Remote sensing-based rice mapping in Brazil: Identifying the best approach for segmenting different spectral compositions using deep learning

This study explored the mapping of irrigated rice fields using remote sensing data and deep learning techniques, focus- ing on the evaluation of various spectral band combinations and polarizations from Sentinel-1 and Sentinel-2 satellites. Three deep learning models, called UNET, FAPNET, and PLANET, were implemented to perform segmentation of rice fields within a region located in southern Brazil. Specifically, the FAPNET model outperformed others when using vegetation-related spectral bands (NIR, RED), while the PLANET model demonstrated greater efficacy with water-related bands (SWIR, VH, VV). However, PLANET struggled with multi-band configurations, and its slower convergence indicated the need for refined training strategies. The integration of optical and SAR data did not lead to significant performance improvements for these models, suggesting that their architectures are limited in processing more than three input channels. In contrast, the UNET model exhibited greater robustness when handling diverse data combinations, achieving balanced performance even with the integration of optical and SAR data. This suggests that while FAPNET and PLANET specialize in extracting features from specific spectral bands, UNET is more adaptable to multi-source inputs. These findings highlight the role of thoughtful model and data choice, illustrating that special- ized structures perform well with particular data setups, whereas more generalized models are superior at synthesizing various data sources. Future research should focus on enhancing PLANET’s multi-band capabilities and improving FAPNET’s sensitivity to SWIR, advancing segmentation precision across a broader range of spectral profiles. This study contributes to the field of crop mapping through remote sensing by providing evidence that indiscriminate data fusion is not always the optimal approach, advocating for model and spectral band choices tailored to the specific application requirements.