© The Authors, 2025, Published by the Universidad del Zulia*Corresponding author: svillaza@uc.edu.ve
Keywords:
Deep soil moisture content
Soil properties
Articial intelligence
AI-based deep soil moisture prediction to assess past and future consistency of NOAA data
Predicción de la humedad del suelo profundo basada en IA para evaluar la coherencia pasada y
futura de los datos de la NOAA
Previsão da umidade profunda do solo baseada em IA para avaliar a consistência passada e futura
dos dados da NOAA
Guillermo Montilla
1
Sergio Villazana
2
*
Vicente Torres
3
Egilda Pérez
2
Héctor Reverón
1
César Seijas
2
Rev. Fac. Agron. (LUZ). 2025, 42(4): e254257
ISSN 2477-9407
DOI: https://doi.org/10.47280/RevFacAgron(LUZ).v42.n4.IVX
Crop production
Associate editor: Dra. Evelyn Pérez Pérez
University of Zulia, Faculty of Agronomy
Bolivarian Republic of Venezuela
1
Yttrium-Technology Corp., Panama City, Panama.
2
Centre of Image Processing, Faculty of Engineering of the
University of Carabobo, Valencia, Venezuela.
3
National Academy of Research and Development,
Cuernavaca, Mexico. Faculty of Engineering, DICT,
National Autonomous University of Mexico.
Received: 15-07-2025
Accepted: 15-11-2025
Published: 11-12-2025
Abstract
Accurate deep soil moisture modeling is essential for agriculture,
especially in regions with unpredictable precipitation impacting
crop health and yield. Understanding these dynamics is crucial
for assessing drought vulnerability and promoting sustainable
agricultural practices. This study evaluated the consistency of
NOAA climate data over ve years using an AI-developed deep
soil moisture model. The objectives were to assess historical and
future NOAA data reliability and predict soil moisture at 40 cm
and 100 cm depths in Panama’s western region. The CNN-BiLSTM
regression model integrated meteorological and soil property data
(clay, silt, sand) from NOAA and the International Soil Reference
and Information Centre. It transformed data into spatial and
temporal features, with training, validation, and testing sets using
2021 data. The generalization capability of the model was assessed
using data from 2019, 2020, 2022, and 2023, validating predictions
with two preceding and two subsequent years. Results show the
40 cm model achieved MAE, RMSE, MAPE, and R
2
of 0.007112
m
3
.m
-3
, 0.012662 m
3
.m
-3
, 3.55 %, and 0.97, respectively. The 100
cm model recorded 0.011019 m
3
.m
-3
, 0.017334 m
3
.m
-3
, 4.92 %, and
0.95, respectively. The model demonstrated coherence over ve
years, conrming NOAA data consistency.
This scientic publication in digital format is a continuation of the Printed Review: Legal Deposit pp 196802ZU42, ISSN 0378-7818.
Rev. Fac. Agron. (LUZ). 2025, 42(4): e254257 October-December. ISSN 2477-9409.
2-7 |
Resumen
La modelación precisa de la humedad del suelo profundo es
fundamental para la agricultura, especialmente en regiones con
precipitaciones impredecibles que afectan la sanidad y el rendimiento
de los cultivos. Comprender esta dinámica es crucial para evaluar la
vulnerabilidad a la sequía y promover prácticas agrícolas sostenibles.
Este estudio evaluo la consistencia de los datos climáticos de la
NOAA a lo largo de cinco años mediante un modelo de humedad del
suelo profundo desarrollado con inteligencia articial. Sus objetivos
incluyeron la evaluación de la coherencia de los datos históricos
y futuros de la NOAA y la predicción de la humedad del suelo a
profundidades de 40 cm y 100 cm en la región occidental de Panamá.
El modelo de regresión CNN-BiLSTM integró datos meteorológicos
y propiedades del suelo (arcilla, limo y arena) provenientes de la
NOAA y del Centro Internacional de Referencia e Información de
Suelos. El modelo transformó estos datos en características espaciales
y temporales, y los conjuntos de entrenamiento, validación y prueba
se basaron en los datos de 2021. La capacidad de generalización del
modelo se evaluó con datos de 2019, 2020, 2022 y 2023, validando
predicciones con dos años anteriores y dos años posteriores. Los
resultados indicaron que el modelo de 40 cm alcanzó valores de
MAE, RMSE, MAPE y R
2
de 0,007112 m
3
.m
-3
, 0,012662 m
3
.m
-3
, 3,55
% y 0,97, respectivamente. El modelo de 100 cm registró valores de
0,011019 m
3
.m
-3
, 0,017334 m
3
.m
-3
, 4,92 % y 0,95, respectivamente. El
modelo demostró coherencia a lo largo de cinco años, lo que conrma
la consistencia de los datos de la NOAA.
Palabras clave: contenido de humedad profunda del suelo,
propiedades del suelo, inteligencia articial.
Resumo
A modelagem precisa da umidade profunda do solo é essencial
para a agricultura, especialmente em regiões onde a precipitação
imprevisível impacta a saúde e a produtividade das culturas.
Compreender essas dinâmicas é crucial para avaliar a vulnerabilidade
à seca e promover práticas agrícolas sustentáveis. Este estudo
analisou a consistência dos dados climáticos da NOAA ao longo
de cinco anos, utilizando um modelo de umidade profunda do solo
desenvolvido com inteligência articial. Os objetivos foram avaliar
a conabilidade dos dados históricos e futuros da NOAA e prever
a umidade do solo nas profundidades de 40 cm e 100 cm na região
oeste do Panamá. O modelo de regressão CNN-BiLSTM integrou
dados meteorológicos e propriedades do solo (argila, silte e areia)
provenientes da NOAA e do Centro Internacional de Referência e
Informação sobre Solos. Transformou esses dados em características
espaciais e temporais, utilizando conjuntos de treinamento, validação
e teste com dados de 2021. A capacidade de generalização do modelo
foi avaliada com dados de 2019, 2020, 2022 e 2023, validando as
predições com dois anos precedentes e dois anos subsequentes. Os
resultados mostram que o modelo de 40 cm obteve valores de MAE,
RMSE, MAPE e R
2
de 0.007112 m
3
.m
-3
, 0.012662 m
3
.m
-3
, 3,55 %
e 0,97, respectivamente. O modelo de 100 cm registrou valores de
0.011019 m
3
.m
-3
, 0.017334 m
3
.m
-3
, 4,92 % e 0,95, respectivamente. O
modelo demonstrou coerência ao longo de cinco anos, conrmando a
consistência dos dados da NOAA.
Palavras-chave: conteúdo de umidade profunda do solo, propriedades
do solo, inteligência articial.
Introduction
Deep soil moisture content is crucial for agricultural land
management, soil and water conservation, and ecological processes
(Tong et al., 2020). It signicantly inuences plant growth and water
resource management, serving as a key indicator of agricultural
productivity since soil water is the main accessible component of
the hydrological cycle for plants (Munro et al., 1998). Soil texture
has a signicant impact on deep soil water content (SWC), with soil
texture-based models enabling the prediction of SWC proles-critical
for monitoring water depletion and informing sustainable land-use
decisions. Estimating deep soil moisture is essential for sustainable
land use in dry and semi-arid regions, as well as for soil and water
conservation (Wang et al., 2016). Understanding the interaction
between soil texture and moisture content, along with the inuences
of climate and land use, is crucial for eective agricultural resource
management. Pan et al. (2015) proposed a soil moisture diagnostic
equation to estimate the soil moisture at depths of 5, 10, 20, 50, and
100 cm in arid and semiarid regions using only precipitation data. The
authors incorporated the daily mean air temperature, precipitation,
and solar radiation to calculate the parameters of the soil moisture
diagnostic equation.
Much research in the literature addresses articial intelligence-
based modeling of soil moisture content. Han et al. (2021) developed
two data-driven models-an articial neural network (ANN) and a Long
Short-Term Memory (LSTM) model-to predict soil moisture up to six
days ahead at depths of 100, 200, 500, and 1,000 mm. Using weather
data (air temperature, precipitation, vapor pressure, soil temperature,
and relative humidity) and soil characteristics, these models were
tested at the Eagle Lake Observatory in California, USA. The study
concluded that the LSTM model consistently outperformed the ANN
model across all depths. Basir et al. (2024) conducted a subsurface
soil moisture forecasting study in Fort Wayne, Indiana, USA, using
nine years of weather data and historical soil moisture measurements.
They developed two models-Vector Auto Regression and LSTM-to
predict subsurface soil moisture at a depth of 20 cm, utilizing inputs
such as total rainfall, ambient temperature, wind speed, relative
humidity, solar radiation, and volumetric water content at 30 cm depth.
The study concluded that the LSTM model outperformed traditional
statistical approaches in forecasting subsurface soil moisture. Geng
et al. (2024) proposed a method that combines machine-learning
techniques with physical laws to improve soil moisture predictions.
They integrated physical concepts and prior knowledge into the
model design and training process to capture multiscale soil moisture
dynamics, thereby predicting soil moisture between 0 and 7 cm using
the ERA5-Land reanalysis dataset. Filipović et al. (2022) developed
an LSTM model to predict soil moisture three days ahead using the
ERA5 reanalysis dataset developed by Muñoz-Sabater et al. (2021)
and Yr database (https://www.yr.no/nb) data from 28 locations in
Serbia across four dierent soil types. Input variables included daily
temperature, precipitation, vapor pressure decit, and soil moisture
content. Trained on 15- to 60-day historical data in 15-day increments,
the LSTM model outperformed random forest and ARIMA models,
achieving the lowest mean absolute error. Yu et al. (2020) developed
a modied Residual Bidirectional Long Short-Term Memory
(ResBiLSTM) model, integrating Residual Networks (ResNet) and
Bidirectional Long Short-Term Memory (BiLSTM), to predict soil
moisture at 10, 20, 30, 40, and 50 cm depths during maize (Zea
mays L.) growth stages. Using data from seven agrometeorological
This scientic publication in digital format is a continuation of the Printed Review: Legal Deposit pp 196802ZU42, ISSN 0378-7818.
Montilla et al. Rev. Fac. Agron. (LUZ). 2025, 42(4): e254257
3-7 |
Figure 1. (a) Crop dominance map from dataset “GFSAD1KCD
v001”. (b) The NOAA-resolution western agricultural
region, cropland selected by majority vote. (b) illustrates
green cells at the NOAA resolution of 0.25x0.25 degrees
for the western agricultural region of the country.
These cells were obtained by counting the 784 cells of
the “GFSAD1KCD v001” product within each cell of the
NOAA grid and classifying each cell as agricultural or
non-agricultural based on a majority vote.
NOAA variables
Climate variables for this study were obtained from NOAA’s
ds083.3 dataset (National Centers for Environmental Prediction,
2015; https://rda.ucar.edu/datasets/d083003/). This global dataset
integrates atmospheric, oceanic, and land data from satellites, weather
balloons, and ground stations on 0.25° x 0.25° grids, updated every
six hours via the Global Data Assimilation System (GDAS), covering
the period from July 8, 2015, to the present.
NOAA’s Climate Forecast System version 2 model, developed
by National Centers for Environmental Prediction (NCEP), estimates
surface and deep volumetric soil moisture by simulating water balance
through precipitation, evapotranspiration, and runo. NOAA’s deep
soil moisture measurements, derived from North-American Land Data
Assimilation System (NLDAS) and Global Land Data Assimilation
System (GLDAS) models (https://ldas.gsfc.nasa.gov/gldas), integrate
satellite data, ground observations, and climate inputs, achieving an
accuracy of ±0.05 to ±0.08 m
3
.m
-3
. These estimates oer a spatial
resolution of 12-25 km and are updated daily.
NOAA provided the climate data for Panama, and table 1 lists all
the variables used in this study.
stations in Hebei, China (2016-2018), including soil water content
and meteorological variables, the model employed three-day input
data to forecast soil moisture for 1- to 6-day horizons.
Deep soil moisture modeling is vital in Panama due to variable
precipitation, enabling optimized agricultural practices, drought
vulnerability assessment, and proactive mitigation for sustainability.
It also supports hydrological assessments by informing groundwater
recharge, soil erosion risks, and ecosystem health.
As a data-handling strategy, we have designed a temporal-
multivariate framework that integrates ten days of prior and current
meteorological data as model inputs. These include precipitation,
maximum and minimum temperature, dew point, eld capacity,
evaporation rate, humidity, heat ux, surface soil moisture, soil
temperature, and surface texture (clay, silt, and sand fractions at 5
cm depth) obtained from the National Oceanic and Atmospheric
Administration (NOAA) (National Centers for Environmental
Prediction, 2015) and the International Soil Reference and
Information Centre (ISRIC) (Poggio et al., 2021). The Convolutional
Neural Network-Bidirectional Long Short-Term Memory (CNN-
BiLSTM) models have been trained using 2021 data and validated
with independent datasets from 2019-2020 and 2022-2023.
A key feature of this strategy is the explicit incorporation of soil
surface texture variables-strongly correlated with soil moisture-
which enhances both accuracy and robustness. In contrast to prior
studies (Yu et al., 2020; Filipović et al., 2022) that excluded these
parameters and therefore limited model generalization to similar
soil types, our approach extends applicability across diverse
agricultural environments, representing a meaningful methodological
advancement.
In this study, we developed and validated two CNN-BiLSTM
regression models capable of estimating deep volumetric soil
moisture (m
3
.m
-3
) at 40 cm and 100 cm depths, demonstrating their
applicability to Panama’s agricultural western region.
Material and methods
Study site
Panama, located in the intertropical region (7°11’34.10”-
9°39’44.70” N, 77°09’30.58”-83°03’3” W) with an area of 75,517
km², is bordered by the Caribbean Sea, Pacic Ocean, Colombia, and
Costa Rica. Its S-shaped topography features a central mountain range
(Talamanca and San Blas), dividing Atlantic and Pacic slopes, with
1,290 km and 1,700 km of coastline, respectively. It has a tropical
climate with high humidity, temperatures of 20 34.5 °C, and rainfall
of 1,000-3,000 mm, varying by topography and ocean currents. It
includes a rainy season (May-mid-December) and a dry season (mid-
December-late April), with the Caribbean side receiving more rain
(Ministerio de Ambiente de Panamá, 2020).
Selection of the agricultural region for this study
The research focuses on western Panama’s agricultural region,
encompassing Chiriquí, Veraguas, Herrera, Coclé, Los Santos, Bocas
del Toro, and Ngöbe-Buglé Comarca (7°24’-9°07’ N, 79°55’-82°55’
W, 100 to 300 m a.s.l.), notable for fertile soils, diverse crops, and
robust infrastructure. The agricultural map (Figure 1a) was created
using the GFSAD1KCD v001 dataset (Thenkabail et al., 2016), with
a 1 km resolution, identifying nine dominant crop types through
integrated remote sensing data from global irrigated and rainfed
croplands.
Soil property variables
Soil texture variables (clay, silt, sand) at 5 cm depth, dened by the
soil texture triangle, were sourced from ISRIC at a 0.25° resolution,
matching NOAA grid cells, downscaled from 0.0025° (Poggio et al.,
2021). These static data remain constant within each grid cell over
time.
Dataset
The study utilizes 16 input variables from NOAA, including
climatic and soil variables listed in table 1, with surface soil texture
(clay, silt, sand) treated as static. Temperature variables were
converted from Kelvin to Celsius. The two CNN-BiLSTM models
target volumetric soil moisture content at 40 cm and 100 cm depths,
with depth intervals simplied to specic depths to enhance result
interpretability and comparability with other studies. The study
region covers 44 NOAA grid cells (Figure 1b), with NOAA variables
recorded every six hours, yielding 1,460 annual samples per variable
per cell, averaged daily. Two CNN-BiLSTM models, estimating deep
soil moisture at 40 cm and 100 cm depths, were trained on 2021 data
and evaluated for generalization using data from 2019-2020 and
2022-2023.
This scientic publication in digital format is a continuation of the Printed Review: Legal Deposit pp 196802ZU42, ISSN 0378-7818.
Rev. Fac. Agron. (LUZ). 2025, 42(4): e254257 October-December. ISSN 2477-9409.
4-7 |
Table 1. NOAA variables used in this study.
Description Vertical Levels Unit
Wind speed Ground or water Surface m.s
-1
Temperature Ground or water surface K
Maximum temperature Specied height above ground: 2 m K
Minimum temperature Specied height above ground: 2 m K
Dewpoint temperature Specied height above ground: 2 m K
Sensible heat ux Ground or water surface W.m
-2
Specic humidity Specied height above ground: 2 m kg.kg
-1
Relative humidity Specied height above ground: 2 m %
Total precipitation Ground or water surface kg.m
-2
(*)
Potential evaporation rate Ground or water surface W.m
-2
Soil temperature. Layer 1 0 to 0.1 m K
Field capacity Ground surface fraction
VSMC (
**
). Layer 1 0 to 0.1 m m
3
.m
-3
VSMC (
**
). Layer 2 0.1 to 0.4 m m
3
.m
-3
VSMC (
**
). Layer 3 0.4 to 1 m m
3
.m
-3
(*) 1 kg.m
-2
is equivalent to a precipitation of 1 mm. (
**
) VSMC: Volumetric soil moisture
content
Figure 2. Block diagram of the CNN-BiLSTM model for
regression proposed by the authors.
Figure 3. A general diagram of the model proposed by the authors.
CNN-and-bidirectional-LSTM: CNN-BiLSTM model
The CNN-BiLSTM model inspired by Yu et al. (2020)
ResBiLSTM, which used parallel ResNet and bidirectional LSTM
(BiLSTM) branches for soil moisture forecasting, features a simplied
architecture. It includes a parallel structure with a single 1D-CNN
layer and two BiLSTM layers, omitting fully connected layers in each
branch. The branches are concatenated, followed by a fully connected
layer and an output neuron (Figure 2). ReLU activation is used for
the 1D-CNN and fully connected layers, with a linear function for the
output neuron. The CNN-BiLSTM model integrates complementary
1D-CNN and BiLSTM branches for deep soil moisture prediction.
The 1D-CNN branch detects spatial patterns in climate data, soil
texture, and soil moisture, capturing input-target dependencies. The
BiLSTM branch models temporal dynamics, leveraging bidirectional
learning to identify long-term dependencies and nonlinear patterns in
time-series data. The concatenation of both branches’ outputs merges
spatial and temporal insights, feeding into nal layers for a richer data
representation, enhancing generalization and prediction accuracy.
The architectures of the models were implemented in Python 3.10.12
using TensorFlow 2.12.0 within the Google Colab environment.
Dataset preprocessing
The data is organized into an n×m matrix, combining
meteorological time series and static soil texture data (Figure 3).
Columns represent m = 16 features, and rows indicate n = 11 time-
steps (the current day plus 10 prior days), determined by Partial
Autocorrelation Function analysis that identies 10 signicant lags
across 44 study cells (Figure 1b). For a non-leap year, 355 n×m
matrices per cell are paired with soil moisture values at 40 cm and
100 cm depths.
The dataset was randomly shued and split into three subsets:
70 % for training, 15 % for validation, and 15 % for testing.
Training set variables were normalized using the Z-score method,
, ,
=
,
−
, ,
�
, where x
i,j
represents the ith
sample of the jth variable, and x
j,mean
and x
j,std
are the mean and standard
deviation of the jth variable, respectively. The output variable was
similarly normalized. The validation and testing sets were normalized
using the mean and standard deviation of the training set.
The optimization algorithm, loss function, and evaluation
metrics
The Adam optimization algorithm and Mean Squared
Error (MSE) loss function were used for training. The models’
performance metrics are Root Mean Square Error (RMSE),
=
1
Τ
�
−
,
2
=
1
, Mean Absolute Error (MAE),
=
1
Τ
�
−
,
=
1
, Mean Absolute Percentage Error
(MAPE) ,
=
1
Τ
�
−
,
Τ
=
1
, and coecient of
determination
2
=
1
−
�
−
,
2
=
1
�
−
mean
2
=
1
�
, where y
i
is
the true value, y
i,est
is the estimated value, y
mean
is the mean value, and
N is the number of samples.
Training of the CNN-BiLSTM models
Hyperparameter tuning was performed using the Keras-Tuner
library to identify the optimal parameter combinations (O’Malley et
al., 2019). The search space for each hyperparameter is detailed in
table 2. Hyperparameter tuning utilized the “BayesianOptimization”
tuner class with 20 “max_trials” and 3 “executions_per_trial” for
stability. The training spanned 100 epochs with early stopping based
on validation loss after a 10-epoch patience period, retaining the best
weights. Optimal hyperparameter values are presented in table 2.
5-7 |
Table 2. Hyperparameters and their optimal values used during
the tuning of the model.
Hyperparameters Search space
Optimal values
at 40 cm at 100 cm
No. of CNN lters 16, 32, 64 32 32
No. of LSTM neurons 32, 64, 128 64 64
No. of FC neurons 32, 64, 128 64 32
Learning rate 0.0001, 0.001, 0.01 0.001 0.001
Regularization constant (L
2
)
0.0001, 0.001, 0.01 0.001 0.001
Batch size 32, 64, 128, 256, 512 32 32
Models for 40 cm and 100 cm depths were trained using optimal
hyperparameters from the tuning stage, with identical epoch counts
and early stopping strategies. Training data was shued per epoch to
prevent local minima convergence.
Results and discussion
Evaluation of the model at 40 cm for the years 2019, 2020,
2022 and 2023
Figure 4 reveals optimal prediction performance in 2021, with
symmetrical behavior across MAPE, MAE, and RMSE. MAPE values
were 4.9 % (2019), 1.7 % (2021), and 4.7 % (2023), indicating better
future prediction performance. R
2
values were 0.95, 0.99, and 0.97
for 2019, 2021, and 2023, respectively, showing asymmetry. This is
attributed to changing climatic conditions, like extended drought, and
complex environmental interactions aecting soil moisture.
Figure 4. Metrics of the model at 40 cm. (a) R
2
and Mean Absolute
Percentage Error (MAPE). (b) Mean Absolute Error (MAE)
and Root Mean Square Error (RMSE).
Figures 5a to 5e present scatter plots showing the 40 cm model
t to the deep soil moisture data across 44 NOAA grid cells over ve
years 2019 to 2023, respectively. The dashed line represents the ideal
model (R
2
= 1), and the solid line corresponds to a linear t using
the RANSAC algorithm (Fischler & Bolles, 1981). RANSAC-based
equations are shown. High R
2
values (Figure 4a) for the 40 cm model
across four test years beyond the training year demonstrate robust
generalization capability.
Figure 5. Soil moisture scatter plots at 40 cm. (a) 2019, (b) 2020,
(c) 2021, (d) 2022, (e) 2023.
Evaluation of the model at 100 cm for the years 2019, 2020,
2022 and 2023
Figure 6 displays performance metrics for the 100 cm depth
model across all study years, with error graphs (MAPE, MAE,
RMSE) showing symmetrical behavior and better future prediction
performance. Asymmetries in R
2
values around 2021 (Figures 4a and
6a) contrast with symmetrical error graphs, possibly due to extended
drought seasons. The metrics MAPE and 1-R
2
are normalized errors,
with MAPE based on variable value and 1-R
2
on variance. Increased
variance, potentially from prolonged drought (up to 45 days in some
agricultural regions, per Sentinel-1 imagery), may correlate with
higher R
2
.
Figure 6. Metrics of the model at 100 cm. (a) R
2
and Mean Absolute
Percentage Error (MAPE). (b) Mean Absolute Error (MAE)
and Root Mean Square Error (RMSE).
Two sequences of 12 zeros and ones, representing dry (0) and wet
(1) months over 12 months, were analyzed. Sequences with 4 and 5
drought months followed by 8 and 7 rainy months yield variances
of 0.222 and 0.243, respectively, indicating a 9.4 % variance
increase with an additional drought month. R
2
is expected to increase
proportionally, but the observed R
2
increase between 2019 and 2023
(Figures 4a and 6a) is lower, as drought duration interacts with other
factors, such as delayed precipitation, aecting R
2
.
Figure 7 presents scatter plots for the 100 cm model across
study years, showing similar behavior to the 40 cm model but with
underestimation in 2019 and 2020 (RANSAC-t line below ideal)
and slight overestimation in 2022 and 2023 (RANSAC-t line above
ideal). The model exhibits strong generalization, as shown in gures
6 and 7. Increased depth correlates with greater deviation from true
values, likely due to the reduced inuence of climatic conditions on
water movement in deeper soil layers.
Consistency of the NOAA data
The deep soil moisture models developed for 40 cm and 100
cm depths were trained using NOAA climatic and soil data from
2021 and subsequently applied to data from 2019, 2020, 2022, and
2023 to evaluate their interannual consistency. This methodological
framework-centred on training with a single reference year-constitutes
an interannual consistency analysis aimed at assessing the temporal
stability and persistence of climatic patterns around an analytical
baseline.
The 2021 model is not an optimised predictive tool based on
multi-year data aggregation, but rather as a methodological anchor
for quantifying the degree of temporal drift and the stability of the
climatic system in agricultural regions of central-western Panama.
This scientic publication in digital format is a continuation of the Printed Review: Legal Deposit pp 196802ZU42, ISSN 0378-7818.
Rev. Fac. Agron. (LUZ). 2025, 42(4): e254257 October-December. ISSN 2477-9409.
6-7 |
Figure 7. Soil moisture scatter plots at 100 cm. (a) 2019, (b) 2020,
(c) 2021, (d) 2022, (e) 2023.
and soil conditions in central-western Panama. The use of 2021 as a
reference year was not intended to optimize predictive performance
through multi-year aggregation but rather to assess the persistence and
temporal validity of the learned climatic–soil relationships. Overall,
the ndings conrm the structural consistency of NOAA climatic data
and demonstrate that the 2021 reference model eectively quanties
temporal drift, providing a robust methodological tool for evaluating
the stability of climatic–soil interactions in tropical agricultural
systems.
Acknowledgment
We acknowledge the partial support from the “Sistema Nacional
de Investigación” (SNI), SENACYT, Panama, and funding from the
Technical Cooperation Agreement No. ATN/ME-19615-PN between
IICA and BIDLab for this research.
Funding source
“Sistema Nacional de Investigación” (SNI), SENACYT, Panama,
Project No. ATN/ME-19615-PN.
Literature cited
Basir, M., Noel, S., Buckmaster, D., & Ashik-E-Rabbani, M. (2024). Enhancing
subsurface soil moisture forecasting: A long short-term memory
network model using weather data. Agriculture, 14(3), 333. https://doi.
org/10.3390/agriculture14030333
Filipović, N., Brdar, S., Mimić, G., Marko, O., & Crnojević, V. (2022). Regional
soil moisture prediction system based on long short-term memory
network. Biosystems Engineering, 213, 30-38. https://doi.org/10.1016/j.
biosystemseng.2021.11.019
Fischler, M., & Bolles, R. (1981). Random sample consensus: A paradigm
for model tting with applications to image analysis and automated
cartography. Communication of the ACM, 24(6), 381-395. https://doi.
org/10.1145/358669.358692
Geng, Q., Yan, S., Li Q., & Zhang, C. (2024). Enhancing data-driven soil moisture
modeling with physically-guided LSTM networks. Frontiers in Forests
and Global Change, 7, 1-12. https://doi.org/10.3389/gc.2024.1353011
Han, H., Choi, C., Kim, J., Morrison, R., Jung, J., & Kim, H. (2021). Multiple-
Depth soil moisture estimates using articial neural network and long
short-term memory models. Water, 13(8), 2584. https://doi.org/10.3390/
w13182584
Ministerio de Ambiente de Panamá. (2020). Cuarta comunicación nacional
sobre cambio climático de Panamá ante la Convención Marco de las
Naciones Unidas sobre el Cambio Climático (CMNUCC). Ministerio
de Ambiente; Programa de las Naciones Unidas para el Desarrollo.
https://transparencia-climatica.miambiente.gob.pa/wp-content/
uploads/2023/08/4CNCC_2023_L.pdf
Munro, R., Lyons, W. F., Shao, Y., Wood, M., Hood, L. M., & Leslie, L. (1998).
Modeling land surface-atmosphere interactions over the Australian
continent with an emphasis on the role of soil moisture. Environmental
Modeling and Software, 13(3-4), 333-339. https://doi.org/10.1016/s1364-
8152(98)00038-3
Muñoz-Sabater, J., Dutra, E., Agustí-Panareda, A., Albergel, C., Arduini, G.,
Balsamo, G., Boussetta, S., Choulga, M., Harrigan, S., Hersbach, H.,
Martens, B., Miralles, D. G., Piles, M., Rodríguez-Fernández, N., Zsoter,
E., Buontempo, C., & Thépaut, J. (2021). ERA5-Land: A state-of-the-
art global reanalysis dataset for land applications. Earth System Science
Data, 13(9), 4349-4383. https://doi.org/10.5194/essd-13-4349-2021
National Centers for Environmental Prediction. (2015). NCEP GDAS/FNL 0.25
Degree Global Tropospheric Analyses and Forecast Grids. rda.ucar.edu
at the National Center for Atmospheric Research, Computational and
Information Systems Laboratory. https://doi.org/10.5065/D65Q4T4Z
O’Malley, T., Bursztein, E., Long, J., Chollet, F., Jin, H., Invernizzi, L., & others.
(2019). KerasTuner [Software]. GitHub. https://github.com/keras-team/
keras-tuner
Pan, F., Nieswiadomy, M., & Shuan, Q. (2015). Application of a soil moisture
diagnostic equation for estimating root-zone soil moisture in arid and
semi-arid regions. Journal of Hydrology, 524, 296-310. https://doi.
org/10.1016/j.jhydrol.2015.02.044
Poggio, L., de Sousa, L., Batjes, N., Heuvelink, G., Kempen, B., Ribeiro, E., &
Rossiter, D. (2021). SoilGrids 2.0: producing soil information for the
globe with quantied spatial uncertainty. SOIL, 7(1), 217-240. https://doi.
org/10.5194/soil-7-217-2021
The models demonstrated strong consistency with the NOAA data
across all evaluated years, yielding high correlation coecients and
low MAE, MAPE, and RMSE values. These outcomes validated
both the robustness of the trained models and the temporal stability
and reliability of the NOAA dataset over the ve-year period. Each
annual dataset incorporated six-hourly climatic variations, ensuring
sucient temporal resolution to capture the underlying dynamics
of the system. Notably, the study encompassed a broad and diverse
geographical region characterised by high temporal and spatial
variability in meteorological and soil conditions. Despite these
complexities, the NOAA climate data maintained a high degree of
structural consistency, supporting the conclusion that the 2021-based
models eectively quantify interannual stability and detect potential
temporal drift in the climatic and soil moisture dynamics governing
the region.
The validated model demonstrates that a single-year baseline
can accurately assess the degree of climate stability during the study
period, oering a practical framework for ecient, resource-light
agricultural water management. From a theoretical perspective, this
temporal persistence conrms that the underlying climate-soil system
in the region operates with signicant predictability, allowing such a
baseline to serve not only for estimation but also as a robust detector
of future climatic regime shifts.
Conclusions
This study presents an innovative articial intelligence framework
for estimating root zone soil moisture at 40 cm and 100 cm depths
by integrating NOAA’s globally recognized climatic data with soil
information from the International Soil Reference and Information
Centre (ISRIC). The approach combines a one-dimensional
convolutional neural network (1D-CNN) with a bidirectional long
short-term memory (BiLSTM) network, eectively capturing
both spatial and temporal dependencies that govern soil moisture
variability. The models were trained exclusively on 2021 data and
validated against 2019-2020 and 2022-2023 datasets to evaluate their
interannual consistency and temporal stability. Results show that
the 40 cm model achieved MAE, RMSE, MAPE, and R
2
values of
0.007112 m
3
.m
-3
, 0.012662 m
3
.m
-3
, 3.55 %, and 0.97, respectively,
while the 100 cm model recorded 0.011019 m
3
.m
-3
, 0.017334 m
3
.m
-3
,
4.92 %, and 0.95, respectively. These metrics conrm the models’ high
predictive accuracy and minimal temporal bias across diverse climatic
7-7 |
Thenkabail, P., Knox, J., Ozdogan, M., Gumma, M., Congalton, R., Wu, Z., Milesi,
C., Finkral, A., Marshall, M., Mariotto, I., You, S., Giri, C., & Nagler, P.
(2016). NASA Making Earth System Data Records for Use in Research
Environments (MEaSUREs) Global Food Security Support Analysis Data
(GFSAD) Crop Dominance 2010 Global 1 km V001 [Data set]. NASA
EOSDIS Land Processes Distributed Active Archive Center. https://doi.
org/10.5067/MEaSUREs/GFSAD/GFSAD1KCD.001
Tong, Y., Wang, Y., Song, Y., Sun, H., Sun, H., & Xu, Y. (2020). Spatiotemporal
variations in deep soil moisture and its response to land-use shifts
in the Wind-Water erosion crisscross region in the critical zone of the
loess Plateau (2011-2015), China. Catena, 193, 104643. https://doi.
org/10.1016/j.catena.2020.104643
Wang, A., Liu, B., Wang, Z., & Liu, G. (2016). Monitoring and predicting the
soil water content in the deeper soil prole of loess Plateau, China.
International Soil and Water Conservation Research, 4(1), 6-11. https://
doi.org/10.1016/j.iswcr.2016.02.001
Yu, J., Tang, S., Zhangzhong, L., Zheng, W., Huang, C., Wang, L., Wong, A.,
& Xu, L. (2020). A deep learning approach for multi-depth soil water
content prediction in summer maize growth period. IEEE Access, 8,
199097-199110. https://doi.org/10.1109/access.2020.3034984