Machine Learning-Based Land-Cover and Crop Types Classification Using Multi-Temporal Sentinel-1 SAR Imagery in Hyper-Arid Environments
Open Access
Abstract
Egypt’s New Delta mega-project represents the largest horizontal agricultural expansion in modern Egyptian history, creating an urgent need for cost-effective land-cover and crop monitoring systems capable of operating under hyper-arid conditions. This study evaluates the operational feasibility of machine learning (ML)-based land-cover and crop classification using multi-temporal Sentinel-1 Synthetic Aperture Radar (SAR) imagery in the El-Daba Development Corridor, Egypt, a hyper-arid component of newly irrigated desert land. Sentinel-1 Ground Range Detected (GRD) data (vertical-vertical (VV) and vertical-horizontal (VH) polarizations) acquired from January to December 2025 were processed within Google Earth Engine (GEE). Monthly composite images were generated to capture phenological development, from which backscatter features, polarization ratios and differences, statistical descriptors, the Radar Vegetation Index (RVI), and Gray-Level Co-occurrence Matrix (GLCM) texture features were extracted. K-Means clustering (K = 10) provided an initial stratification of the feature space, while Random Forest (RF) classification was trained on 840 samples (70%) and validated on 358 independent samples (30%). Temporal backscatter analysis revealed distinct phenological signatures: wheat and sugar beet peaked in March (VV = −7 dB), maize reached maximum backscatter in July (VH = −12 dB, RVI = 0.70), and perennial crops maintained stable signatures year-round. Feature importance analysis identified VH in July (8.7%), VV in March (7.9%), and RVI in July (7.2%) as the most discriminative variables, confirming that combining observations from multiple phenological periods substantially enhances crop separability. The RF classifier achieved 90.5% overall accuracy with a Kappa coefficient of 0.895, with per-class accuracies ranging from 77.8% (vegetables) to 100% (non-agricultural classes). Optimal discrimination windows were identified as February-March for winter cereals, June-July for summer maize, and January/May-June for potato. These results demonstrate that SAR-only land-cover and crop mapping can deliver operationally useful accuracy in hyper-arid environments where optical imagery is constrained, providing a scalable foundation for agricultural monitoring in Egypt’s desert reclamation frontiers and comparable arid systems globally.Keywords
This publication has 89 references indexed in Scilit:
- AI-enabled predictive pipelines for early warning of agricultural pests, plant diseases, and droughtSmart Agricultural Technology, 2026
- Smart water for sustainable agriculture through climate resilient assessment and integrated soil water crop managementDiscover Water, 2026
- Field crop mapping using machine learning and multi-sensor satellite fusion: toward dynamic agricultural monitoringSmart Agricultural Technology, 2025
- Deep Learning for Time Series Forecasting: Review and Applications in Geotechnics and GeosciencesArchives of Computational Methods in Engineering, 2025
- Analysis of RVI for rice crops in small-scale agricultural fields using Sentinel-1 SAR data: case study on LAI retrieval using regression algorithmsPaddy and Water Environment, 2024
- Convergence of successive linear programming algorithms for noisy functionsComputational Optimization and Applications, 2024
- A machine learning approach for accurate crop type mapping using combined SAR and optical time series dataAdvances In Space Research, 2022
- Spatial cross-validation is not the right way to evaluate map accuracyEcological Modelling, 2021
- Application of decision tree-based ensemble learning in the classification of breast cancerComputers in Biology and Medicine, 2021
- Classification assessment methodsApplied Computing and Informatics, 2020