Multi-season rice mapping using deep learning models with multitemporal Sentinel-1 SAR data in the Kuttanad Delta, Kerala

Abstract

Timely and precise monitoring of rice paddies is essential for sustaining production, ensuring food security, and addressing climate challenges, as rice is a significant contributor to greenhouse gas emissions. Accurate rice mapping, facilitated by Sentinel-1 SAR, unaffected by weather is used in Machine learning (ML) and Deep learning (DL) models for multiclass classification of rice cropping seasons by analysing temporal backscatter patterns. A modified Dual-Branch BiLSTM model is developed to capture VH backscatter variations across the homogeneous and heterogeneous rice-growing landscapes. The study compares the performance of ML models, Random Forest (RF) and Support Vector Machine (SVM), with DL models, BiGRU and BiLSTM-BiGRU, for mapping Rabi, Kharif, double-cropping rice fields, and non-rice areas in the Kuttanad Delta region. A thorough evaluation of the proposed models was conducted using metrics like Precision, Recall, and F1 Score to assess their effectiveness. The results show that the Modified Dual-branch BiLSTM model attains F1 scores as high as 0.97 in homogeneous regions and 0.94 in heterogeneous rice-growing landscapes, highlighting its robustness and strong generalisation in mapping rice in varied landscape areas, particularly in the cloudy tropical and subtropical regions where optical data are often not consistently available during the rice cultivation season. © 2025 Informa UK Limited, trading as Taylor & Francis Group.