pdf
State of Hawaiʻi Department of Health. (2022). Act 238, Session Laws of Hawaii. Retrieved from
https://health.hawaii.gov/cab/files/2023/05/
SLH2022_Act238.pdf
State of Hawaiʻi Department of Health, Clean Air Branch. (2025).
Hawaiʻi greenhouse gas emissions report for 2022
. Prepared by ICF and
University of Hawaiʻi Economic Research Organization (UHERO). Retrieved from
https://health.hawaii.gov/cab/files/2025/04/Final‐2022‐
GHG‐Emissions‐Report.pdf
The Hawaii Soil Health Database. (2025). University of Hawaii at Mānoa [Unpublished data].
Trumbore, S. E. (2000). Age of soil organic matter and soil respiration: Radiocarbon constraints on belowground C dynamics.
Ecological Ap-
plications
,
10
(2), 399–411.
https://doi.org/10.1890/1051‐0761(2000)010[0399:AOSOMA]2.0.CO;2
U.S. Geological Survey, 3D Elevation Program (3DEP). (n.d.). 1/3 arc‐second (10‐meter) digital elevation model [Dataset]. Retrieved from
https://www.usgs.gov/3d‐elevation‐program
U.S. Forest Service. (2024). Forest inventory and analysis database [Datasets: SOILS_LAB, SOILS_SAMPLE_LOC].
U.S. Department of
Agriculture, Forest Service, Pacific Northwest Research Station
. Retrieved from
https://apps.fs.usda.gov/fia/datamart/datamart.html
Vaswani, A., Shazeer, N., Parmar, N., Uszkoreit, J., Jones, L., Gomez, A. N., et al. (2017). Attention is all you need.
Advances in Neural In-
formation Processing Systems
,
30
.
https://doi.org/10.48550/arXiv.1706.03762
Virtanen, P., Gommers, R., Oliphant, T. E., Haberland, M., Reddy, T., Cournapeau, D., et al. (2020). SciPy 1.0: Fundamental algorithms for
scientific computing in Python.
Nature Methods
,
17
(3), 261–272.
https://doi.org/10.1038/s41592‐019‐0686‐2
Wells, J. M., Crow, S. E., Meki, M. N., Sierra, C. A., Carlson, K. M., Youkhana, A., et al. (2017). Maximizing soil carbon sequestration: Assessing
procedural barriers to carbon management in cultivated tropical perennial grass systems. In
Recent advances in carbon capture and storage
(pp.
151–172). InTech.
Earth's Future
10.1029/2026EF008563
PETER‐CONTESSE ET AL.
18 of 18
23284277, 2026, 9, Downloaded from https://agupubs.onlinelibrary.wiley.com/doi/10.1029/2026EF008563 by Oregon State University, Wiley Online Library on [19/10/2026]. See the Terms and Conditions (https://onlinelibrary.wiley.com/terms-and-conditions) on Wiley Online Library for rules of use; OA articles are governed by the applicable Creative Commons License
Document Outline