Environmental and Sustainability Indicators 27 (2025) 100797

Contents lists available at ScienceDirect

Environmental and Sustainability Indicators

journal homepage: www.sciencedirect.com/journal/environmental-and-sustainability-indicators

Land use/cover classification using machine learning algorithms and their

impacts on land surface temperature and soil moisture in the Alawuha
Watershed, Ethiopia
Getachew Bayable a,* , Getie Gebrie b, Tadele Melese a , Alebel Melaku c
a
College of Agriculture and Environmental Science, Bahir Dar University, Bahir Dar, Ethiopia
b
School of Land Administration, Woldiya University, Woldiya, Ethiopia
c
College of Agriculture and Natural Resources, Debre Markos University, Debre Markos, Ethiopia

A R T I C L E I N F O A B S T R A C T

Keywords: Land use/cover (LULC) mapping is vital for natural resource management and environmental monitoring in
Land surface temperature rapidly developing regions such as Ethiopia’s Northern Highlands. This study pioneers the integration of
Land use/cover classification Sentinel-1 Synthetic Aperture Radar and Sentinel-2 Level 2A MultiSpectral Instrument data via Google Earth
Machine learning algorithms
Engine to achieve high-accuracy LULC classification in the Alawuha Watershed, evaluating Classification and
Soil moisture index
Spatiotemporal analysis
Regression Trees (CART), Random Forest (RF), and Support Vector Machines (SVM). It also examined spatio­
temporal variations in land surface temperature (LST) and the soil moisture index (SMI) across LULC types using
Landsat 8. The Radial Basis Function (RBF) SVM outperformed RF and CART, achieving an average overall
accuracy (OA) of 89.6 % and an F1 score of 89.5 % across 2019 and 2024, compared to 88.2 % OA and 88.1 % F1
for RF, and 83.8 % OA and 83.3 % F1 for CART. Spatiotemporal analysis revealed urban expansion, increased
forest cover, and stable farmland, with farmland consistently dominant in the watershed. LST decreased
significantly from 2014 to 2025, with built-up areas showing the highest values at 41.4 ◦ C (2019) and 38.1 ◦ C
(2024) and forests the lowest at 30.4 ◦ C (2019) and 27.8 ◦ C (2024). SMI increased significantly (2014–2025),
with forests recording the highest values at 0.59 (2019) and 0.66 (2024), and built-up and bare lands the lowest.
These findings highlight LULC’s role in regulating microclimates and water balance, offering key insights for
sustainable land-use planning and environmental management.

1. Introduction amplifies these issues by raising LST and lowering soil moisture, which
are critical biophysical parameters that influence agricultural produc­
Land use/cover (LULC) classification is a cornerstone for effective tivity, water availability, and climate resilience (Effati et al., 2021;
natural resource management and environmental monitoring (Z. Zhao Sridhar et al., 2019; Z. Zhao et al., 2024).
et al., 2024). Accurate LULC maps are vital for socioeconomic planning, LST is an essential parameter in surface radiation and water balance
resource optimization, and mitigating urbanization impacts such as (Janani et al., 2024). LST, the radiative skin temperature of the land, is a
biodiversity loss, increased land surface temperatures (LST), and altered key climate variable that indicates how hot or cold the Earth’s surface is
local microclimates (Mumtaz et al., 2020; Nega and Balew, 2022; (Ghaderpour et al., 2024; Z. L. Li et al., 2023), and is crucial for inves­
Stehfest et al., 2019). In rapidly developing regions like the Northern tigating climate (Bayable and Alemu, 2022). It is sensitive to vegetation
Highlands of Ethiopia, where diverse topography and climatic vari­ density and varies significantly with land cover, altitude, and time of
ability coexist, robust LULC mapping becomes even more critical for day (Ghaderpour et al., 2023, 2024). LST is closely correlated with soil
sustainable development (Semaw et al., 2022). Urbanization transforms moisture and is used to monitor droughts, water resources, and hazards
thermal environments, leading to surface urban heat islands that exac­ such as landslides (Khamidov et al., 2023; Loche et al., 2022). Soil
erbate warming, disrupt hydrological processes, and reduce biodiversity moisture, a critical parameter in global hydrological cycles, has broad
(Jabbar et al., 2023). The transition to impervious surfaces further applications in hydrology, meteorology, climatology, and water

* Corresponding author.
E-mail address: bayable.geta@gmail.com (G. Bayable).

https://doi.org/10.1016/j.indic.2025.100797
Received 13 April 2025; Received in revised form 10 July 2025; Accepted 14 July 2025
Available online 15 July 2025
2665-9727/© 2025 The Authors. Published by Elsevier Inc. This is an open access article under the CC BY license (http://creativecommons.org/licenses/by/4.0/).
G. Bayable et al. Environmental and Sustainability Indicators 27 (2025) 100797

resource management (Anderson et al., 2007; Bastiaanssen et al., 2000; Normalized Difference Built-up Index (NDBI) were incorporated to
Dobriyal et al., 2012; Pauwels et al., 2002; Western et al., 2004; Yin enhance classification performance. Furthermore, the impacts of LULC
et al., 2023). change on land surface temperature (LST) and soil moisture index (SMI)
Both LST and SMI reflect underlying LULC patterns, influencing have not been assessed in the Alawuha Watershed, providing additional
surface energy and water balances (Janani et al., 2024; Jia et al., 2021; motivation for this study. Accordingly, the main objectives and contri­
Tronquo et al., 2022). Vegetation moderates surface temperatures bution of this study were to:
through cooling processes like evapotranspiration, while its removal
elevates LST and diminishes soil moisture retention (Goward et al., • Compare the accuracy of three nonparametric machine learning al­
2002; Weng et al., 2004). Croplands, supported by irrigation and canopy gorithms (CART, RF, SVM) for LULC classification using Sentinel-1
cover, typically maintain moderate to high soil moisture, whereas urban SAR and Sentinel-2 MSI data.
and bare lands, characterized by impervious surfaces and minimal • Analyze spatiotemporal trends in LST and SMI between 2014 and
vegetation, show higher LST and lower moisture levels (Peng et al., 2025, linking observed changes to vegetation dynamics and the ef­
2012; Weng et al., 2004; Zhang et al., 2011). Recent studies reinforce fects of Ethiopia’s Green Legacy Initiative.
these relationships such as Shokati et al. (2024) applied Random Forests • Evaluate how different LULC classes influence LST and SMI, identi­
using Sentinel-2, Landsat-8/9, and UAV hyperspectral data, demon­ fying patterns such as urban heat islands and forest-driven moisture
strating higher moisture in vegetated areas compared to bare soils, while retention.
Aliabad and Ghaderpour (2025) used MODIS data to model soil heat • Provide policy-relevant recommendations for forest conservation,
flux, showing lower heat flux in vegetated residential zones than in sustainable land management, and climate-resilient urban planning.
non-vegetated covers. These dynamics are directly relevant to the Ala­
wuha Watershed, where LULC-driven variations in LST and SMI serve as These findings provide valuable insights for sustainable land-use
critical indicators of ecosystem health, agricultural productivity, and planning, enhancing agricultural resilience, and guiding environ­
water availability (Goward et al., 2002; A. Saha et al., 2018). Therefore, mental management strategies in the Ethiopian highlands. To present
investigating how LST and soil moisture vary across LULC classes is these results, this manuscript is structured as follows: Section 1 in­
essential for understanding land–atmosphere interactions, guiding troduces the research background, motivation, and objectives. Section 2
climate adaptation, and promoting sustainable land-use practi­ details the study area, datasets (Sentinel-1, Sentinel-2, SRTM), spectral
ces—especially in Ethiopia’s rain-fed agricultural context. indices, and methods for classification and trend analysis. Section 3
Advances in remote sensing and geospatial technologies provide presents the results, including classification accuracy, spatiotemporal
effective tools for monitoring LULC changes and their impact on LST and LULC changes, and trends in LST and SMI. Section 4 discusses the
soil moisture. Medium-resolution datasets, such as Sentinel-2 and findings in the context of existing literature, their implications, and
Landsat-8, are widely used for their cost-effectiveness and accessibility, study limitations. Section 5 concludes with key insights and policy
enabling detailed LULC mapping and derivation of soil moisture indices recommendations.
(Z. Zhao et al., 2024). Machine learning algorithms increasingly
outperform traditional parametric methods like the Maximum Likeli­ 2. Materials and methods
hood Classifier (MLC) in LULC classification, offering superior efficiency
and accuracy (Gibson et al., 2020; Liu et al., 2022; Pham et al., 2020; 2.1. Study area
Sheykhmousa et al., 2020; Sun et al., 2019). Among these, Classification
and Regression Trees (CART), Random Forest (RF), and Support Vector The study was conducted in the Alawuha Watershed, located in the
Machines (SVM) excel in remote sensing image analysis due to their high North Wello Zone of the Amhara National Regional State, Ethiopia. The
classification accuracy (Bayable et al., 2023; Loukika et al., 2021). Deep watershed covers an area of approximately 780 km2, with elevations
learning techniques like convolutional neural networks (CNNs), while ranging from 1293 m to 3808 m above mean sea level (a.m.s.l.) (Fig. 1).
promising, are computationally intensive and require extensive re­ The landscape is marked by deep gorges and significant erosion, with the
sources and training data (Sewak et al., 2018). eastern boundary forming part of the southwestern block of the Afar
Due to variations in the spectral and textural characteristics of depression (Semaw et al., 2022). The watershed experiences a bimodal
ground features, the performance of machine learning algorithms for rainfall pattern with two rainy seasons. The short rainy season, Belg,
LULC classification varies regionally. This necessitates comparative occurs from March to May, while the long rainy season, Kiremt, lasts
studies to identify the most accurate methods across different land­ from June to September, contributing most of the annual rainfall
scapes. While parametric approaches such as Maximum Likelihood (1000–1100 mm) (Loakes et al., 2018). Mean monthly temperatures
Classification (MLC) dominate LULC studies in Ethiopia (Semaw et al., range between 21 and 25 ◦ C (Semaw et al., 2022). The area’s diverse
2022), the application of nonparametric machine learning algorithms topography significantly influences agricultural practices. Highland re­
using Sentinel-2 Level 2A MultiSpectral Instrument (MSI) and Sentinel-1 gions (Dega, 3200–3700 m) rely on Belg rains for crop production, while
Synthetic Aperture Radar (SAR) data remains underexplored particu­ mid-altitude areas (Woyina Dega, 1500–3200 m) and lowland areas
larly in the Northern Highlands, where persistent cloud cover limits (Kolla, below 1500 m) depend primarily on rainfall from the Meher
optical remote sensing. Although several studies in Ethiopia have season (Mengistie and Kidane, 2016). Agriculture, comprising crop
employed MLC for LULC classification, limited research has evaluated farming and animal husbandry, serves as the primary livelihood for
the performance of nonparametric algorithms in the Northern High­ residents. This sector is predominantly small-scale, labor-intensive, and
lands, including the Alawuha Watershed. Previous studies in this region rain-fed, making it vulnerable to seasonal variability and climatic
(Semaw et al., 2022; Zeleke et al., 2022) primarily relied on Landsat-5 fluctuations.
and Landsat-8 imagery, which offer lower spatial resolution than
Sentinel-2 Level-2A data and are more prone to classification errors due 2.2. Data and processing
to MLC’s sensitivity to outliers. To address these limitations, this study
evaluates three widely used nonparametric classifiers, Classification and This study utilized multiple datasets to evaluate LULC classification
Regression Trees (CART), Random Forest (RF), and Support Vector and its relationship with LST and soil moisture. The datasets employed
Machine (SVM), by integrating Sentinel-2 MSI optical data, Sentinel-1 include Sentinel-2 level 2A MultiSpectral Instrument (MSI) optical im­
SAR data, and Shuttle Radar Topography Mission (SRTM) elevation agery, Sentinel-1 Synthetic Aperture Radar (SAR), Landsat 8 imagery,
data. Environmental indices such as the Normalized Difference Vege­ and Shuttle Radar Topography Mission (SRTM) elevation data (Table 1;
tation Index (NDVI), Normalized Difference Water Index (NDWI), and Fig. 2). For LULC mapping, spectral bands from Sentinel-2 level 2A MSI

2
G. Bayable et al. Environmental and Sustainability Indicators 27 (2025) 100797

Fig. 1. Location Map of Alewuha Watershed, including elevation.

Table 1
Summary of the datasets utilized in this study, including product name, spatial resolution, temporal resolution, and period used.
Product Name Spatial Resolution Temporal Resolution Period Used Accessed From Data set provider

Sentinel-2 Level 2A MSI 10–20 m 5 days 2019, 2024 GEE European Union/ESA/Copernicus
Sentinel-1 SAR 10 m 6–12 days 2019, 2024 GEE European Union/ESA/Copernicus
Landsat OLI/TIRS 30 m (OLI), resampled from 100m to 30m (TIRS) 16 days 2014–2025 GEE USGS
SRTM Digital Elevation 30 m Static (2000) 2000 GEE NASA/USGS

imagery, including blue, green, red, red edge 1, red edge 2, red edge 3, Sentinel-2 imagery, field surveys, high-resolution Google Earth images,
near-infrared, red edge 4, shortwave infrared 1, and shortwave infrared and existing literature. Initially, irregular polygons representing each
2, were used. To address challenges related to cloud cover, Sentinel-1 LULC class were manually delineated using a visual interpretation of
SAR data, which is unaffected by atmospheric conditions, was inte­ Sentinel-2 imagery. These polygons were exported in KML format and
grated with the Sentinel-2 level 2A imagery. The vertical-vertical (VV) validated using Google Earth Pro to ensure the reliability and accuracy
and vertical-horizontal (VH) polarization channels of Sentinel-1 SAR of the delineations. Then, a stratified random sampling approach was
provided additional structural information, thereby improving classifi­ employed within these validated polygons to generate representative
cation accuracy. Additionally, SRTM elevation data with a 30-m reso­ sample points for the six LULC classes. The number of training samples
lution, along with derived slope data, were incorporated to capture for each class was determined based on its proportional area, ensuring
terrain variability. Furthermore, Landsat 8 imagery was used to estimate sufficient representation. As a rule of thumb, at least 50 sample points
LST and the soil moisture index (SMI), enabling the examination of are recommended per class (Thomas et al., 2004). Using a trial-and-error
thermal dynamics and moisture variations across different LULC types. approach, we experimented with various sample sizes and ultimately
Data preprocessing was conducted on the Google Earth Engine (GEE) selected 1280 sample points: urban (180), water (150), forest (200),
platform to standardize datasets and ensure reliability. Cloud and cirrus farmland (crop) (400), shrubland (250), and bareland (100). Of these,
masking was applied to Sentinel-2 imagery using the QA60 band, and 70 % were allocated for training and 30 % for validation across the
median composites were generated for Sentinel-2 and Sentinel-1 data­ machine learning algorithms. This sampling approach ensured a
sets to reduce temporal noise and provide stable LULC representations. balanced representation across all LULC classes, minimized classifica­
Derived indices, including the Normalized Difference Vegetation Index tion bias, and supported robust and accurate model performance
(NDVI), Modified Normalized Difference Water Index (MNDWI), and evaluation.
Normalized Difference Built-up Index (NDBI) (Table 2), were computed
to enhance feature discrimination and facilitate the differentiation of 2.4. Machine learning algorithms
vegetation, water bodies, built-up areas, and other LULC types.
Normalization techniques were also applied across all datasets to miti­ Three non-parametric common machine-learning algorithms, Clas­
gate scale differences, ensuring compatibility and minimizing potential sification and Regression Trees (CART), Random Forest (RF), and Sup­
biases in classification and analysis. port Vector Machines (SVM), were employed in this study, utilizing the
Google Earth Engine (GEE) platform. These machine-learning algo­
2.3. Sample point generation and classification method rithms were selected for their robustness in handling outliers, ability to
effectively classify complex LULC types, and suitability for high-
The LULC map was classified into six categories: forest, shrubland, dimensional datasets (Bayable et al., 2023; Loukika et al., 2021; Z.
built-up areas, barren land, farmland, and water bodies. Sample points Zhao et al., 2024).
were generated through a combination of visual interpretation of Classification and Regression Trees (CART): The CART technique is

3
G. Bayable et al. Environmental and Sustainability Indicators 27 (2025) 100797

Fig. 2. Diagram illustrating the workflow for LULC classification, LST, and SMI analysis in the Alawuha Watershed.

Support Vector Machines (SVM): SVM is a supervised classification

Table 2
algorithm that identifies the optimal hyperplane to separate classes by
Equations for spectral indices.
maximizing the margin between the hyperplane and the nearest data
Index Equation Source points, known as support vectors (VAPNIK & CORTES, 1995). SVM
NDVI (NIR − Red)/(NIR + Red) handles both linear and non-linear data effectively by utilizing various
Tucker (1979)
kernels, including linear, polynomial, and radial basis function (RBF)
NDBI (SWIR − NIR)/(SWIR + NIR)
Z. Zhao et al. (2024) kernels. The cost parameter (C), gamma, and kernel functions are the
MNDWI (Green − SWIR)/(Green + SWIR)
Xu (2006)
most important parameters for selecting support vectors (Loukika et al.,
2021). The RBF kernel was employed in this study due to its superior
performance in capturing non-linear relationships, particularly for LULC
known for handling large datasets more efficiently than traditional classification (Z. Zhao et al., 2024). Hyperparameter tuning was con­
methods like the Maximum Likelihood Classifier (MLC) (Hartfield et al., ducted using grid search optimization to determine the best values for
2011; Z. Zhao et al., 2024). CART is a binary decision tree algorithm that parameters such as cost (C) and gamma. This ensured optimal classifi­
recursively partitions data into smaller subsets based on input features, cation performance across different LULC categories.
optimizing splits using criteria such as Gini impurity or information gain
(Shetty, 2019). Known for its simplicity and interpretability, CART is
2.5. Accuracy assessment
particularly effective for large datasets but can be prone to overfitting,
especially in high-dimensional settings (Loukika et al., 2021). To miti­
The classification results were evaluated using several performance
gate overfitting, hyperparameter tuning was performed by optimizing
evaluation metrics, including Overall Accuracy (OA), Kappa Coefficient
parameters such as the maximum number of nodes and the minimum
(K), User Accuracy (UA), Producer Accuracy (PA), and F1 Score (Nasiri
population of leaf nodes. This approach ensured high classification ac­
et al., 2022, Table 3). OA measured the proportion of correctly classified
curacy while maintaining model simplicity.
samples relative to the total samples, while the Kappa Coefficient
Random Forest (RF): RF, introduced by Breiman (2001), is an
assessed the agreement between observed and predicted classifications
ensemble learning method that combines multiple decision trees to
beyond random chance. UA and PA provided class-specific accuracy
improve classification accuracy and generalization. By using random
evaluations. UA measures the proportion of correctly classified samples
subsets of data and aggregating predictions from individual trees, RF
for each class relative to the total number of samples assigned to that
enhances noise resistance and mitigates overfitting compared to stand­
class, while PA measures the proportion of correctly classified samples
alone decision trees (Belgiu and Drăgu, 2016). RF is particularly effec­
relative to the total number of reference samples for that class. The F1
tive for high-dimensional datasets and eliminates the need for manual
Score, calculated as the harmonic mean of UA and PA, offered a
feature selection, streamlining the classification process. In this study,
balanced measure of classification performance. A confusion matrix,
hyperparameter tuning was performed, focusing on optimizing the bag
generated in GEE, was used to calculate these metrics and provided a
fraction and the number of trees, which improved overall model
detailed breakdown of classification results for each LULC class.
performance.
where r represents the number of rows and columns in the error

4
G. Bayable et al. Environmental and Sustainability Indicators 27 (2025) 100797

Table 3 3. Results
Equations for performance evaluation metrics (Loukika et al., 2021; Nasiri et al.,
2022). 3.1. Performance evaluation of machine learning algorithms
Metric Equation

OA Number of Correctly Classified Samples

The performance of three common machine learning classi­
× 100 fiers—CART, RF, and RBF SVM—was evaluated for LULC classification
Number of Total Samples
using Level-2A processed Sentinel-2 MSI and Sentinel-1 SAR imagery
∑ ∑r
Kc N ri=1 xii − (x × x )
∑r 1=1 i+ +i
N2 − 1=1 (xi+ × x+i ) from 2019 to 2024. In 2019, CART achieved an overall accuracy (OA) of
UA Number of Correctly Classified Samples in each class 80.26 % and a Kappa coefficient (Kc) of 0.75, with the highest F1 scores
Number of Samples Classified to that class
PA Number of Correctly Classified Samples in each class
for water (96.55 %) and urban (88.55 %) classes and the lowest for bare
Number of Samples from Reference Data in each Class land (64.29 %). Similarly, RF recorded an OA of 87.44 % and a Kc of
F1 Score 2 × UA × PA 0.84, with high F1 scores for water (98.82 %) and urban (96.72 %)
UA + PA classes and a lower score for shrubland (74.65 %). Moreover, RBF SVM
outperformed both CART and RF, with an OA of 89 % and a Kc of 0.86,
achieving high F1 scores for water (96.47 %) and urban (98.41 %)
matrix, Xii represents the number of observations in row i and column i,
classes, with shrubland having the lowest score (78.08 %) (Table 5).
Xi ˖ represents the marginal total of row i, X˖i represents the marginal
Likewise, in 2024, CART improved to an OA of 87.26 % and a Kc of 0.84,
total of column i, and N represents the total number of observations.
with the highest F1 scores for water (98.73 %) and urban (94.23 %) and
the lowest for bare land (76.36 %). RF achieved an OA of 88.89 % and a
2.6. Land surface temperature (LST) and soil moisture index (SMI) Kc of 0.86, with high F1 scores for water (97.44 %) and urban (94.00 %),
estimation and a lower score for shrubland (81.38 %). RBF SVM maintained su­
perior performance with an OA of 90.24 % and a Kc of 0.88, with the
Landsat-8 Band 10 from the Thermal Infrared Sensor (TIRS) and highest F1 scores for water (98.73 %) and urban (96.91 %), and the
Bands 4 and 5 from the Operational Land Imager (OLI) were used for this lowest for shrubland (81.69 %) (Table 5). Overall, RBF SVM consistently
analysis (Kim et al., 2022). The digital numbers (DN) from Band 10 were outperformed RF and CART in both years, demonstrating its effective­
first converted to Top-of-Atmosphere (TOA) radiance using multiplica­ ness for LULC classification.
tive and additive rescaling factors (Avdan and Jovanovska, 2016; Roy
and Bari, 2022). TOA radiance was then used to calculate brightness
3.2. Land use/cover (LULC) change
temperature (BT) with thermal conversion constants. NDVI, derived
from Bands 4 and 5, was used to estimate the proportion of vegetation
The spatial coverage of LULC classes from 2019 to 2024 varied across
(Pv), which informed the calculation of surface emissivity (ϵ). Finally,
classifiers (Table 5). According to CART, urban areas increased from
LST was estimated using the brightness temperature and surface emis­
27.41 km2 to 31.81 km2, water bodies decreased from 20.79 km2 to
sivity equation (Table 4; Avdan and Jovanovska, 2016; Roy and Bari,
11.24 km2, forests expanded from 58.05 km2 to 70.74 km2, farmland
2022).
reduced from 389.94 km2 to 377.22 km2, shrubland grew from 243.32
Where Ml is the multiplicative rescaling factor of band 10, Qcal is the
km2 to 277.11 km2, and bare land shrank from 39.38 km2 to 10.77 km2
digital number (DN) for band 10, and Al is the additive rescaling factor
(Table 5; Fig. 3). Similarly, RF estimated urban areas increasing from
of band 10. K1 is the thermal conversion constant one of band 10, and K2
15.90 km2 to 17.1 km2, water bodies slightly decreasing from 14.44 km2
is the thermal conversion constant two of band 10 (Roy and Bari, 2022).
to 13.25 km2, forests increasing from 41.08 km2 to 71.39 km2, farmland
The soil moisture index (SMI) used in this study was derived using
declining from 414.30 km2 to 367.22 km2, shrubland expanding from
LST data obtained from Landsat-8 imagery. The index ranges from 0 to
277.82 km2 to 297.77 km2, and bare land decreasing from 15.34 km2 to
1, where 0 indicates extreme dryness and 1 indicates extreme wetness
12.17 km2 (Table 5; Fig. 3). RBF SVM reported urban areas increasing
(SUTARİYA et al., 2021). The SMI is calculated using the following from 16.44 km2 to 19.89 km2, water bodies decreasing from 13.51 km2
formula (Janani et al., 2024; Mohamed et al., 2020; A. Saha et al., 2018): to 11.36 km2, forests increasing from 42.59 km2 to 78.55 km2, farmland
LSTmax − LST remaining relatively stable (393.00 km2–393.08 km2), shrubland
SMI = decreasing from 281.85 km2 to 257.22 km2, and bare land declining
LSTmax − LSTmin
from 31.51 km2 to 18.8 km2 (Table 5; Fig. 3). Farmland was the domi­
where SMI represents the soil moisture index, LSTmax denotes the nant class, comprising approximately 50 % of the watershed across all
maximum LST, and LSTmin denotes the minimum LST. Furthermore, the classifiers (Table 5; Fig. 3). The findings of this study highlight the
modified Mann-Kendall trend test and linear regression model were variability in classification results of the temporal changes in LULC
applied to analyze trends in LST and SMI from 2014 to 2024. The among different algorithms.
magnitude of change was quantified using Sen’s slope, a robust
nonparametric method. 3.3. Spatiotemporal distribution of land surface temperature and soil
moisture
Table 4
LST estimation equations. The spatial distribution of land surface temperature (LST) and soil
Step Equation moisture index (SMI) in the Alawuha Watershed from 2014 to 2024 is
illustrated in Fig. 4. From a spatial perspective, LST values were
TOA Radiance TOA = Ml × Qcal + Al
Brightness Temp
⎛ ⎞ generally higher in the eastern regions, while the western, central, and
⎜
BT = ⎝ (
K2
) + 1⎟
⎠ − 273.15 southwestern areas exhibited comparatively lower LST values (Fig. 4a).
K1
ln In contrast, SMI was highest in western and central regions and lowest in
TOA
NDVI NDVI = (NIR − Red)/(NIR + Red) eastern areas (Fig. 4b). These spatial patterns suggest an inverse rela­
Proportion of Vegetation
(
(NDVI − NDVImin)
)2 tionship between LST and SMI across the watershed, reflecting distinct
Pv =
(NDVImax − BNDVImin) environmental gradients.
Emissivity ϵ = 0.004 × Pv + 0.986
( ( ) Time-series analysis (Fig. 5) illustrates the temporal variation of LST,
LST BT SMI, and NDVI from January 2014 to June 2025. The spatially averaged
(BT)/ 1 + 0.00115 × *ln(ϵ))
1.4388
LST showed a significant decreasing trend (p < 0.05) according to the

5
G. Bayable et al. Environmental and Sustainability Indicators 27 (2025) 100797

Table 5
Accuracy assessment of LULC, including area in km2 and percentage.
Classifiers Classes Area (km2) Percentage UA (%) PA (%) F1 Score (%) OA (%) Kc

2019 2024 2019 2024 2019 2024 2019 2024 2019 2024 2019 2024 2019 2024

CART Urban 27.41 31.81 3.52 4.08 84.06 90.74 93.55 98 88.55 94.23 80.256 87.26 0.75 0.84
Water 20.79 11.24 2.67 1.44 95.45 100.00 97.67 97.5 96.55 98.73 ​ ​ ​ ​
Forest 58.05 70.74 7.45 9.08 84.75 84.21 75.75 78.68 80.00 81.35 ​ ​ ​ ​
Farmland 389.94 377.22 50.06 48.43 80.83 85.60 83.62 92.24 82.20 88.8 ​ ​ ​ ​
Shrubland 243.32 277.11 31.24 35.58 66.67 84.06 65.75 80.55 66.20 82.27 ​ ​ ​ ​
Bareland 39.38 10.77 5.06 1.38 69.23 84.00 60.00 70 64.29 76.36 ​ ​ ​ ​
RF Urban 15.90 17.1 2.04 2.2 98.33 94 95.16 94 96.72 94 87.44 88.89 0.84 0.86
Water 14.44 13.25 1.85 1.7 100 100 97.67 95 98.82 97.44 ​ ​ ​ ​
Forest 41.08 71.39 5.27 9.17 77.77 88.14 84.84 85.25 81.15 86.67 ​ ​ ​ ​
Farmland 414.30 367.22 53.19 47.15 87.3 88.62 94.82 93.97 90.90 91.21 ​ ​ ​ ​
Shrubland 277.82 297.77 35.67 38.23 76.81 80.82 72.6 81.94 74.65 81.38 ​ ​ ​ ​
Bareland 15.34 12.17 1.97 1.56 100 88.46 70 76.67 82.35 82.14 ​ ​ ​ ​
RBF SVM Urban 16.44 19.89 2.11 2.55 96.875 100 100 94 98.41 96.91 88.97 90.24 0.86 0.88
Water 13.51 11.36 1.73 1.46 97.619 100 95.348 97.5 96.47 98.73 ​ ​ ​ ​
Forest 42.59 78.55 5.47 10.08 87.5 85.71 84.8484 88.52 86.15 87.1 ​ ​ ​ ​
Farmland 393.00 393.08 50.46 50.47 89.256 90.24 93.1 95.69 91.14 92.89 ​ ​ ​ ​
Shrubland 281.85 257.22 36.19 33.02 78.082 82.86 78.08 80.55 78.08 81.69 ​ ​ ​ ​
Bareland 31.51 18.8 4.05 2.41 88.46 88.89 76.666 80 82.14 84.21 ​ ​ ​ ​

Fig. 3. LULC maps for 2019 and 2024 generated using three machine learning classifiers: Classification and Regression Trees (CART), Radial Basis Function Support
Vector Machines (RBF SVM), and Random Forest (RF).

modified Mann–Kendall test (Fig. 5b). In contrast, SMI exhibited a sig­ LULC classes, the RBF SVM classification result was used, as it achieved
nificant increase over the same period (p < 0.05) (Fig. 5c). Similarly, the highest accuracy compared to RF and CART. The variations of LST
NDVI increased significantly (2014–2025), indicating enhanced vege­ within each LULC class are shown in Fig. 7a, with higher LST observed in
tation cover in the watershed (Fig. 5a). 2019 compared to 2024 across all classes. The maximum LSTs were
Spatial trends in LST and SMI, assessed using Sen’s Slope and the recorded in built-up areas, with values of 41.4 ◦ C in 2019 and 38.1 ◦ C in
modified Mann-Kendall trend test, are shown in Fig. 6. LST increased in 2024, whereas the minimum LSTs were observed in forest areas, with
the northwestern and central regions, particularly around urban and values of 30.4 ◦ C in 2019 and 27.8 ◦ C in 2024 (Fig. 7a). Conversely, SMI
farmland areas, but decreased significantly in western forested and increased from 2019 to 2024 across all classes within the watershed
southeastern irrigated farm areas (Fig. 6a; Fig. 6c; Fig. 3). In contrast, (Fig. 7b). The highest SMI values were found in forest areas, at 0.59 in
SMI trends exhibited the opposite pattern, with decreases in northern, 2019 and 0.66 in 2024, while built-up areas (0.30 in 2019 and 0.39 in
northwestern, western, and central urban and farmland areas and sig­ 2024) and bare land areas (0.32 in 2019 and 0.40 in 2024) exhibited the
nificant increases in western forested and southeastern irrigated areas. lowest SMI values (Fig. 7b).
Overall, the spatial pattern of SMI Sen’s Slope was inversely related to
that of LST (Fig. 6). 4. Discussion

4.1. Machine learning performance and LULC change

3.4. Land surface temperature and soil moisture across LULC classes
The Radial Basis Function Support Vector Machine (RBF SVM)
To investigate the variation of LST and soil moisture across different

6
G. Bayable et al. Environmental and Sustainability Indicators 27 (2025) 100797

Fig. 4. Spatial distribution of (a) land surface temperature (LST) and (b) soil moisture index (SMI) derived from Landsat-8 imagery for the period 2014–2024. The
selected months were chosen based on relatively low cloud cover compared to other months of the year.

outperformed other classifiers in the Alawuha Watershed, achieving an Jodzani et al. (2019) reported SVM’s effectiveness in urban LULC clas­
average Overall Accuracy (OA) of 89.6 % and an F1 score of 89.5 %, sification compared to RF. Mengesha et al. (2024) reported that SVM
compared to Random Forest (RF; OA: 88.2 %, F1: 88.1 %) and Classi­ outperformed RF and CART in Ethiopian agricultural mapping using
fication and Regression Trees (CART; OA: 83.8 %, F1: 83.3 %). RF integrated Sentinel-1 SAR and Sentinel-2 MSI data, while Woldemariam
ranked second, while CART had the lowest performance among the et al. (2022) noted SVM’s higher accuracy over RF in the Lake Haramaya
evaluated classifiers. These findings are consistent with prior studies Watershed. Shao and Lunetta (2012) also noted SVM’s superior accu­
demonstrating SVM’s higher accuracy in LULC mapping (Jodzani et al., racy over CART for LULC classification with limited training data using
2019; Mengesha et al., 2024; Woldemariam et al., 2022). For instance, MODIS imagery. The superior performance of the RBF SVM in this study

7
G. Bayable et al. Environmental and Sustainability Indicators 27 (2025) 100797

Fig. 5. Time series of: (a) NDVI, (b) LST, and (c) SMI from January 2014 to June 2025, with significant trends (P < 0.05, Modified Mann-Kendall test).

Fig. 6. Maps display: (a) Sen slopes for LST, (b) Sen slopes for SMI, (c) Sen slopes and statistical significance (P < 0.05) of LST using the Modified Mann-Kendall test,
and (d) Sen slopes and statistical significance (P < 0.05) of SMI using the Modified Mann-Kendall test, from 2014 to 2024. Positive Sen slopes indicate increasing
trends, while negative slopes indicate decreasing trends.

is likely attributable to its non-linear kernel, which effectively captures cover patterns due to its non-linear modeling and resistance to over­
complex LULC patterns, and its strong generalization ability, which fitting in smaller datasets. These findings underscore the importance of
mitigates the impact of limited training sample sizes (Shao and Lunetta, evaluating multiple classifiers to select the most effective method,
2012). However, some studies report contrasting results. For instance, ensuring precise LULC mapping and reliable trend analysis for changes
Bayable et al. (2023) reported that RF outperformed SVM and CART for such as urban expansion and afforestation.
detecting Eichhornia crassipes in Lake Tana, while Loukika et al. (2021)
and Shetty et al. (2021) revealed higher RF accuracy in LULC mapping.
These discrepancies may arise from differences in data characteristics, 4.2. Spatiotemporal variability of land surface temperature, soil moisture,
such as feature complexity or training sample size, highlighting the and vegetation cover
context-specific nature of classifier performance.
Analysis of LULC changes in the Alawuha Watershed revealed urban Land Surface Temperature (LST) in the Alawuha Watershed exhibi­
expansion from 16.44 km2 to 19.89 km2 and forest cover growth from ted a significant overall decrease (p < 0.05, Modified Mann-Kendall test;
42.59 km2 to 78.55 km2 between 2019 and 2024 (Table 5; Fig. 3). These Fig. 5b) from 2014 to 2025, with spatial patterns likely attributed to
changes align with Eshetie et al. (2025) and Zeleke et al. (2022), who elevation and vegetation cover. Western and southwestern regions,
linked similar patterns to population growth and Ethiopia’s Green characterized by higher elevations and dense forests (Figs. 1 and 3),
Legacy Initiative (Jalleta, 2021). Farmland, occupying approximately maintained lower LSTs, which could be due to the cooling effects of
50 % of the watershed (Table 5; Fig. 3), underscores its socioeconomic evapotranspiration and the inverse relationship between elevation and
importance, consistent with Mengistie and Kidane (2016). The vari­ temperature (Li et al., 2015; He et al., 2019). In contrast, eastern re­
ability in LULC estimates across classifiers emphasizes the need for gions, dominated by cropland, bare land, and urban areas (Fig. 3), dis­
robust algorithms like RBF SVM, which excels in capturing complex land played higher LSTs, possibly attributed to low vegetation cover and
lower elevations that enhance heat absorption and emission (Sharma

8
G. Bayable et al. Environmental and Sustainability Indicators 27 (2025) 100797

2025. The analysis revealed a strong negative correlation at lag 0 (r =

− 0.55, p < 0.05), indicating that higher NDVI indicative of denser
vegetation corresponds to lower LST within the same month, consistent
with the cooling effects of evapotranspiration and canopy cover (Li
et al., 2015; Gohain et al., 2021). A significant one-month phase delay
was also observed, where NDVI at time (t) correlated negatively with
LST at time (t+1) (r = − 0.46, p < 0.05), suggesting that vegetation
growth in one month contributes to reduced LST in the following month.
This delay likely results from ecological processes such as increased
evapotranspiration and albedo effects following vegetation develop­
ment. Similar phase-shift dynamics have been reported in other regions.
Ghaderpour et al. (2023) observed in Italy that surface temperature
peaked shortly after NDVI reached its maximum and that NDVI reached
its minimum just after the annual temperature minimum, demonstrating
a lagged response between vegetation and temperature cycles. Likewise,
Dastour et al. (2022) identified a 2–4-week lag in NDVI response to
climate variables across the cold-climate Athabasca River Basin in
Canada. These findings align with our results in Alawuha and highlight
that vegetation-induced cooling effects are delayed by several weeks.
This temporal lag can be strategically leveraged in reforestation and
greening interventions such as Ethiopia’s Green Legacy Initiative to
maximize cooling impacts in urban hotspots, where LST reached 41.4 ◦ C
in 2019 and 38.1 ◦ C in 2024 (Fig. 7a).

4.3. Land surface temperature and soil moisture across LULC classes

LULC changes in the Alawuha Watershed, notably urban expansion

Fig. 7. Variations in LST (a) and SMI (b) across different LULC classes. from 16.44 km2 to 19.89 km2 and forest cover increase from 42.59 km2
to 78.55 km2 between 2019 and 2024 (Fig. 3, Table 5), significantly
and Vashishtha, 2023). The overall decline in LST could be linked to influence LST and SMI. LST varies across different LULC types (Das and
increased vegetation cover, as suggested by rising NDVI trends (Fig. 5a), Angadi, 2020). Built-up areas recorded the highest LSTs (41.4 ◦ C in
likely driven by Ethiopia’s Green Legacy Initiative launched in 2019 2019, 38.1 ◦ C in 2024), reflecting urban heat island effects (Jabbar et al.,
(Jalleta, 2021; Bayable et al., 2024). However, localized LST increases 2023; Laskar et al., 2024), while forests exhibited the lowest (30.4 ◦ C in
were observed in urban and farmland areas (Fig. 6), where reduced 2019, 27.8 ◦ C in 2024), likely due to canopy cover and evapotranspi­
vegetation could promote heat emission (Sharma and Vashishtha, ration (Li et al., 2015; Gohain et al., 2021). Impervious surfaces in
2023), whereas forested and irrigated farmland showed a significant built-up areas, such as concrete and asphalt, absorb and retain heat, as
declining LST trend (Fig. 6), possibly due to enhanced evapotranspira­ noted by Laskar et al. (2024), whereas vegetated areas facilitate cooling
tion and surface cooling (Berhe et al., 2022). (Li et al., 2015). Similarly, bare land and cropland showed higher LSTs
Complementing LST patterns, the Soil Moisture Index (SMI) exhibi­ than forests and water bodies, consistent with trends observed in African
ted an inverse spatial distribution to LST (Fig. 4). Western regions with and South Asian contexts (Nega and Balew, 2022; Saha et al., 2024).
dense forest cover displayed higher SMI, which could be attributed to Soil moisture also varied significantly across LULC classes (Janani
extensive root systems that enhance moisture retention (Goward et al., et al., 2024), with forests displaying the highest SMI (0.59 in 2019, 0.66
2002; Weng et al., 2004). Similarly, Shokati et al. (2024) reported in 2024), likely attributed to their capacity to retain water through
higher SMI in vegetated areas compared to bare soil. Eastern areas, extensive root systems (Zhao et al., 2017; Shokati et al., 2024). In
primarily cropland and bare land (Fig. 3), showed lower SMI, likely due contrast, built-up and bare land areas exhibited the lowest SMI, possibly
to reduced vegetation and rapid surface heating that limits moisture due to impervious surfaces and reduced vegetation that impede infil­
retention (Sharma and Vashishtha, 2023). The watershed’s average SMI tration (Guo et al., 2020). These variations could be influenced by soil
increased significantly from 2014 to 2025 (p < 0.05, Modified texture and LULC-specific properties affecting moisture retention. These
Mann-Kendall test), possibly driven by expanded forest cover and rising LULC-driven LST and SMI trends underscore the profound influence of
NDVI (Figs. 5a and 3). However, SMI trends varied spatially: forested land use changes on environmental conditions, suggesting that sus­
and irrigated farmland showed significant increases, whereas urban tainable urban planning, such as integrating green infrastructure, may
areas and non-irrigated farmland in western and central regions expe­ be critical to mitigate urban heat islands, enhance water retention, and
rienced declines (Fig. 6). Enhanced irrigation helps mitigate support sustainable water resources in the Alawuha Watershed.
soil-moisture stress—studies in Ethiopia report significantly improved
soil moisture and crop water productivity under small-scale irrigation 4.4. Limitations of the study
schemes (Berhe et al., 2022). The SMI Sen’s Slope map (Fig. 6b) reveals
an inverse relationship with the LST Sen’s Slope map (Fig. 6a): areas This study has several limitations that may affect the reliability of its
with rising SMI likely driven by vegetation and irrigation correlate with results. The 12-year time series (2014–2025) used for NDVI, LST, and
decreasing LST, while areas with declining SMI—possibly due to urban SMI, derived from Landsat 8, may limit the robustness of trend analyses
expansion and intensified farming—show increasing LST. These results with the Modified Mann-Kendall test, as two to three decades are typi­
underscore that vegetation and irrigation are pivotal for microclimate cally recommended for detecting long-term trends in environmental
regulation in the Alawuha Watershed, highlighting the importance of variables (Ghaderpour et al., 2024). Cloud contamination during the
sustainable land management for ecological stability. rainy season (May–September) introduces uncertainties in Landsat 8
To investigate potential phase delays between vegetation and LST and Sentinel-2 data, potentially affecting LULC, LST, and SMI estimates.
cycles in the Alawuha Watershed, a cross-correlation analysis was con­ Additionally, the training sample collection for LULC classification
ducted using monthly NDVI and LST data from January 2014 to June relied on visual interpretation, supported by field surveys and Google

9
G. Bayable et al. Environmental and Sustainability Indicators 27 (2025) 100797

Earth imagery, which could introduce subjectivity and lead to delinea­ soil moisture sensors and remote sensing should be supported to opti­
tion errors or misclassification. Although stratified random sampling mize irrigation. Training programs and expanded extension services are
was used to enhance class representation, we acknowledge moderate essential to reach smallholder farmers and encourage the uptake of
confidence in these samples due to potential delineation errors, partic­ climate-smart agriculture. Similarly, efforts in other developing coun­
ularly for smaller LULC classes with limited intra-class variability. The tries should prioritize building the capacity of smallholder farmers,
study used common machine learning algorithms to classify LULC; to providing access to affordable technologies and financial incentives to
improve accuracy, more sophisticated methods such as deep learning support the adoption of sustainable practices.
models should be used. Additionally, the medium-resolution Sentinel-2 Leveraging advanced technologies for land use and environ­
imagery (10–20 m) may limit the detection of small or fragmented mental monitoring: The study’s application of machine learning al­
land-use types. To address these limitations, future studies could gorithms to satellite imagery underscores their potential for
incorporate longer time series (e.g., ≥20 years), enhance ground-truth environmental monitoring. Governments and environmental agencies
validation, and use high-resolution imagery (<10 m) to improve LULC should invest in machine learning-driven platforms to enable up-to-date,
classification and spatiotemporal analyses of LST and SMI in the Ala­ near-real-time tracking of land use and environmental changes. Inte­
wuha Watershed. grating machine learning with remote sensing imagery provides
spatially explicit insights that enhance decision-making. Establishing
4.5. Policy implications regional centers of excellence in geospatial analysis, supported by in­
ternational partnerships, could strengthen local capacity for environ­
The findings of this study provide critical insights for policymakers mental monitoring and informed decision-making.
and stakeholders involved in sustainable land management and climate Implementing integrated cross-sectoral policies: The interplay of
resilience in the Northern Highlands of Ethiopia and other similar re­ LULC changes, LST, and SMI demands coordinated strategies across
gions. Based on the results of the study, the following policy directions agriculture, urban development, and environmental sectors, requiring
are derived. integrated land-use planning aligned with SDGs 11 (Sustainable Cities
Enhancing forest conservation and reforestation programs: and Communities) and 15 (Life on Land). Furthermore, establishing
Forests exhibited the lowest LST and highest soil moisture index (SMI), multi-stakeholder platforms, including inter-ministerial task forces,
underscoring their role in regulating microclimates and maintaining government agencies, local communities, private sector actors, and in­
water balance. To boost climate resilience, policymakers should priori­ ternational organizations, can foster collaboration to co-design policies
tize forest protection and reforestation. Ethiopia’s National Adaptation that balance economic development and environmental sustainability
Plan and Climate Resilient Green Economy (FDRE, 2011, 2019) high­ while addressing the challenges of rapid urbanization and environ­
light afforestation as a key adaptation strategy, reinforced by this mental change.
study’s evidence. The IPCC’s (2023) push for nature-based solutions
further supports integrating forest conservation into national climate 5. Conclusion
action plans. Policies should tackle deforestation drivers, like unsus­
tainable agricultural expansion, through community-based manage­ In Ethiopia, rapid land-use changes and climatic variability chal­
ment, incentivized by payment for ecosystem services (PES) schemes. lenge sustainable natural resource management. This study highlights
Buffer zones around forests could also curb agricultural and urban the potential of machine learning algorithms and remote sensing tech­
encroachment. Stakeholders, local communities, NGOs, and govern­ nology to inform effective environmental strategies. It demonstrated the
ment, should scale up initiatives like participatory forest management efficacy of machine learning algorithms for LULC classification in the
(PFM), which has proven effective in restoring degraded lands. Remote Alawuha Watershed, Northern Highlands of Ethiopia, with RBF SVM
sensing with GEE can monitor forest cover changes in near-real-time, outperforming RF and CART. LULC analysis revealed significant shifts,
ensuring compliance with reforestation targets like the Green Legacy including urban expansion, increased forest cover, stable farmland, and
Initiative. Policymakers should develop targeted policies that promote farmland’s dominant coverage in the watershed. From 2014 to 2025,
the restoration of degraded land and the expansion of green cover in LST decreased significantly, with built-up areas recording the highest
urban and rural areas. Ethiopia’s Green Legacy Initiative could be values and forests the lowest, while SMI increased significantly, with
expanded to include the restoration of degraded lands throughout the forests showing the highest soil retention and built-up and bare lands the
country and the promotion of forestry and agroforestry systems. lowest. These findings emphasize vegetation’s unmatched ability to cool
Promoting sustainable urban planning: Built-up areas recorded microclimates and retain moisture, starkly contrasted by the heat and
the highest LST, necessitating urban planning with green infrastructure. dryness of urban and bare landscapes. They advocate for scaling up
Policies should prioritize tree planting, green roofs, and urban parks to reforestation efforts, such as Ethiopia’s Green Legacy Initiative, and
mitigate heat island effects, aligning with UN-HABITAT’s (2024) vision embedding green infrastructure in urban planning to bolster climate
of cities as climate solutions rather than seeing cities as problems. Local resilience, offering a vital roadmap for sustainable land-use policies in
governments could enforce zoning regulations mandating green spaces regions facing similar environmental pressures.
in new urban developments, leveraging Ethiopia’s Urban Development
Policy (FDRE, 2019) for compliance. Such measures reduce LST while CRediT authorship contribution statement
enhancing livability, public health, and well-being, supported by public
awareness campaigns to boost community involvement. Getachew Bayable: Writing – review & editing, Writing – original
Adopting sustainable agricultural practices to balance produc­ draft, Visualization, Validation, Software, Methodology, Investigation,
tivity and ecological integrity: Agricultural land dominated the study Formal analysis, Data curation, Conceptualization. Getie Gebrie:
area, underscoring its socioeconomic importance. Climate-smart prac­ Writing – review & editing, Methodology, Investigation, Formal anal­
tices, such as agroforestry and conservation agriculture, should be pro­ ysis, Data curation. Tadele Melese: Writing – review & editing, Visu­
moted to enhance soil health, improve water retention, and boost alization, Supervision, Methodology, Formal analysis, Data curation,
agricultural productivity, in line with recommendations from the Food Conceptualization. Alebel Melaku: Writing – review & editing, Meth­
and Agriculture Organization (FAO, 2022). The FAO (2022) has advo­ odology, Investigation, Formal analysis, Data curation,
cated for these practices as a pathway to achieving sustainable food Conceptualization.
systems. Ethiopia’s Climate-Resilient Green Economy Strategy could
incentivize the adoption of such techniques to balance food security
with ecological integrity. Additionally, precision agriculture tools like

10
G. Bayable et al. Environmental and Sustainability Indicators 27 (2025) 100797

Consent of participation FAO, 2022. The state of food and agriculture 2022; leveraging automation in agriculture
for transforming agrifood systems. In: The State of Food and Agriculture 2022.
https://doi.org/10.4060/cb9479en.
Not applicable. FDRE, 2011. Ethiopia’s Climate-Resilient Green Economy Green Economy Strategy.
FDRE, 2019. Ethiopia’s Climate Resilient Green Economy Strategy. National Adaptation
Plan.
Ethical approval Ghaderpour, E., Mazzanti, P., Bozzano, F., Scarascia Mugnozza, G., 2024. Trend analysis
of MODIS land surface temperature and land cover in central Italy. Land 13 (6),
Not applicable. 1–15. https://doi.org/10.3390/land13060796.
Ghaderpour, E., Mazzanti, P., Mugnozza, G.S., Bozzano, F., 2023. Coherency and phase
delay analyses between land cover and climate across Italy via the least-squares
Funding wavelet software. Int. J. Appl. Earth Obs. Geoinf. 118 (February), 103241. https://
doi.org/10.1016/j.jag.2023.103241.
Gibson, R., Danaher, T., Hehir, W., Collins, L., 2020. A remote sensing approach to
No funding was received for this study. mapping fire severity in south-eastern Australia using sentinel 2 and random forest.
Rem. Sens. Environ. 240 (December 2019), 111702. https://doi.org/10.1016/j.
rse.2020.111702.
Declaration of competing interest Gohain, K.J., Mohammad, P., Goswami, A., 2021. Assessing the impact of land use land
cover changes on land surface temperature over pune city, India. Quat. Int. 575–576,
259–269. https://doi.org/10.1016/j.quaint.2020.04.052.
The authors declare that they have no financial interests or personal Goward, S.N., Xue, Y., Czajkowski, K.P., 2002. Evaluating land surface moisture
relationships that could have appeared to influence the work reported in conditions from the remotely sensed temperature/vegetation index measurements:
this paper. an exploration with the simplified simple biosphere model. Rem. Sens. Environ. 79
(2–3), 225–242. https://doi.org/10.1016/S0034-4257(01)00275-9.
Guo, X., Fu, Q., Hang, Y., Lu, H., Gao, F., Si, J., 2020. Spatial variability of soil moisture
Data availability in relation to land use types and topographic features on hillslopes in the Black soil
(mollisols) area of northeast China. Sustainability 12 (9), 8–10. https://doi.org/
10.3390/SU12093552.
Data will be made available on request.
Hartfield, K.A., Landau, K.I., van Leeuwen, W.J.D., 2011. Fusion of high resolution aerial
multispectral and lidar data: land cover in the context of urban mosquito habitat.
References Remote Sens. 3 (11), 2364–2383. https://doi.org/10.3390/rs3112364.
He, J., Zhao, W., Li, A., Wen, F., Yu, D., 2019. The impact of the terrain effect on land
surface temperature variation based on Landsat-8 observations in mountainous
Anderson, M.C., Norman, J.M., Mecikalski, J.R., Otkin, J.A., Kustas, W.P., 2007.
areas. Int. J. Rem. Sens. 40 (5–6), 1808–1827. https://doi.org/10.1080/
A climatological study of evapotranspiration and moisture stress across the
01431161.2018.1466082.
Continental United States based on thermal remote sensing: 2. Surface moisture
IPCC, 2023. Climate Change 2023: Synthesis Report; Summary for Policymakers (H. Lee
climatology. J. Geophys. Res. Atmos. 112 (11). https://doi.org/10.1029/
And).
2006JD007507.
Jabbar, H.K., Hamoodi, M.N., Al-Hameedawi, A.N., 2023. Urban heat islands: a review of
Aliabad, F.A., Ghaderpour, E., 2025. Modeling soil heat flux from MODIS products for
contributing factors, effects and data. IOP Conf. Ser. Earth Environ. Sci. 1129 (1),
arid regions. Ecol. Inform. 86 (January), 103005. https://doi.org/10.1016/j.
1–9. https://doi.org/10.1088/1755-1315/1129/1/012038.
ecoinf.2025.103005.
Jalleta, A.K., 2021. The legal protection of forests: ethiopian green legacy vs.
Avdan, U., Jovanovska, G., 2016. Algorithm for automated mapping of land surface
international environmental regimes. Beijing Law Rev. 12 (3), 725–749. https://doi.
temperature using LANDSAT 8 satellite data. J. Sens. 2016. https://doi.org/
org/10.4236/blr.2021.123038.
10.1155/2016/1480307.
Janani, N., Kannan, B., Nagarajan, K., Thiyagarajan, G., Duraisamy, M.R., 2024. Soil
Bastiaanssen, W.G.M., Molden, D.J., Makin, I.W., 2000. Remote sensing for irrigated
moisture mapping for different land-use patterns of lower bhavani river basin using
agriculture: examples from research and possible applications. Agric. Water Manag.
vegetative index and land surface temperature. Environ. Dev. Sustain. 26 (2),
46 (2), 137–155. https://doi.org/10.1016/S0378-3774(00)00080-9.
4533–4549. https://doi.org/10.1007/s10668-022-02896-1.
Bayable, G., Alemu, G., 2022. Spatiotemporal variability of land surface temperature in
Jia, A., Ma, H., Liang, S., Wang, D., 2021. Cloudy-sky land surface temperature from
north-western Ethiopia. Environ. Sci. Pollut. Control Ser. 29 (2), 2629–2641.
VIIRS and MODIS satellite data using a surface energy balance-based method. Rem.
https://doi.org/10.1007/s11356-021-15763-9.
Sens. Environ. 263 (September 2020), 112566. https://doi.org/10.1016/j.
Bayable, G., Cai, J., Mekonnen, M., Legesse, S.A., Ishikawa, K., Imamura, H.,
rse.2021.112566.
Kuwahara, V.S., 2023. Detection of water hyacinth (Eichhornia crassipes) in Lake
Jodzani, S.E., Johnson, B.A., Chen, D., 2019. Comparing deep neural networks ,
Tana, Ethiopia, using machine learning algorithms. Water (Switzerland) 15 (5).
ensemble classifiers , and support vector machine algorithms. Remote Sens. 11 (1),
https://doi.org/10.3390/w15050880.
1–24.
Bayable, G., Cai, J., Mekonnen, M., Legesse, S.A., Ishikawa, K., Sato, S., Kuwahara, V.S.,
Khamidov, M., Ishchanov, J., Hamidov, A., Shermatov, E., Gafurov, Z., 2023. Impact of
2024. Spatiotemporal variability of Lake surface water temperature and water
soil surface temperature on changes in the groundwater level. Water (Switzerland)
quality parameters and its interrelationship with water hyacinth biomass in Lake
15 (21), 1–16. https://doi.org/10.3390/w15213865.
Tana, Ethiopia. Environ. Sci. Pollut. Control Ser. 31 (33), 45929–45953. https://doi.
Kim, M., Kim, D., Kim, G., 2022. Examining the relationship between land use/land
org/10.1007/s11356-024-34212-x.
cover (LULC) and land surface temperature (LST) using explainable artificial
Belgiu, M., Drăgu, L., 2016. Random forest in remote sensing: a review of applications
intelligence (XAI) models: a case study of Seoul, South Korea. Int. J. Environ. Res.
and future directions. ISPRS J. Photogrammetry Remote Sens. 114, 24–31. https://
Publ. Health 19 (23). https://doi.org/10.3390/ijerph192315926.
doi.org/10.1016/j.isprsjprs.2016.01.011.
Laskar, S., Mishra, P., Behera, B., 2024. Impact of land use dynamics on land surface
Berhe, G.T., Baartman, J.E.M., Veldwisch, G.J., Grum, B., Ritsema, C.J., 2022. Irrigation
temperature: a multi-decadal analysis in Indian sundarbans region. Environ. Dev.
development and management practices in Ethiopia: a systematic review on existing
Sustain., 0123456789 https://doi.org/10.1007/s10668-024-05710-2.
problems, sustainability issues and future directions. Agric. Water Manag. 274
Li, Y., Zhao, M., Motesharrei, S., Mu, Q., Kalnay, E., Li, S., 2015. Local cooling and
(March), 107959. https://doi.org/10.1016/j.agwat.2022.107959.
warming effects of forests based on satellite observations. Nat. Commun. 6. https://
Breiman, L., 2001. Random forests. In: Schapire, R.E. (Ed.), Machine Learning, vol. 45.
doi.org/10.1038/ncomms7603.
Kluwer Academic Publishers. https://doi.org/10.1007/978-3-030-62008-0_35.
Li, Z.L., Wu, H., Duan, S.B., Zhao, W., Ren, H., Liu, X., Leng, P., Tang, R., Ye, X., Zhu, J.,
Das, S., Angadi, D.P., 2020. Land use-land cover (LULC) transformation and its relation
Sun, Y., Si, M., Liu, M., Li, J., Zhang, X., Shang, G., Tang, B.H., Yan, G., Zhou, C.,
with land surface temperature changes: a case study of barrackpore subdivision,
2023. Satellite remote sensing of global land surface temperature: definition,
West Bengal, India. Remote Sens. Appl.: Society and Environment 19 (July 2019),
methods, products, and applications. Rev. Geophys. 61 (1), 1–77. https://doi.org/
100322. https://doi.org/10.1016/j.rsase.2020.100322.
10.1029/2022RG000777.
Dastour, H., Ghaderpour, E., Zaghloul, M.S., Farjad, B., Gupta, A., Eum, H., Achari, G.,
Liu, L., Guo, Y., Li, Y., Zhang, Q., Li, Z., Chen, E., Yang, L., Mu, X., 2022. Comparison of
Hassan, Q.K., 2022. Wavelet-based spatiotemporal analyses of climate and
machine learning methods applied on multi-source medium-resolution satellite
vegetation for the athabasca River basin in Canada. Int. J. Appl. Earth Obs. Geoinf.
images for Chinese pine (Pinus tabulaeformis) extraction on google Earth engine.
114 (September), 103044. https://doi.org/10.1016/j.jag.2022.103044.
Forests 13 (5). https://doi.org/10.3390/f13050677.
Dobriyal, P., Qureshi, A., Badola, R., Hussain, S.A., 2012. A review of the methods
Loakes, K.L., Ryves, D.B., Lamb, H.F., Schäbitz, F., Dee, M., Tyler, J.J., Mills, K.,
available for estimating soil moisture and its implications for water resource
McGowan, S., 2018. Late Quaternary climate change in the north-eastern highlands
management. J. Hydrol. 458–459, 110–117. https://doi.org/10.1016/j.
of Ethiopia: a high resolution 15,600 year diatom and pigment record from Lake
jhydrol.2012.06.021.
hayk. Quat. Sci. Rev. 202, 166–181. https://doi.org/10.1016/j.
Effati, F., Karimi, H., Yavari, A., 2021. Investigating effects of land use and land cover
quascirev.2018.09.005.
patterns on land surface temperature using landscape metrics in the city of Tehran,
Loche, M., Scaringi, G., Yunus, A.P., Catani, F., Tanyaş, H., Frodella, W., Fan, X.,
Iran. Arabian J. Geosci. 14 (13). https://doi.org/10.1007/s12517-021-07433-4.
Lombardo, L., 2022. Surface temperature controls the pattern of post-earthquake
Eshetie, G.G., Alemie, B.K., Wubie, A.M., 2025. Assessing spatio-temporal dynamics of
landslide activity. Sci. Rep. 12 (1), 1–11. https://doi.org/10.1038/s41598-022-
land degradation neutrality using goggle Earth engine in the alawuha watershed of
04992-8.
north wello zone, Ethiopia. Environ. Chall. 20 (May), 101202. https://doi.org/
10.1016/j.envc.2025.101202.

11
G. Bayable et al. Environmental and Sustainability Indicators 27 (2025) 100797

Loukika, K.N., Keesara, V.R., Sridhar, V., 2021. Analysis of land use and land cover using mapping using multi-temporal remote sensing data and google Earth engine. Remote
machine learning algorithms on google Earth engine for munneru river basin, India. Sens. 13 (8). https://doi.org/10.3390/rs13081433.
Sustainability 13 (24). https://doi.org/10.3390/su132413758. Sheykhmousa, M., Mahdianpari, M., Ghanbari, H., Mohammadimanesh, F., Ghamisi, P.,
Mengesha, T.E., Desta, L.T., Gamba, P., Ayehu, G.T., 2024. Multi-temporal passive and Homayouni, S., 2020. Support vector machine versus random forest for remote
active remote sensing for agricultural mapping and acreage estimation in context of sensing image classification: a meta-analysis and systematic review. IEEE J. Sel. Top.
small farm holds in Ethiopia. Land 13 (3). https://doi.org/10.3390/land13030335. Appl. Earth Obs. Rem. Sens. 13, 6308–6325. https://doi.org/10.1109/
Mengistie, D., Kidane, D., 2016. Assessment of the impact of small-scale irrigation on JSTARS.2020.3026724.
household livelihood improvement at gubalafto district, north wollo, Ethiopia. Shokati, H., Mashal, M., Noroozi, A., Abkar, A.A., Mirzaei, S., Mohammadi-Doqozloo, Z.,
Agriculture (Switzerland) 6 (3). https://doi.org/10.3390/agriculture6030027. Taghizadeh-Mehrjardi, R., Khosravani, P., Nabiollahi, K., Scholten, T., 2024.
Mohamed, E.S., Ali, A., El-Shirbeny, M., Abutaleb, K., Shaddad, S.M., 2020. Mapping soil Random forest-based soil moisture estimation using Sentinel-2, Landsat-8/9, and
moisture and their correlation with crop pattern using remotely sensed data in arid UAV-based hyperspectral data. Remote Sens. 16, 1962.
region. Egyptian Journal of Remote Sensing and Space Science 23 (3), 347–353. Sridhar, V., Kang, H., Ali, S.A., 2019. Human-induced alterations to land use and climate
https://doi.org/10.1016/j.ejrs.2019.04.003. and their responses for hydrology and water management in the mekong river basin.
Mumtaz, F., Tao, Y., Leeuw, G. De, Zhao, L., Fan, C., Elnashar, A., Bashir, B., Wang, G., Water (Switzerland) 11 (6). https://doi.org/10.3390/w11061307.
Li, L.L., Naeem, S., Arshad, A., 2020. Modeling spatio-temporal land transformation Stehfest, E., van Zeist, W.J., Valin, H., Havlik, P., Popp, A., Kyle, P., Tabeau, A., Mason-
and its associated impacts on land surface temperature (LST). Remote Sens. 12 (18). D’Croz, D., Hasegawa, T., Bodirsky, B.L., Calvin, K., Doelman, J.C., Fujimori, S.,
https://doi.org/10.3390/RS12182987. Humpenöder, F., Lotze-Campen, H., van Meijl, H., Wiebe, K., 2019. Key
Nasiri, V., Deljouei, A., Moradi, F., Seyed Mohammad Moein, S., Borz, S.A., 2022. Land determinants of global land-use projections. Nat. Commun. 10 (1), 1–10. https://doi.
use and land cover mapping using Sentinel-2 , Landsat-8 two composition methods. org/10.1038/s41467-019-09945-w.
Remote Sens. 14 (9), 1977. Sun, Z., Xu, R., Du, W., Wang, L., Lu, D., 2019. High-resolution urban land mapping in
Nega, W., Balew, A., 2022. The relationship between land use land cover and land China from sentinel 1A/2 imagery based on google Earth engine. Remote Sens. 11
surface temperature using remote sensing: systematic reviews of studies globally (7), 1–22. https://doi.org/10.3390/rs11070752.
over the past 5 years. Environ. Sci. Pollut. Control Ser. 29 (28), 42493–42508. Sutariya, S., Hirapara, A., Meherbanali, M., Tiwari, M.k., Singh, V., Kalubarme, M., 2021.
https://doi.org/10.1007/s11356-022-19997-z. Soil moisture estimation using Sentinel-1 SAR data and land surface temperature in
Pauwels, V.R.N., Hoeben, R., Verhoest, N.E.C., De Troch, F.P., Troch, P.A., 2002. panchmahal district, Gujarat state. Int. J. Electron. Govern. 8 (1), 65–77. https://doi.
Improvement of TOPLATS-Based discharge predictions through assimilation of ERS- org/10.30897/ijegeo.777434.
Based remotely sensed soil moisture values. Hydrol. Process. 16 (5), 995–1013. Thomas, L., Ralph, W., Kiefer, J.C., 2004. Remote sensing and image interpretation (fifth
https://doi.org/10.1002/hyp.315. edition). Geogr. J. 146, 448–449.
Peng, S., Piao, S., Ciais, P., Friedlingstein, P., Ottle, C., Bréon, F.M., Nan, H., Zhou, L., Tronquo, E., Lievens, H., Bouchat, J., Defourny, P., Baghdadi, N., Verhoest, N.E.C., 2022.
Myneni, R.B., 2012. Response to comment on “Surface urban heat island across 419 Soil moisture retrieval using multistatic L-Band SAR and effective roughness
global big cities.”. Environ. Sci. Technol. 46 (12), 6889–6890. https://doi.org/ modeling. Remote Sens. 14 (7), 1–23. https://doi.org/10.3390/rs14071650.
10.1021/es301811b. Tucker, C.J., 1979. Red and photographic infrared linear combinations for monitoring
Pham, T.D., Yokoya, N., Xia, J., Ha, N.T., Le, N.N., Nguyen, T.T.T., Dao, T.H., Vu, T.T.P., vegetation. Rem. Sens. Environ. 8 (2), 127–150. https://doi.org/10.1016/0034-
Pham, T.D., Takeuchi, W., 2020. Comparison of machine learning methods for 4257(79)90013-0.
estimating mangrove above-ground biomass using multiple source remote sensing UN-HABITAT, 2024. World Cities Report 2024; Cities and Climate Action.
data in the Red River Delta biosphere reserve, Vietnam. Remote Sens. 12 (8), 1–24. Vapnik, C., Cortes, V., 1995. Support-vector networks. In: Saitta, L. (Ed.), Machine
https://doi.org/10.3390/RS12081334. Leaming, vol. 20. Kluwer Academic Publisher. https://doi.org/10.1088/1742-6596/
Roy, B., Bari, E., 2022. Examining the relationship between land surface temperature and 628/1/012073.
landscape features using spectral indices with google Earth engine. Heliyon 8 (9), Weng, Q., Lu, D., Schubring, J., 2004. Estimation of land surface temperature-vegetation
e10668. https://doi.org/10.1016/j.heliyon.2022.e10668. abundance relationship for urban heat island studies. Rem. Sens. Environ. 89 (4),
Saha, A., Patil, M., Goyal, V.C., Rathore, D.S., 2018. Assessment and Impact of Soil 467–483. https://doi.org/10.1016/j.rse.2003.11.005.
Moisture Index in Agricultural Drought Estimation Using Remote Sensing and GIS Western, A.W., Zhou, S.L., Grayson, R.B., McMahon, T.A., Blöschl, G., Wilson, D.J., 2004.
Techniques 2. https://doi.org/10.3390/ecws-3-05802. March 2019. Spatial correlation of soil moisture in small catchments and its relationship to
Saha, J., Ria, S.S., Sultana, J., Shima, U.A., Hasan Seyam, M.M., Rahman, M.M., 2024. dominant spatial hydrological processes. J. Hydrol. 286 (1–4), 113–134. https://doi.
Assessing seasonal dynamics of land surface temperature (LST) and land use land org/10.1016/j.jhydrol.2003.09.014.
cover (LULC) in bhairab, Kishoreganj, Bangladesh: a geospatial analysis from 2008 Woldemariam, G.W., Tibebe, D., Mengesha, T.E., Gelete, T.B., 2022. Machine-learning
to 2023. Case Stud. Chem. Environ. Eng. 9 (August 2023), 100560. https://doi.org/ algorithms for land use dynamics in Lake haramaya watershed, Ethiopia. Modeling
10.1016/j.cscee.2023.100560. Earth Systems and Environment 8 (3), 3719–3736. https://doi.org/10.1007/s40808-
Semaw, F., Zeleke, G., Balew, A., 2022. Evaluating flood risk management practices and 021-01296-0.
vulnerability mapping in alawuha watershed (north wollo zone, Ethiopia) using GIS Xu, H., 2006. Modification of normalised difference water index (NDWI) to enhance open
and remote sensing. Applied Geomatics 14 (2), 347–367. https://doi.org/10.1007/ water features in remotely sensed imagery. Int. J. Rem. Sens. 27 (14), 3025–3033.
s12518-022-00429-z. https://doi.org/10.1080/01431160600589179.
Sewak, M., Sahay, S.K., Rathore, H., 2018. Comparison of deep learning and the classical Yin, D., Song, X., Zhu, X., Guo, H., Zhang, Y., Zhang, Y., 2023. Spatiotemporal analysis of
machine learning algorithm for the malware detection. Proceedings - 2018 IEEE/ soil moisture variability and its driving factor. Remote Sens. 15 (24). https://doi.
ACIS 19th International Conference on Software Engineering, Artificial Intelligence, org/10.3390/rs15245768.
Networking and Parallel/Distributed Computing 293–296. https://doi.org/10.1109/ Zeleke, G., Teshome, M., Ayele, L., 2022. Assessment of land degradation neutrality
SNPD.2018.8441123. SNPD 2018, July. status in north wello zone, northern Ethiopia. Ethiopian Renaissance Journal of
Shao, Y., Lunetta, R.S., 2012. Comparison of support vector machine, neural network, Social Sciences and Humanities 9 (1), 47–75. https://doi.org/10.20372/
and CART algorithms for the land-cover classification using limited training data erjssh.2022.0901.03.
points. ISPRS J. Photogrammetry Remote Sens. 70, 78–87. https://doi.org/10.1016/ Zhang, L., Dirmeyer, P.A., Wei, J., Guo, Z., Lu, C.H., 2011. Land-atmosphere coupling
j.isprsjprs.2012.04.001. strength in the global forecast system. J. Hydrometeorol. 12 (1), 147–156. https://
Sharma, A., Vashishtha, D., 2023. Spatio-temporal assessment of land use land cover doi.org/10.1175/2010JHM1319.1.
changes and their impact on variations of land surface temperature in aligarh Zhao, C., Jia, X., Zhu, Y., Shao, M., 2017. Long-term temporal variations of soil water
municipality. Journal of the Indian Society of Remote Sensing 51 (4), 799–827. content under different vegetation types in the loess Plateau, China. Catena 158
https://doi.org/10.1007/s12524-022-01652-2. (January), 55–62. https://doi.org/10.1016/j.catena.2017.06.006.
Shetty, S., 2019. Analysis of machine learning classifiers for LULC classification on Zhao, Z., Islam, F., Waseem, L.A., Tariq, A., Nawaz, M., Islam, I.U., Bibi, T., Rehman, N.
google Earth engine, 8 (23), 1–22. U., Ahmad, W., Aslam, R.W., Raza, D., Hatamleh, W.A., 2024. Comparison of three
Shetty, S., Gupta, P.K., Belgiu, M., Srivastav, S.K., 2021. Assessing the effect of training machine learning algorithms using google Earth engine for land use land cover
sampling design on the performance of machine learning classifiers for land cover classification. Rangel. Ecol. Manag. 92, 129–137. https://doi.org/10.1016/j.
rama.2023.10.007.

12