论文信息
论文解读
AI 辅助整理- 方法
- 用Sentinel-1/2、CHIRPS、NASADEM多源遥感与地面样本,嵌套
- 发现
- GTB总体精度最高(0.96),RF的AUC-ROC达0.99,共识图识别灌溉面积58.85万公顷。
- 机会点
- 可探索多模型集成与时间窗口优化在数据稀缺流域的迁移性,并引入作物物候特征提升泛化。
论文摘要
可靠的灌溉农业制图对可持续水资源管理、水资源配置和农业生产力至关重要,尤其是在上青尼罗河流域(Upper Blue Nile Basin)等数据稀缺的流域。尽管该流域具有显著的灌溉潜力,但准确且空间明确的灌溉面积地图仍然有限。本研究评估并比较了随机森林(Random Forest, RF)、支持向量机(Support Vector Machine, SVM)、梯度树提升(Gradient Tree Boosting, GTB)和双向长短期记忆网络(Bidirectional Long Short-Term Memory, BiLSTM)等机器学习分类器,利用多源遥感数据集,对埃塞俄比亚上青尼罗河流域2025年12月至2026年4月灌溉季的灌溉区域进行制图。分类器使用了两组互补数据:多源遥感及辅助预测变量,包括Sentinel-2多光谱影像、Sentinel-1合成孔径雷达后向散射、CHIRPS降雨估计和NASADEM衍生的地形变量;以及用于模型训练和验证的灌溉田与雨养田地面参考样本。模型性能采用总体精度、Kappa系数、F1分数、精确率、召回率和AUC-ROC等性能指标进行评估,并辅以五折外层空间嵌套空间交叉验证。在四个分类器中,GTB的总体性能最高,其精度、Kappa系数和F1分数分别为0.96、0.91和0.96。RF排名第二,其AUC-ROC值最高,为0.99,其次是BiLSTM和SVM。分析表明,2月Sentinel-2光谱指数是最具影响力的预测变量,2月是GTB和RF最重要的影像获取月份。采用多模型集成生成了保守的一致性地图。该一致性地图将上青尼罗河流域588,545公顷土地分类为灌溉农业。
全文
Research SquareComparative Evaluation of Machine Learning
Algorithms for Irrigated Area Mapping in the Upper
Blue Nile Basin
Abdurezak Arage Hamza
Addis Ababa University
Zeleke Agide Dejen
Addis Ababa University
Sirak Tekleab Gebrekristos
Addis Ababa University
Haileyesus Belay Lakew
Addis Ababa University
Tewele Gebretsadkan Haile
Adigrat University
Research Article
Keywords: Irrigated area, Machine Learning, Random Forest, Gradient Tree Boosting, Remote Sensing,
Upper Blue Nile Basin
Posted Date: November 13th, 2026
DOI: https://doi.org/10.21203/rs.3.rs-11524145/v1
License: This work is licensed under a Creative Commons Attribution 4.0 International License.
Read Full License
Additional Declarations: No competing interests reported.
1
Comparative Evaluation of Machine Learning Algorithms
for Irrigated Area Mapping in the Upper Blue Nile Basin
Abdurezak Arage Hamza 1,3, Zeleke Agide Dejen 2, Sirak Tekleab Gebrekristos 2, Haileyesus Belay
Lakew 2, Tewele Gebretsadkan Haile 4
1 Africa Center of Excellence for Water Management, Addis Ababa University, Addis Ababa, Ethiopia
2 Center for Water Research, Institute of Water, Environment and Climate Research, Addis Ababa
University, Addis Ababa, Ethiopia
3 School of Water Resource and Environmental Engineering, Haramaya Institute of Technology,
Haramaya University, Haramaya, Ethiopia
4 College of Engineering and Technology, Adigrat University, Adigrat, Ethiopia
Email: abdurezakhwre@gmail.com
Abstract
Reliable mapping of irrigated agriculture is important for sustainable water resource management,
allocation and agricultural productivity, particularly in data-scarce river basins such as Upper Blue
Nile Basin. Despite the basin’s significant irrigation potential, accurate and spatially explicit
irrigated area maps remain limited. This study evaluated and compared machine learning
classifiers including Random Forest (RF), Support Vector Machine (SVM), Gradient Tree
Boosting (GTB), and Bidirectional Long Short-Term Memory (BiLSTM) for mapping irrigated
areas in the Upper Blue Nile Basin, Ethiopia during the December 2025 to April 2026 irrigation
season using multi source remote sensing datasets. The classifiers used two complementary data
groups: multi-source remote-sensing and ancillary predictor variables, including Sentinel- 2
multispectral imagery, Sentinel-1 synthetic aperture radar backscatter, CHIRPS rainfall estimates,
and NASADEM-derived topographic variables; and ground-reference samples of irrigated and
rainfed fields for model training and validation. The model performance was evaluated using the
performance metrics overall accuracy, kappa coefficient, F1-score, precision, recall, and AUC-
ROC, supported by nested spatial cross-validation with five outer spatial folds. Among the four
classifiers, GTB scored the highest overall performance with accuracy, Kappa coefficient, F1-
score, with values of 0.96, 0.91, and 0.96, respectively. RF ranked second and scored the highest
AUC-ROC value of 0.99, followed by BiLSTM and SVM. Analysis showed that February
2
Sentinel-2 spectral indices were the most influential predictors and February was the most
important acquisition month for GTB and RF. Multi model ensemble was used to produce a
conservative consensus map. The consensus map classified 588,545 ha as irrigated agriculture in
the Upper Blue Nile Basin.
Keywords: Irrigated area, Machine Learning, Random Forest, Gradient Tree Boosting, Remote
Sensing, Upper Blue Nile Basin
- Introduction
Globally, Irrigated agriculture uses approximately 70% of the available freshwater withdrawals [1]
and up to 90% in some least income countries [2]. Due to population growth and climate variability
efficient irrigation system management and monitoring have become critical to ensure food
security and sustainable water resource utilization [3]. To understand the impact of consumptive
water use by agriculture development of high-resolution irrigated area maps is essential [4]. Recent
technological advancement in irrigated agriculture significantly improves water productivity and
reduces environmental impacts [1].
Challenges in food security and increasing water scarcity need accurate irrigation mapping for
sustainable agricultural water management [5]. Traditional ground-based irrigation mapping is
inadequate for large scale irrigation assessment due to lack of spatial and temporal extent and high
associated cost [6]. As a result, remote sensing technologies have become reliable tools for
providing spatially comprehensive and temporally dynamic information on irrigated agriculture
[7]. However, in heterogeneous smallholder farms, distinguishing irrigated areas from rainfed
agriculture areas pose significant challenge [8].
Upper Blue Nile is a transboundary basin that is one of the essential water resources feeding into
the Nile River and supporting around 250 million people. The basin experiences limited research
on irrigated areas and water demands due to lack of data and faces significant environmental
pressures due to poor land management practices [9]. Recent studies in the basin indicate dramatic
land use changes especially increasing the cultivated lands [10]. The basin’s irrigation potential is
largely untapped, except small fractions of suitable land being currently under irrigation [11].
3
Although irrigation is expanding in the basin, its actual extent remains uncertain because many
smallholder irrigated fields are scattered and fragmented across the basin. This makes it
challenging to conduct irrigation planning, monitor the growth of irrigation, assess agricultural
water demands, and analyze the hydrological impacts of irrigation.
Remote sensing has become a powerful tool for monitoring irrigated agriculture across various
climatic conditions. Over the past decades advancement of remote sensing technologies
established vegetation indices as the reliable source of identifying irrigation fields. Studies show
that multi-temporal remote sensing analysis are effective for irrigated area mapping rather than
single date satellite imagery that enable general irrigation classification [7]. Higher spatial
resolution satellite missions such as Sentinel 1 and Sentinel 2 have transformed field scale
agricultural analysis, which is especially important in smallholder and fragmented agricultural
areas. High resolution imagery provided by Sentinel 2 is the highest spatial resolution 10 m from
the open-source sensors that is considerable improvement for agricultural landscapes [12].
The integration of remote sensing data with machine learning algorithms enhances the accuracy
and robustness of irrigated area classification [13]. Random Forest has been implemented in
satellite remote sensing-based land classification studies for classification of agricultural land uses
and suitable to handle high dimensional data [4]. Support Vector Machine has shown strong
performance for large scale irrigation mapping [14]. Recent studies in Ethiopia shows that
application of machine learning is effective for accurately mapping irrigation areas in different
watersheds [15]. Advanced machine learning approaches, particularly ensemble methods and deep
learning are essential for crop mapping because they can capture by nonlinear relationships among
predictor variables [16].
Therefore, the aim of this study is to identify irrigated areas in the Upper Blue Nile Basin during
the December 2025–April 2026 irrigation season by comparatively evaluating Bidirectional Long
Short-Term Memory algorithms, Random Forest, Support Vector Machine, and Gradient Tree
Boosting models.
4
- Materials and Methods
2.1. Study Area
The upper Blue Nile Basin (UBNB) is part of the Blue Nile Basin contained within the Ethiopian
national boundary. The main river originates in Lake Tana and join the White Nile at Khartoum,
Sudan. The basin covers a large part of the Ethiopian Highlands, from an altitude of less than 500
meter above mean sea level (m amsl) at the Sudanese border to more than 4200 m amsl in the
Central and the Eastern escarpment of the Ethiopian Highlands [17]. This steep elevation gradient
defines its rugged topography of escarpments, high plateaus, and deep gorges, which strongly
influence hydrological responses and sediment transport processes [18]. The climate of the basin
is governed by the seasonal migration of the Intertropical Convergence Zone (ITCZ) and
associated monsoon systems. The main rainy season (Kiremt) extends from June to September and
contributes roughly 70–80% of annual rainfall and streamflow, followed by the dry season (Bega)
from October to January. A smaller rainy season (Belg) occurs from February to May. Annual
rainfall varies greatly, from 800 mm in the drier lowlands to more than 2,200 mm in the highlands
[19]. Temperature also varies with elevation, ranging from 15–25°C in highland zones to over
35°C in the lowlands during peak dry months [20] . The Hydrology of the Upper Blue Nile River
basin exhibits large seasonal and interannual variations. The average annual discharge at the
Sudanese outlet has been estimated 48-52 billion cubic meters with more than two-thirds of the
discharge during the rainy(Kiremt) season [21]. Recently, the Blue Nile River Basin has been
experiencing changes in the regimes of rainfall and temperatures influenced by climate change.
Ensemble projections show increase in rainfall by 10-25% in some scenarios, while the variability
of the distribution of the latter will raise uncertainties of water availability [22]. The rise in
temperatures will increase evapotranspiration demand with implications for crop water demand
and for the operation of the reservoirs [23].
5
Figure 1: Location map of the Upper Blue Nile Basin study area
2.2. Data Sources
Various static and time series data were used for the machine learning classification models. Multi
source geospatial datasets including topographic, remote sensing rainfall data, and remote sensing
imagery products are integrated to map irrigated area in the upper Blue Nile basin.
2.2.1. Remote Sensing Rainfall Data
The accuracy of irrigation mapping during the irrigation season is affected by the amount of
rainfall. Satellite derived rainfall estimates provide precipitation data at different temporal and
spatial scales and play a key role for hydrological applications. CHIRPS provides high-resolution
precipitation estimates (0.05° or about 5 km) daily, pentadal, and monthly time steps from 1981 to
the current year. CHIRPS outperforms other satellite rainfall estimates indicating strong regional
preference in Upper Blue Nile Basin [24].
2.2.2. Remote Sensing Imagery Products
6
Satellite imagery products provide spatially explicit information about land cover, vegetation, and
moisture status. They are typically static or derived from repeated acquisitions and are essential
for irrigated area mapping and suitability analysis. Sentinel-2 MSI has the capability to provide
multi-spectral images with resolutions of 10 to 20 m since 2015. The main applications include
irrigated area mapping, vegetation index computation, and vegetation growth and health status.
Sentinel-1C-band Synthetic Aperture Radar Ground Range Detected (SAR GRD) which is
calibrated and terrain corrected was obtained from Google Earth Engine. SAR provide dual
polarization backscatter, Vertical Transmit Vertical Receive (VV) and Vertical Transmit Horizontal
Receive (VH). VV is mainly sensitive to soil moisture condition and VH is sensitive to crop canopy
characteristics. The combination used of the two polarizations, VV and VH, improves the detection
of agricultural dynamics.
2.2.3. Topography
Topography data was obtained from the NASA Digital Elevation Model (NASADEM) through
Google Earth Engine (GEE). Digital Elevation Model influence surface hydrology, agricultural
suitability and soil moisture distribution [25]. Slope is a critical factor for irrigation feasibility with
flat terrains are more suitable for irrigation development.
2.2.4. Reference Data
The reference dataset included 482 polygons of field in the basin agricultural areas. Polygon
delineation was performed using visual interpretation of very high-resolution imagery from
Google Earth using the field boundary, the vegetation condition during dry season, water-related
structures and agriculture landscape pattern. Irrigated fields showed stable greenness in dry season
and patterns of irrigation, and rainfed fields showed poor greenness during dry season and absence
of irrigation structure. Predictor values were extracted from pixels within each polygon boundary
on the analysis grid. Samples were partitioned by whole polygons and spatial blocks rather than
individual pixels, ensuring that all observations from a given field were assigned exclusively to
either the model-development or validation set.
2.2.5. Land Use Land cover
A 2024 land use land cover map of Upper Blue Nile was used as ancillary dataset to identify
agricultural land. The purpose of the LULC map was to mask nonagricultural land cover such as
7
forests, shrubland, water bodies, and other classes before model training and prediction. The
agricultural land derived from the LULC map was used to filter the predictor variables and
reference samples before model development.
Figure 2: Land use/land cover map of the Upper Blue Nile Basin (2024) used to derive the
agricultural mask applied in irrigation classification
2.3. Preprocessing of Remote Sensing and Ancillary Data
Sentinel-1C-band Synthetic Aperture Radar Ground Range Detected (SAR GRD) is suitable for
all-whether agricultural monitoring because SAR observations are largely unaffected by cloud
cover [26] . To reduce the skewness in the data distribution Radar backscatter were converted to
decibel scale using logarithmic transformation. To improve the class separability between rainfed
and irrigated the VH/VV ratio was computed. Sentinel 2 surface reflectance data were
atmospherically corrected and radiometrically consistent from the Sentinel 2A and Sentinel 2B
satellites [27]. Cloud masking was applied using the quality assessment (QA60) band. Based on
bitwise filtering, the cloud and cirrus pixels were masked. After cloud masking, vegetation and
moisture indices derived from multispectral satellite data were used to identify irrigated areas.
8
Irrigated croplands generally maintain higher vegetation vigor and sustained moisture conditions
compared to rainfed systems, particularly during dry periods [28]. Therefore, integrating multiple
spectral indices and their temporal dynamics improves the reliability of irrigation mapping.
The Normalized Difference Vegetation Index (NDVI) was derived from Sentinel-2 images to
characterize vegetation greenness and crop development. NDVI was included because irrigated
fields usually maintain higher vegetation activity than rain-fed fields during the dry season. The
index was calculated from red and near-infrared reflectance values.
𝑁𝐷 =𝑁 − 𝐷𝑁 + 𝐷 (1)
However, NDVI can be influenced by soil background effects, particularly in sparsely vegetated
areas. To address this limitation, the Soil Adjusted Vegetation Index (SA VI) is incorporated,
which minimizes soil brightness effects and improves vegetation detection under varying soil
conditions:
= 𝑁 − 𝐷(𝑁 + 𝐷 + ) ∗ (1 + ) (2)
where L is a soil adjustment factor (commonly 0.5).
Normalized Difference Water Index (NDWI) was used as a measure of surface moisture
conditions. Higher NDWI values are commonly found in irrigated fields due to the presence of
water during the irrigation season.
𝑁𝐷 =𝑁 − 𝑁𝑁 + 𝑁 (3)
The Normalized Difference Moisture Index (NDMI) was used to represent vegetation moisture
status. As irrigated crops generally exhibit higher canopy moisture than rainfed crops, NDMI
provides useful information for distinguishing irrigated areas.
𝑁𝐷 =𝑁 − 𝑁 + (4)
9
Monthly and seasonal composites of NDVI, NDWI, NDMI, and SA VI were generated for the
irrigation season. Summary statistics including mean, maximum, and standard deviation were
calculated to capture temporal variations in vegetation and moisture conditions. These variables
were subsequently used as predictor features in the machine learning models.
2.4. Machine Learning Classifications Models
Four machine learning classification algorithms were evaluated for irrigated area mapping in the
Upper Blue Nile Basin. Random Forest (RF), Support Vector Machine (SVM), Gradient Tree
Boosting (GTB), and Bidirectional Long Short-Term Memory (BiLSTM). These methods
represent different learning approaches, including ensemble learning, margin-based classification,
boosting, and deep learning. Random Forest is ensembles learning method that builds multiple
decision trees for training, and it combines its prediction through majority voting for classification
[29]. The Random Forest algorithm is widely used in remote sensing applications because it can
handle high-dimensional data and nonlinear relationships among predictor variables [30], [4] .
Support Vector Machine identifies an optimal hyperplane that increases the margin between classes
in multidimensional feature space [31]. The method has been successfully applied in irrigation
mapping and other remote sensing classification studies because of its ability to perform well with
complex datasets and limited training samples [14]
Gradient Tree Boosting is an ensemble machine learning algorithm that sequentially incorporates
multiple weak learners, typically low decision trees to create strong predictive models. This
algorithm iteratively trains new trees to minimize errors by optimizing a differentiable loss
function through gradient descent [32]. The step-by-step learning strategy of GTB enable to
capture complex nonlinear relationships for effective classification [33]. Bidirectional Long Short-
Term Memory is a deep learning architecture derived from Recurrent Neural Network and
specially designed for temporal and sequential data modeling. It incorporates forward and
backward parallel recurrent processing simultaneously to learn the model temporal dependencies
from both past and future observations [34]. BiLSTM models are effective for time series analysis
because they can collect seasonal and temporal variations [35].
The same predictor dataset was used to develop the four classifiers and included Sentinel-1 SAR
variables, Sentinel-2 spectral indices, CHIRPS rainfall data, topographic variables, and seasonal
metrics. The models were evaluated with nested spatial cross-validation with five outer folds. The
10
performance was compared by overall accuracy, Kappa coefficient, precision, recall, F1-score and
AUC-ROC. This comparative framework allowed us to evaluate the relative performance of
ensemble-learning, margin-based and deep-learning approaches for mapping of irrigated areas in
the Upper Blue Nile Basin. For RF and GTB, predictor importance was obtained from the fitted
tree-based models. For SVM, permutation importance was computed on the full reference dataset
using the final tuned models as the mean decrease in accuracy over 10 permutations per predictor;
values were not averaged across outer folds.
2.5. Evaluation Metrics and Model Selection
Model performance was assessed using multiple evaluation metrics to ensure a multi-dimensional
comparison of the four classifier models. Overall accuracy is the proportion of correctly classified
samples compared to the total number of samples. Precision, recall, F1-score were calculated for
potential class imbalance, where F1-score served as balance between precision and recall avoiding
misleading conclusion that overall accuracy may produce in imbalanced datasets. The Area Under
the Receiver Operating Characteristic Curve (AUC-ROC) evaluates each model’s discriminative
ability across all possible classification thresholds. Cohen’s Kappa coefficient was used to quantify
the degree of agreement between model predictions and reference data. The performance
evaluation metrics were evaluated with nested spatial cross validation with five outer spatial folds
to ensure that the model performance reflects true spatial generalizability rather than overfitting
spatially autocorrelated training samples [36]. The final model selection was based on combination
of performance across all metrics. The evaluation metrics were calculated using Equations (5)-(9)
= + 𝑁 + 𝑁 + + 𝑁 (5)
!"#"$% = + (6)
!'' = + 𝑁 (7)
1 = 2 ∗ !"#"$% ∗ !''!"#"$% + !'' (8)
11
- =+ − ,
1 − ,
(9)
where TP, TN, FP, FN, Po, and P e denote true positives, true negatives, false positives, and false
negatives, observed agreement, agreement expected by chance respectively.
2.6. Multi Model Ensemble Method
To enhance the individual model irrigated area map in the Upper Blue Nile Basin, a Multi Model
Ensemble Mapping approach was developed by integrating predictions from the four machine
learning models. The ensemble applied a conservative 3 of 4 voting rules to retain areas supported
by at least three models and minimize the misclassification errors from the individual models. Its
performance was evaluated using pooled out of fold predictions from the five spatial cross
validation folds. Akbar et al., (2025) show combining output of multiple machine learning
classifiers significantly improve individual models in irrigation area mapping compared to ground
truth accuracy. Ensemble mapping ensures the irrigation map classification was made on those
locations where strong cross model consensus exists.
- Results and Discussions
3.1. Model Performance and Statistical Evaluation
The comparison of the four machine learning algorithms for December 2025 to April 2026
irrigation season showed clear differences in descriptive performance metrics. GTB scored the
highest average performance of 0.96, which indicates it correctly classified 96% in the evaluation
set. Followed by Random Forest within an accuracy of 0.95. BiLSTM attaining an accuracy of
0.95 and support vector machine scored 0.91. The kappa coefficient adjusts the possibility of
agreement occurring by chance and it is robust indicator of model reliability. The kappa coefficient
result shows notable drop from the top GTB and RF models. The AUC-ROC result indicates highly
uniform for all models with RF and GTB scored 0.99, respectively. BiLSTM scored AUC-ROC of
0.98, and SVM recorded 0.97. AUC-ROC values near to 1 indicate strong ability to discriminate
between negative and positive classes across all classification threshold for all models. The F1-
score indicates the harmonic mean of precision and recall that balance false positives and false
negatives. The result confirmed the superiority of GTB and RF. This indicates RF and GTB
12
maintained strong balance between the false negative and false positive to minimize the
classification errors. The performance metrics of the four classifiers are summarized in Figure 3.
Figure 3: Performance metrics of the evaluated machine learning models
The GTB consistently scores high performance across all performance evaluation metrics, and it
is the most reliable general-purpose classifier in this study. This model’s precision score shows as
key finding, suggests that its learned decision boundary generated positives predictions when the
input features provided convincing evidence. GTB scored a recall of 0.95, indicating that
approximately 5% of reference irrigated samples were classified as non-irrigated. This precision-
recall tradeoff is a fundamental challenge in classification. From an architectural perspective , its
powerful performance is consistent with its ability to capture complex nonlinear relationships
through sequentially constructed decision trees.
Random Forest model ranked second and showed powerful performance across all evaluation
metrics. The result indicates a balanced ability to decrease false positive classifications while
correctly detecting irrigated fields. Although RF performed slightly below GTB in accuracy,
Kappa, and F1-score, its high AUC-ROC shows strong discrimination between irrigated and
rainfed fields across the classification thresholds. Lu et al. (2020) achieved an over accuracy of
93.3% and a Kappa coefficient of 0.916 using RF and multispectral imagery, while Magidi et al.
(2021) reported an overall accuracy of 88% using Landsat and Sentinel data.
13
BiLSTM model detects most irrigated fields correctly but omitted approximately 6.07% of the
actual irrigated samples. Although performing slightly below RF and GTB, BiLSTM shows strong
classification performance. Previous studies show that BiLSTM based models can effectively learn
temporal patterns from multispectral remote sensing data. Shi et al. (2026) reported that an LSTM
based framework integrating Sentinel-1 and Sentinel-2 observations improved crop classification
performance with an overall accuracy of 93.61% and a Kappa coefficient of 91.66%. This finding
supports the broader usefulness of LSTM based architecture for detecting temporal agricultural
signals.
SVM model demonstrates comparatively lower performance across the evaluation metrics and is
ranked fourth among the four models. SVM produced comparatively few false positive irrigated
classifications but omitted approximately 14% of the actual irrigation samples. The SVM
performance result remains consistent with previous evidence that SVM can effectively classify
irrigated and non-irrigated fields using multitemporal remote sensing data. Gao et al. (2018)
obtained an overall accuracy of 81.08% using SVM sentinel-1 time series data. Another study
using both sentinel-1 and sentinel-2 data obtained an accuracy of 88.4% for classifying irrigation.
In this study SVM showed strong classification ability, but its low recall shows that it failed to
detect more irrigated fields than other models.
Overall, the performance evaluation shows that model selection should consider multiple metrics
rather than overall accuracy alone. The difference in recall, precision, and discriminatory ability
shows that each classifier captures distinct aspect of irrigated field characteristics. These
harmonious classification behaviors support integrating the four models output through multi
model ensemble mapping to reduce errors across individual model’s classifiers. A more detailed
comparison of model performance is presented in Figure 4.
14
Figure 4: Detail Performance analysis
The radar chart and heatmap show that GTB, RF, and BiLSTM achieved balanced performance
across the six-evaluation metrics, while SVM lower recall produced the clearest performance
imbalance. The descriptive mean summarizes the overall pattern, but it should be treated as a
comparative indicator rather than an independent validation metric.
3.2. Spatial Cross Validation
Model performance was further evaluated using nested spatial cross-validation with five outer
spatial folds. Four spatial folds were used in each iteration for model development, while the
remaining fold was kept out for independent evaluation. Hyperparameter tuning and model
selection were conducted within the outer training data through the inner cross validation loop.
Data imputation, scaling, and other fitted preprocessing operations were estimated from the
training folds and then applied to the held-out fold. This prevents the outer test folds from affecting
model selection and decreases the optimistic performance estimates associated with using the same
observation for evaluation and tuning [36]. Spatial Random cross validation may overestimate
predictive performance when geographically adjacent training and test samples are spatially
dependent [41].
The four-classifiers showed geographically variable performance across the five outer folds. All
four classifier models performed particularly well in Fold 4. However, performance was lower in
Fold 1 for several models. SVM showed lower scores than the other models in this fold. SVM
performance improved in Folds 3 and 4 but declined again in Fold 5, that indicates greater
geographic variability than the other classifier models. BiLSTM had a lower mean kappa
15
coefficient than GTB and RF but showed the smallest fold- to-fold kappa variation, with standard
deviation of 0.05. In contrast, SVM had the largest fold to fold Kappa variation with standard
deviation of 0.12. GTB scored the highest mean performance with moderate variation, indicating
comparatively strong balance between average accuracy and spatial stability. The difference
among spatial folds indicates that the classification of irrigated and rainfed fields was not uniform
across the study area. Such variation could potentially be associated with geographic differences
in crop phenology or irrigation practice. The near perfect scores in Fold 4 show that irrigated and
rainfed fields were highly separable within that held out region. The nested spatial cross validation
results support GTB as the strongest average classifier, while the comparatively stable
performance of BiLSTM and greater geographic sensitivity of SVM. The fold-level performance
obtained from nested spatial cross-validation is presented in Figure 5.
Figure 5: Fold-level performance
Figure 6 summarizes the mean performance and fold-to-fold variability of the evaluated
classifiers.
16
Figure 6: Mean performance and fold to fold standard deviation.
3.3. Statistical Comparison
The pairwise exact McNemar tests indicate that GTB and RF had significantly lower paired
classification error than SVM after Holm correction. BiLSTM scored more correct predictions
than SVM among the discordant cases. This discrepancy was not significant after adjustment. No
statistically significant differences were detected within GTB, RF and BiLSTM, indicating that
these models are under statistically comparable leading groups. The pairwise exact McNemar
test was used to compare the classification errors of the evaluated models based on paired
predictions from the nested spatial cross-validation.
Table 1: Pairwise exact McNemar tests.
Model A Model B A+/B- A-/B+ Exact p-value Holm p-value α = 0.05
GTB SVM 29 8 0.0007 0.0045 Significant
RF SVM 25 7 0.0021 0.0105 Significant at 5%
BiLSTM SVM 27 11 0.0138 0.0554 Non-significant
GTB BiLSTM 9 4 0.2668 0.8005 Non-significant
RF GTB 4 7 0.5488 1.0000 Non-significant
RF BiLSTM 9 7 0.8036 1.0000 Non-significant
17
3.4. Feature Importance
Feature importance score was extracted for SVM, GTB and RF to understand the relative
contribution of each feature variable in the classification process. This analysis was used to identify
relative contribution of predictor variables. Because the feature importances were obtained from
model specific approaches. GTB was dominated by Sentinel-2 NDMI feature followed by
February EVI and January NDVI. Random Forest score had a much more evenly distributed
feature importance. The top ranked features are February NDMI, February SA VI, February NDVI,
and March NDMI. The near equal importance of this February indices shows that Random Forest
uses complementary information from multiple spectral indices rather than depending on a single
dominant feature. The SVM feature importance depends on with selected Sentinel-1, Sentinel-2,
seasonality variable rainfall and topographic variable. This suggests the model leverages on
environmental conditioning variables. For SVM, predictor ranking was obtained using
permutation feature importance applied to the final tuned model on the reference dataset. From
the spectral features SVM uses a broader spectral signal from February to April rather than on
concentrating on single temporal window. Figure 7 presents the model-specific predictor rankings.
18
Figure 7: Model-specific predictor importance. RF and GTB use tree-based importance; SVM
uses permutation sensitivity
Sentinel-2 contributed approximately 97% of GTB. The other variables Sentinel-1, DEM, and
Rainfall made small contributions. This indicates GTB based classification decisions depend on
a limited set of optical indicators representing vegetation condition and moisture. Sentinel-2 is
also dominant in RF contributing approximately 94% of the grouped importance. Its individual
importance was distributed across different variables particularly NDMI, SA VI, NDVI and EVI.
Therefore, Random Forest did not depend on one predictor. SVM individual feature importance
shows rainfall variables contribute highly rather than the Sentinels imagery. The monthly
acquisition result shows February was the most important month for GTB and RF. For GTB and
RF approximately 87% and 46% respectively comes from its grouped importance of February
dataset. The aggregated importance of predictor variables by data source and acquisition month
is presented in Figure 8.
Figure 8: Importance aggregated by source and acquisition month
3.5. Irrigated Area Map of Upper Blue Nile basin
All models consistently identified irrigated agriculture as spatially limited and unevenly distributed
across the basin. The irrigated areas were scattered concentrated primarily in the central, northern,
southwestern, and southern parts of the Upper Blue Nile. The rainfed agriculture dominated in the
northern and central highlands of the basin, indicating high dependence on seasonal rainfall for
19
agricultural production in the basin. The spatial distribution of irrigated and rainfed agriculture
predicted by the four machine learning models is presented in Figure 9.
Figure 9: Map of Irrigated and Rainfed agricultural lands
All four classifiers showed the same general spatial pattern, but ranged considerably in estimated
irrigated extent. The highest estimation was given by BiLSTM (871,049 ha) followed by GTB
(793,809 ha), RF (670,626 ha) and SVM (614,261 ha). The difference between the highest and
lowest individual-model estimates was 256,788 ha, indicating that classifier choice can
considerably impact basin-scale area estimations, even when traditional accuracy metrics are
equal. The GTB model produced extensive and concentrated irrigated clusters, particularly in the
central parts of the basin. RF model produced broadly similar patterns but classified some irrigated
areas more conservatively. BiLSTM detects many similar clusters, while differences occurred in
smaller and spatially scattered areas. SVM has shown the most limited and fragmented irrigated
20
area distribution, which is consistent with its comparatively low recall and indicates that it ignored
more irrigated fields than other models. The spatial differences between the maps reflect variation
in how the classifiers used the predictor variables. GTB depended strongly on February Sentinel-
2 moisture and vegetation indices, while RF distributed predictors over various spectral indices
and months. SVM assigned comparatively greater importance to rainfall variables. Previous
studies in the Upper Blue Nile watershed similarly found that monthly radar and vegetation index
information improved irrigation classification [42].
GTB obtained the highest overall performance among the evaluated models. However, the
comprehensive spatial agreement with GTB, RF, and BiLSTM together with the lack of
statistically significant differences in their paired error rates, indicates that the important irrigation
patterns detected by these classifiers are strong. The significance of the irrigation map results show
irrigation development has limited extent relative to the basin’s large agricultural area. Overall, the
maps demonstrate consistent irrigation signals and model reliant spatial ambiguity. Areas classified
as irrigated by various models have stronger cross model support, while areas predicted only by
one model represent higher classification uncertainty. The observed agreement and disagreement
indicate the importance of multi model ensemble mapping, which can maintain locations supported
by multiple models while decreasing errors related to individual model predictions.
3.6. Model Ensemble Mapping
The application of Multi Model Ensemble Mapping produced a high agreement irrigated area map
by combining the predictions of GTB, RF, BiLSTM, and SVM. The pixel level agreement map
showed considerable spatial variability in agreement among the classifiers across the basin. The
ensemble approach addresses the individual single model output limitation. From the model
performance metrics result, there is variation tradeoff between precision and recall in individual
models. Quantitative validation using pooled out-of-fold predictions from the five spatial cross-
validation folds showed that the 3-of-4 ensemble achieved an overall accuracy of 95.64%,
precision of 97.48%, recall of 93.93%, an F1-score of 0.9567, Cohen’s Kappa of 0.9129, and an
AUC-ROC of 0.9904. The ensemble achieved the highest precision among the evaluated
approaches, while its overall accuracy equaled that of GTB. Therefore, it is interpreted as a
conservative, high-precision consensus model rather than as universally superior to every
individual classifier.
21
Figure 10: Model ensemble irrigated and rainfed area map in the Upper Blue Nile
The consensus irrigation map using conservative model ensemble found 588,545 ha (5,885.45
km2) for irrigation and around 8,290,763 ha as rain-fed. The minimum and maximum individual
irrigation model output was from SVM which had 614,261 ha and BiLSTM which showed 871,049
ha, thus giving a difference of around 256,788 ha. This difference is a result of variability among
the model outputs. The consensus irrigation area is smaller than all the individual model irrigation
area due to the 3-of-4 majority voting rule. This conservative approach, together with the high
overall accuracy (95.64%) and irrigated-class precision (97.48%), supports the reliability of the
consensus map.
Gebeyehu, (2018) mapped 310,707 ha of irrigated land in the Upper Blue Nile Basin for
2016/2017 using Sentinel-2 imagery and object-based image analysis. This is considerably lower
than the 588,545 ha estimated in the present study. Yimere & Assefa, (2022) reported 738,184 ha,
but their estimate covered both existing irrigation and future irrigation potential. The differences
in these estimates could be related to the methods and periods considered in each study, as well as
22
the difficulty of mapping small and scattered irrigated fields. The high classification accuracy and
agreement among at least three models suggest that the estimated irrigated area is reliable.
Table 2: Estimated irrigated and rainfed areas classified by each model
Model Irrigated (ha) Rainfed (ha)
Random Forest 670,627 8,208,682
GTB 793,810 8,085,499
BiLSTM 871,049 8,008,259
SVM 614,261 8,265,047
Ensemble 588,545 8,290,763
- Conclusion
Irrigation is important in sustaining agricultural productivity by ensuring food security, especially
where water availability fluctuates due to climatic variability in Upper Blue Nile Basin. However,
up to date information on irrigation mapping is unavailable which pose challenge in water resource
management and allocation in the basin. The aim of this study is to evaluate and compare four
machine learning models for irrigated area mapping in Upper Blue Nile Basin using multi source
remote sensing datasets. The combination of these data sets allowed a comprehensive
characterization of bio-physical conditions that differentiate irrigated land from rainfed agriculture
across the basin. Among the four machine learning classifiers models, the GTB model
demonstrated the highest performance across all performance evaluation metrics. The result
indicates that GTB was strongest individual classifier for irrigated mapping in this study. However,
the McNemar tests found no statistically significant differences in paired error rates among GTB,
RF, and BiLSTM, implying that these three models formed statistically similar leading group.
The feature-importance analysis results show that Sentinel-2 spectral indices were the most
important influential predictors for GTB and RF. February was the dominant acquisition month,
highlighting this period as an optimal phenological window for classification of irrigation from
rainfed in Upper Blue Nile Basin. The GTB model depends strongly on February moisture and
23
vegetation indicators, particularly NDMI, whereas RF model feature importance is more balanced
and more evenly across multiple spectral indices and months. SVM uniquely depends on rainfall
and topographic variables beside spectral features. These distinctive features are important to
profile emphasize the value of combining models with different learning strategies in an ensemble
approach. The Multi-Model Ensemble Mapping approach had a requirement of consensus from at
least three out of the four classifiers before the pixel could be classified as irrigated. It produced
relatively conservative consensus value of 588,545 ha.
Mapping irrigated land in the Upper Blue Nile Basin is challenging because the majority of the
irrigated fields are small and located in different parts of the basin. And the lack of trustworthy
records of irrigation makes it even tougher to tell how much land is actually irrigated. In this study
we show that combining machine learning algorithms with satellite data can be used to identify
these fields and increase our understanding of where irrigation is occurring. The irrigation map
generated in this study can be used to track irrigation expansion and for providing information for
estimating irrigation water demand. It can also help water resource managers pinpoint potential
problem areas in planning irrigation development and water resource allocation.
Author contributions
A.A.H. contributed to conceptualization, methodology, data collection and processing, model
development, data analysis, validation, visualization, and writing the original manuscript. Z.A.D.,
S.T.G., and H.B.L. supervised the research and contributed to reviewing and editing the
manuscript. T.G.H. contributed to data collection and preparation. All authors reviewed and
approved the final manuscript.
Funding
This research has been produced with the financial assistance of the European Union (Contract
number: AFS2023/444-387). The contents of this document are the sole responsibility of authors
and can under no circumstances be regarded as reflecting the position of the European Union.
Data availability
24
The Sentinel-1 SAR, Sentinel-2 surface reflectance, CHIRPS rainfall, and NASADEM elevation
datasets used in this study were accessed through Google Earth Engine and are publicly available
through the following links:
https://developers.google.com/earth-engine/datasets/catalog/COPERNICUS_S1_GRD
https://developers.google.com/earth-
engine/datasets/catalog/COPERNICUS_S2_SR_HARMONIZED
https://developers.google.com/earth-engine/datasets/catalog/UCSB-CHG_CHIRPS_PENTAD
https://developers.google.com/earth-engine/datasets/catalog/NASA_NASADEM_HGT_001
The processed datasets and irrigation maps generated during this study are available from the
corresponding author upon reasonable request.
Declarations
Ethical approval
Not applicable. This study did not involve human participants or animals.
Consent to participate
Not applicable. This study did not involve human participants.
Consent to publish
Not applicable. This manuscript does not contain identifiable personal information requiring
consent for publication.
Clinical trial registration
Not applicable. This study did not involve a clinical trial.
Competing interests
The authors declare no competing interests.
25
References
[1] E. A. Abioye et al., “Precision Irrigation Management Using Machine Learning and
Digital Farming Solutions,” AgriEngineering, vol. 4, no. 1, pp. 70–103, Feb. 2022, doi:
10.3390/agriengineering4010006.
[2] P. Doungmanee, “The nexus of agricultural water use and economic development level,”
Kasetsart Journal of Social Sciences, vol. 37, no. 1, pp. 38–45, Jan. 2016, doi:
10.1016/j.kjss.2016.01.008.
[3] C. Massari et al., “A Review of Irrigation Information Retrievals from Space and Their
Utility for Users,” Remote Sens. (Basel)., vol. 13, no. 20, p. 4112, Oct. 2021, doi:
10.3390/rs13204112.
[4] D. Ketchum, K. Jencso, M. P. Maneta, F. Melton, M. O. Jones, and J. Huntington,
“IrrMapper: A Machine Learning Approach for High Resolution Mapping of Irrigated
Agriculture Across the Western U.S.,” Remote Sens. (Basel)., vol. 12, no. 14, p. 2328, Jul.
2020, doi: 10.3390/rs12142328.
[5] D. G. Eshete, B. G. Sinshaw, and K. G. Legese, “Critical review on improving irrigation
water use efficiency: Advances, challenges, and opportunities in the Ethiopia context,”
Water-Energy Nexus, vol. 3, pp. 143–154, 2020, doi: 10.1016/j.wen.2020.09.001.
[6] T. Conlon, C. Small, and V. Modi, “A Multiscale Spatiotemporal Approach for
Smallholder Irrigation Detection,” Frontiers in Remote Sensing, vol. 3, Apr. 2022, doi:
10.3389/frsen.2022.871942.
[7] M. Ozdogan, Y. Yang, G. Allez, and C. Cervantes, “Remote Sensing of Irrigated
Agriculture: Opportunities and Challenges,” Remote Sens. (Basel)., vol. 2, no. 9, pp.
2274–2304, Sep. 2010, doi: 10.3390/rs2092274.
[8] A. A. Mohammedshum, B. H. P. Maathuis, C. M. Mannaerts , and D. Teka, “Mapping
Small-Scale Irrigation Areas Using Expert Decision Rules and the Random Forest
Classifier in Northern Ethiopia,” Remote Sens. (Basel)., vol. 15, no. 24, p. 5647, Dec.
2023, doi: 10.3390/rs15245647.
[9] Y. T. Dile et al., “Advances in water resources research in the Upper Blue Nile basmin
and the way forward: A review,” J. Hydrol. (Amst)., vol. 560, pp. 407–423, May 2018,
doi: 10.1016/j.jhydrol.2018.03.042.
[10] B. G. Tikuye , M. Rusnak, B. R. Manjunatha, and J. Jose, “Land Use and Land Cover
Change Detection Using the Random Forest Approach: The Case of The Upper Blue Nile
River Basin, Ethiopia,” Global Challenges, vol. 7, no. 10, Oct. 2023, doi:
10.1002/gch2.202300155.
26
[11] A. Yimere and E. Assefa, “ASSESSING AND MAPPING IRRIGATION POTENTIAL
IN THE ABBAY RIVER BASIN, ETHIOPIA,” Russ. J. Agric. Socioecon. Sci., vol. 114,
no. 6, pp. 97–109, Jun. 2021, doi: 10.18551/rjoas.2021-06.11.
[12] M. F. A. Vogels, S. M. de Jong, G. Sterk, H. Douma, and E. A. Addink, “Spatio-Temporal
Patterns of Smallholder Irrigated Agriculture in the Horn of Africa Using GEOBIA and
Sentinel-2 Imagery,” Remote Sens. (Basel)., vol. 11, no. 2, p. 143, Jan. 2019, doi:
10.3390/rs11020143.
[13] J. M. Deines, A. D. Kendall, M. A. Crowley, J. Rapp, J. A. Cardille, and D. W. Hyndman,
“Mapping three decades of annual irrigation across the US High Plains Aquifer using
Landsat and Google Earth Engine,” Remote Sens. Environ., vol. 233, p. 111400, Nov.
2019, doi: 10.1016/j.rse.2019.111400.
[14] Q. Gao, M. Zribi, M. J. Escorihuela, N. Baghdadi, and P. Q. Segui, “Irrigation Mapping
Using Sentinel-1 Time Series at Field Scale,” Remote Sens. (Basel)., vol. 10, no. 9, p.
1495, Sep. 2018, doi: 10.3390/rs10091495.
[15] A. K. Yimer, A. T. Haile, S. D. Hatiye, S. Ragettli, and M. T. Taye, “Comparative
evaluation of the accuracy of mapping irrigated areas using sentinel 1 images in the Bilate
and Gumara watersheds, Ethiopia,” Cogent Eng., vol. 11, no. 1, Dec. 2024, doi:
10.1080/23311916.2024.2357728.
[16] A. Joshi, B. Pradhan, S. Gite, and S. Chakraborty, “Remote-Sensing Data and Deep-
Learning Techniques in Crop Mapping and Yield Prediction: A Systematic Review,”
Remote Sens. (Basel)., vol. 15, no. 8, p. 2014, Apr. 2023, doi: 10.3390/rs15082014.
[17] T. Lemann, V. Roth, G. Zeleke, A. Subhatu, T. Kassawmar, and H. Hurni, “Spatial and
Temporal Variability in Hydrological Responses of the Upper Blue Nile basin, Ethiopia,”
Water (Basel)., vol. 11, no. 1, p. 21, Dec. 2018, doi: 10.3390/w11010021.
[18] E. Garzanti, S. Andò, G. Vezzoli, A. Ali Abdel Megid, and A. El Kammar, “Petrology of
Nile River sands (Ethiopia and Sudan): Sediment budgets and erosion patterns,” Earth
Planet. Sci. Lett., vol. 252, no. 3–4, pp. 327–341, Dec. 2006, doi:
10.1016/j.epsl.2006.10.001.
[19] Z. T. Segele and P. J. Lamb, “Characterization and variability of Kiremt rainy season over
Ethiopia,” Meteorology and Atmospheric Physics, vol. 89, no. 1–4, pp. 153–180, Jun.
2005, doi: 10.1007/s00703-005-0127-x.
[20] P.- E. Mellander, S. G. Gebrehiwot, A. I. Gärdenäs, W. Bewket, and K. Bishop, “Summer
Rains and Dry Seasons in the Upper Blue Nile Basin: The Predictability of Half a Century
of Past and Future Spatiotemporal Patterns,” PLoS One, vol. 8, no. 7, p. e68461, Jul. 2013,
doi: 10.1371/journal.pone.0068461.
27
[21] G. G. Chakilu , S. Sándor, T. Zoltán, and K. Phinzi, “Climate change and the response of
streamflow of watersheds under the high emission scenario in Lake Tana sub-basin, upper
Blue Nile basin, Ethiopia,” J. Hydrol. Reg. Stud., vol. 42, p. 101175, Aug. 2022, doi:
10.1016/j.ejrh.2022.101175.
[22] U. Kim and J. J. Kaluarachchi, “Climate Change Impacts on Water Resources in the
Upper Blue Nile River Basin, Ethiopia 1,” JAWRA Journal of the American Water
Resources Association, vol. 45, no. 6, pp. 1361–1378, Dec. 2009, doi: 10.1111/j.1752-
1688.2009.00369.x.
[23] S. G. Setegn, D. Rayner, A. M. Melesse, B. Dargahi, and R. Srinivasan, “Impact of
climate change on the hydroclimatology of Lake Tana Basin, Ethiopia,” Water Resour.
Res., vol. 47, no. 4, Apr. 2011, doi: 10.1029/2010WR009248.
[24] D. A. Malede, T. A. Agumassie, J. R. Kosgei, T. G. Andualem, and I. Diallo, “Recent
Approaches to Climate Change Impacts on Hydrological Extremes in the Upper Blue Nile
Basin, Ethiopia,” Earth Systems and Environment, vol. 6, no. 3, pp. 669–679, Sep. 2022,
doi: 10.1007/s41748-021-00287-6.
[25] Y. Deng, J. P. Wilson, and J. C. Gallant, “Terrain Analysis,” in The Handbook of
Geographic Information Science, Wiley, 2007, pp. 417–435. doi:
10.1002/9780470690819.ch23.
[26] E. Beriaux, A. Jago, C. Lucau- Danila, V. Planchon, and P. Defourny, “Sentinel-1 Time
Series for Crop Identification in the Framework of the Future CAP Monitoring,” Remote
Sens. (Basel)., vol. 13, no. 14, p. 2785, Jul. 2021, doi: 10.3390/rs13142785.
[27] M. Drusch et al., “Sentinel-2: ESA’s Optical High-Resolution Mission for GMES
Operational Services,” Remote Sens. Environ., vol. 120, pp. 25–36, May 2012, doi:
10.1016/j.rse.2011.11.026.
[28] M. Mohammed, B. Birhanu, and F. Abegaz, “Multi-spectral remote sensing for current
irrigated area mapping of the Rift Valley Lakes Basin in Ethiopia,” AQUA — Water
Infrastructure, Ecosystems and Society, vol. 72, no. 12, pp. 2396–2407, Dec. 2023, doi:
10.2166/aqua.2023.413.
[29] L. Breiman, “Random Forests,” Mach. Learn., vol. 45, no. 1, pp. 5–32, Oct. 2001, doi:
10.1023/A:1010933404324.
[30] M. Pal, “Random forest classifier for remote sensing classification,” Int. J. Remote Sens.,
vol. 26, no. 1, pp. 217–222, Jan. 2005, doi: 10.1080/01431160412331269698.
[31] C. Cortes and V. Vapnik, “Support-vector networks,” Mach. Learn., vol. 20, no. 3, pp.
273–297, Sep. 1995, doi: 10.1007/BF00994018.
28
[32] J. H. Friedman, “Greedy function approximation: A gradient boosting machine.,” The
Annals of Statistics, vol. 29, no. 5, Oct. 2001, doi: 10.1214/aos/1013203451.
[33] A. Natekin and A. Knoll, “Gradient boosting machines, a tutorial,” Front. Neurorobot.,
vol. 7, 2013, doi: 10.3389/fnbot.2013.00021.
[34] S. Hochreiter and J. Schmidhuber, “Long Short-Term Memory,” Neural Comput., vol. 9,
no. 8, pp. 1735–1780, Nov. 1997, doi: 10.1162/neco.1997.9.8.1735.
[35] A. Graves and J. Schmidhuber, “Framewise phoneme classification with bidirectional
LSTM and other neural network architectures,” Neural Networks, vol. 18, no. 5–6, pp.
602–610, Jul. 2005, doi: 10.1016/j.neunet.2005.06.042.
[36 P. Schratz, J. Muenchow, E. Iturritxa, J. Richter, and A. Brenning, “Hyperparameter
tuning and performance assessment of statistical and machine-learning algorithms using
spatial data,” Ecol. Modell., vol. 406, pp. 109–120, Aug. 2019, doi:
10.1016/j.ecolmodel.2019.06.002.
[37] M. U. Akbar et al., “Multi-model ensemble mapping of irrigated areas using remote
sensing, machine learning, and ground truth data,” Agric. Water Manag., vol. 312, p.
109416, May 2025, doi: 10.1016/j.agwat.2025.109416.
[38] Y. Lu, W. Song, Z. Su, J. Lü, Y. Liu, and M. Li, “MAPPING IRRIGATED AREAS
USING RANDOM FOREST BASED ON GF-1 MULTI-SPECTRAL DATA,” The
International Archives of the Photogrammetry, Remote Sensing and Spatial Information
Sciences, vol. XLIII-B2-2020, pp. 697–702, Aug. 2020, doi: 10.5194/isprs-archives-
XLIII-B2-2020-697-2020.
[39] J. Magidi, L. Nhamo, S. Mpandeli, and T. Mabhaudhi , “Application of the Random Forest
Classifier to Map Irrigated Areas Using Google Earth Engine,” Remote Sens. (Basel)., vol.
13, no. 5, p. 876, Feb. 2021, doi: 10.3390/rs13050876.
[40] S. Shi, Q. Liu, and Z. Yan, “Crop Type Mapping in an Irrigation District Using Multi-
Source Remote Sensing and LSTM-Based Time Series Analysis,” Agriculture, vol. 16, no.
9, p. 920, Apr. 2026, doi: 10.3390/agriculture16090920.
[41] D. R. Roberts et al., “Cross‐validation strategies for data with temporal, spatial,
hierarchical, or phylogenetic structure,” Ecography, vol. 40, no. 8, pp. 913–929, Aug.
2017, doi: 10.1111/ecog.02881.
[42] Y. G. Mekonnen, T. Alamirew, A. D. Chukalla, A. T. Hunegnaw, and D. A. Malede,
“Comparison of Google Earth Engine Machine Learning Algorithms for Mapping
Smallholder Irrigated Areas in a Mountainous Watershed, Upper Blue Nile Basin,
29
Ethiopia,” Journal of the Indian Society of Remote Sensing, vol. 52, no. 3, pp. 569–582,
Mar. 2024, doi: 10.1007/s12524-024-01846-w.
[43] Y. Gebeyehu, “Irrigation Potential Mapping using GIS-based Parametric Evaluation
Approach in Upper Blue Nile River Basin, Ethiopia ,” 2018.
[44] A. Yimere and E. Assefa, “Current and Future Irrigation Water Requirement and Potential
in the Abbay River Basin, Ethiopia,” Air, Soil and Water Research, vol. 15, Jan. 2022,
doi: 10.1177/11786221221097929.
论文原文与出版信息
Research Square