Advancing irrigated area mapping using machine learning and deep learning approaches: The case of Awash Valley, Ethiopia

Abstract

Accurate and timely data on irrigated cropland in arid lowlands requires in Ethiopia to strengthen national food security resilience. However, conventional mapping methods remain inconsistent to provide accurate information about irrigated farming practices. This study addresses these gaps by evaluating multisource Earth Observation (EO) data using advanced machine learning (such as SVM, RF) and state-of-the-art deep learning (ResUNet). Using Sentinel-2 Level 2(S2A) imagery, we designed three classification scenarios: i) six spectral bands; ii) spectral bands plus three vegetation indices, including the Enhanced Vegetation Index (EVI), Normalized Difference Vegetation Index (NDVI), and Land Surface Water Index (LSWI); and iii) an expanded feature set incorporating spectral texture metrics such as variance, contrast and dissimilarity. These models provide a robust framework for quantifying irrigated areas in the Awash basin, overcoming conventional data limitations. Model outputs were rigorously validated using ground reference data. This study evaluates the transformative potential of the ResUNet deep learning architecture for precise irrigation and land-use mapping in Ethiopia’s heterogeneous Awash basin landscape. By integrating multispectral bands, vegetation indices and spectral features, ResUNet achieved a superior Overall Accuracy of 84.13% and Kappa of 0.83, significantly outperforming SVM (67.54%) and Random Forest (82.33%). The model’s residual learning framework proved essential for capturing complex spatial-spectral hierarchies in semi-arid zones, achieving a peak F1-score of 0.89 for irrigated area detection. Beyond technical performance, this automated approach provides critical monitoring capabilities to support water resource management, drought early-warning systems, and food security forecasting for smallholder farmers.