Integrating Dual-pol SAR parameters and Multi-spectral vegetation indices for Wheat crop classification in fragmented land parcels

Mapping crop inventory in fragmented and heterogeneous landscapes from either SAR or optical sensors is a major challenge due to limited information when single data source is used. Remote sensing science and evidence indicates that the phenological features can be identified by synergistic use of SAR and optical data, specifically for larger mapping extent. This study aims to explore the combination of Sentinel-1 dual-pol SAR polarimetric parameters and Sentinel-2 optical vegetation indices for binary classification of wheat crop parcels at spatially large scale in central India. The dual-pol SAR parameters used in this study are the radar backscatter, Stokes parameters and radar vegetation indices. Random Forest (RF) classification is deployed for feature selection and accuracy assessment of wheat versus non-wheat pixels. Among the dual-pol SAR parameters, the combination of first two Stokes parameters g 0 and g 1 demonstrates highest overall accuracy (OA) of 81.45%. In optical descriptors, a combination of Difference Vegetation Index of GREen (DVIGRE), Green Normalized Difference Vegetation Index (GNDVI) achieved highest OA of 83.66%. However, OA improves to 84.63% with the fusion of SAR-based g 0 and g 1 , and optical-based GNDVI and DVIGRE parameters. Additionally, this study investigates the impact of field parcel-size on classification accuracy assessment distinguishing use of optical and/or SAR satellite data. The SAR-parameter based classification outperforms for fields with parcel size greater than 3 ha, while for smaller fields, optical as well as fusion based classification works better. Using SAR data, the model achieves better accuracy with an improved Kappa value of 0.127 during the initial growth stages (BBCH 00-49), whereas fusion provides better accuracy during later growth stages (BBCH 50-99) with an improved Kappa value of 0.017 for 2023–2024.