Multitemporal Machine Learning Approach for 10-m Pixel-Based Crop Classification in the Fragmented Agricultural Fields of the Nile Delta, Egypt

Abstract

Annual agricultural censuses in Egypt are often delayed, leaving crop data outdated and less reliable. This study addresses the challenge of fine-scale crop mapping in the fragmented Nile Delta, where accurate classification is vital for water resource management and food security. It uses 10-m multitemporal Sentinel-1 SAR and Sentinel-2 optical imagery, combined with 11 000 field samples from the 2021 and 2023 summer seasons, to train and independently validate models. The focus is on maize and rice, the main economically significant summer crops, together representing about 70% of summer cropping in the Nile Delta. To capture smallholder farm heterogeneity and improve on manual collection methods, a pixel-based, multisensor time-series approach was applied using spectral bands and vegetation indices linked to chlorophyll content and biomass. Six machine learning (ML) models were compared: decision tree (DT), random forest (RF), XGBoost, K-nearest neighbors, support vector classifier (SVC), and long short-term memory (LSTM). XGBoost achieved the best performance, reaching 94.3% overall accuracy on 2021 data, but dropped to 74% on independent 2023 data because phenological changes caused missing observations in June and September. DT and RF also showed strong potential for agricultural monitoring. Results were validated through area-based aggregation of predicted crop coverage and statistical correlation with official government area statistics. The approach offers rapid information for water management, agricultural policy, and crop yield estimation, and it appears suitable for integration into national agricultural monitoring systems in the Nile Delta’s smallholder-fragmented, mixed-irrigation landscape.