Assessing the Potentials of Multi-temporal Sentinel-1 SAR Data for Paddy Yield Forecasting Using Artificial Neural Network

Accurate yield estimation of paddy crop plays an important role in forecasting paddy productivity for ensuring regional or national food security of the country. Although the crop growth models provide accurate yield forecasting, these models are difficult to implement in developing countries like India due to inhomogeneous or/and lack of required information about crop, soil, weather, etc. On the contrary, remotely sensed imagery available homogeneously provides valuable inputs for this purpose. Particularly, synthetic aperture radar (SAR) data proved to have great potential for paddy growth monitoring and biophysical parameters retrieval over optical data. Moreover, the effective use of artificial neural network (ANN) may enable us to understand the complex relation between parameters as well as improve the forecasting performance than using empirical-/semiempirical-based approaches. Thus, the study aims to analyze multi-temporal dual-polarization C-band Sentinel-1 SAR data for paddy yield forecasting using ANN model. In this study, smart sampling based on the normalized difference vegetation (NDVI) and normalized difference water index has been considered to obtain in situ yield measurements in the study area. The peak stage signature of backscattering coefficients is considered to estimate yield due to the maximum possibility of signal to interact with crop cover characteristics. It is observed that the VH-polarization-based ANN model provides better accuracy with coefficient of determination (R-2) and root mean square error (RMSE) of 0.72 and 600.11 kg/ha, respectively, in comparison with VV polarization which has shown 0.26 and 948.46 kg/ha, respectively. Overall, the study demonstrates that the effective use of ANN model may provide reliable yield estimation accuracy from remotely sensed imagery alone.