Deep Learning based Paddy Land Abandonment Detection Using Multitemporal Polarimetric SAR Patterns

The abandonment of agricultural land in India, especially paddy fields, has emerged as a significant challenge for food security and ecosystem sustainability in the country. Although rice production is vital for national food security, research on paddy land abandonment in India remains limited. Some Indian states have reported an alarming decline in paddy cultivation area over the past two decades. The study employs the Udupi district of Karnataka, India, a high-rainfall coastal region where paddy has traditionally been the dominant crop and where paddy land abandonment has been observed, as the study area. This study addresses crucial research gaps by framing these objectives for the study: (1) developing a deep learning framework that utilizes both intensity and phase information from polarimetric Synthetic Aperture Radar (SAR) data for abandoned paddy land detection, (2) leveraging recurrent neural networks (RNNs) to capture temporal patterns in abandonment, and (3) demonstrating an automated, all-weather monitoring approach that overcomes the limitations of traditional optical remote sensing in tropical regions.

Conventional monitoring approaches struggle with persistent cloud cover in tropical regions which limits effective assessment of abandonment patterns. SAR data provides unique capabilities for continuous monitoring under all weather conditions, making it particularly well-suited for tropical regions. However, previous studies have primarily underutilized SAR's potential by concentrating solely on backscattering intensity from ground range detected (GRD) products, overlooking the valuable phase information that could offer deeper insights into land use changes.  In this study, we employ Sentinel-1 Single Look Complex (SLC) data, which offers both intensity and phase information. Considering the temporal nature of paddy land abandonment, we developed a deep learning framework utilizing RNNs viz. LSTM, BiLSTM and BiGRU to effectively capture time-series patterns in the data. This framework analyzes backscattering coefficients (VV and VH polarizations) and polarimetric parameters (entropy, anisotropy and alpha angle) derived from SLC data collected during the Kharif seasons from 2017 to 2024. We carried out extensive ground truth data collection of active and abandoned paddy lands to train and validate our models. The backscattering coefficients were processed through orbit correction, radiometric calibration, TOPSAR deburst, multi-looking, speckle filtering and terrain correction. For deriving the polarimetric parameters, after basic preprocessing steps, the covariance matrix was generated followed by the polarimetric decomposition of the phase-preserved data. Results indicate that our RNN models show promising performance in detecting temporal patterns of paddy land abandonment. The method exhibits a robust ability to produce reliable abandoned land maps in regions prone to cloudy and rainy conditions. Future research should explore polarimetric features across various vegetation types in abandoned lands, expand the methodology to other agricultural systems, and examine the impact of socio-economic and topographical factors on abandonment patterns to support evidence-based land management policies.