Publicação
An approach based on Synthetic Aperture Radar (SAR) and optical images for monitoring rice growth in mangrove swamp rice sites in Guinea-Bissau
| Resumo: | Mangrove Swamp Rice in Guinea-Bissau is crucial for food and nutritional security but faces significant challenges related to rainfall variability, soil salinity, and brackish water flooding events. Remote sensing through PlanetScope (PS) and Sentinel-1 (S1) imagery provides valuable insights into crop phenology, supporting better-informed decision-making for smallholders and research and development organizations. This study develops a spatiotemporal monitoring approach to predict sowing and harvesting dates using high-resolution optical and Synthetic Aperture Radar (SAR) data, integrating a 3D Convolutional Neural Network. Vegetation indices and polarimetric parameters were structured into temporal sequences to generate spatio-temporal monitoring maps of sowing and harvesting events. Feature selection involved correlation analysis and Principal Component Analysis to identify the most informative spectral and SAR bands. The model was trained and validated using cross-validation, with an independent test plot reserved to assess model performance. Results demonstrated that the optical sensor PS provides more precise date predictions than S1, achieving an absolute RMSE below 12.08 days for both sowing and harvesting events. Visual validation using the NDRE vegetation index confirmed the model’s ability to align predictions about sowing and harvest events. Inundation events and continuous bare soil areas were identified as key factors affecting prediction accuracy, as part of the source of error analysis. The findings highlight the potential of deep learning for crop monitoring in agricultural systems. Future research should enhance computational efficiency, explore multi-sensor data fusion, and refine the model to account for environmental disruptions such as tidal flooding. The models should also be assessed to extend their applications in more agricultural contexts. |
|---|---|
| Autores principais: | Rivera,Emmanuel Jesús Céspedes |
| Assunto: | 3D convolutional neural network synthetic aperture radar remote sensing time-series Oryza spp. rede neuronal convolucional 3D radar de abertura sintética deteção remota séries temporais |
| Ano: | 2025 |
| País: | Portugal |
| Tipo de documento: | dissertação de mestrado |
| Tipo de acesso: | acesso aberto |
| Instituição associada: | Universidade de Lisboa |
| Idioma: | inglês |
| Origem: | Repositório da Universidade de Lisboa |
| Resumo: | Mangrove Swamp Rice in Guinea-Bissau is crucial for food and nutritional security but faces significant challenges related to rainfall variability, soil salinity, and brackish water flooding events. Remote sensing through PlanetScope (PS) and Sentinel-1 (S1) imagery provides valuable insights into crop phenology, supporting better-informed decision-making for smallholders and research and development organizations. This study develops a spatiotemporal monitoring approach to predict sowing and harvesting dates using high-resolution optical and Synthetic Aperture Radar (SAR) data, integrating a 3D Convolutional Neural Network. Vegetation indices and polarimetric parameters were structured into temporal sequences to generate spatio-temporal monitoring maps of sowing and harvesting events. Feature selection involved correlation analysis and Principal Component Analysis to identify the most informative spectral and SAR bands. The model was trained and validated using cross-validation, with an independent test plot reserved to assess model performance. Results demonstrated that the optical sensor PS provides more precise date predictions than S1, achieving an absolute RMSE below 12.08 days for both sowing and harvesting events. Visual validation using the NDRE vegetation index confirmed the model’s ability to align predictions about sowing and harvest events. Inundation events and continuous bare soil areas were identified as key factors affecting prediction accuracy, as part of the source of error analysis. The findings highlight the potential of deep learning for crop monitoring in agricultural systems. Future research should enhance computational efficiency, explore multi-sensor data fusion, and refine the model to account for environmental disruptions such as tidal flooding. The models should also be assessed to extend their applications in more agricultural contexts. |
|---|