Multi-crop identification using semi-supervised deep learning framework

Abstract

Accurate crop classification from satellite imagery is essential for agricultural monitoring, yet mapping the crop type in data-scarce regions remains challenging due to limited ground truth, spectral similarities between crops, and strong temporal variability.  Supervised deep learning approaches are further constrained by their dependence on large annotated datasets. In this study, a semi-supervised framework based on R3Net, a residual–recurrent refinement network, was developed to leverage both optical and SAR signals for pixel-level segmentation of Rabi season crops (wheat, mustard, and other) using multi-temporal satellite imagery. The architecture integrates parallel temporal encoders with recurrent units to capture spatial–spectral features and phenological dynamics across key growth stages. A confidence-aware pseudo-labeling strategy was introduced to utilize large volumes of unlabeled data while suppressing uncertain predictions. The framework was evaluated against three baselines: fully supervised training, standard pseudo-labeling, and the Mean Teacher method. The supervised model achieved 65.21% accuracy and a mean IoU of 48.55%, whereas semi-supervised approaches significantly improved performance. The confidence-aware R3Net achieved the best results, with 76.05% accuracy and a mean IoU of 59.03%, including notable improvements in mustard segmentation. Testing on an independent geographic region (IARI fields, 2022–23) further demonstrated robust generalization, producing segmentation maps consistent with field boundaries and crop phenology.