Paddy characterization and yield estimation using optical and SAR data
Loading...
Date
2023-02
Authors
Journal Title
Journal ISSN
Volume Title
Publisher
G. B. Pant University of Agriculture and Technology, Pantnagar
Abstract
Reliable and accurate monitoring of rice crop is vital for food security and the global economy. This study aims to characterize rice types and provide an early yield estimation using optical and Synthetic Aperture Radar (SAR) data. Dense time series Sentinel-1 SAR backscatter data were analyzed for summer and kharif rice to discriminate between different rice types and identify their optimum stages to represent crop growth profiles. Critical stages of discrimination were determined statistically for all rice types. Knowledge-based decision tree algorithm was employed to classify all major types of rice grown during both the seasons. Savitzky-Golay fitted temporal profile of remotely sensed indices from Sentinel-1 and Sentinel-2 satellite sensors were used to extract seasonality parameters using TIMESAT. Correlation analysis was performed at different phenophases to identify sensitive biophysical and remote sensing parameters. Finally, machine learning models were used to predict yield of rice at 45, 60 and 90 days after transplanting using remote sensing indices and biophysical parameters as inputs. The performances of ML models for early yield estimation were evaluated using standardized ranking performance index. Results revealed significant differences between backscatter profiles of different rice types especially in early vegetative stages with transplanting being the most critical stage. Transplanting/seeding, tillering, panicle initiation, peak vegetative/flowering and maturity were identified to be the most important stages to represent the backscatter profile of entire rice growth. Decision tree algorithm discriminated rice types with high overall accuracy and Kappa coefficient, viz., 94.74% and 0.94 for summer rice and 91.80% and 0.90 for kharif rice, respectively. The extracted seasonality parameters revealed that start time derived from SAR indices was found to accurately distinguish different types of rice. The correlation strength generally increased with the progress of crop growth in both seasons. All significant correlations were positive except that with moisture content. NRPB was strongly correlated with VV and VV/VH ratio at all stages in both the seasons. LAI and dry biomass were found to be the most sensitive biophysical parameters towards remote sensing data. Area under rice
cultivation in the study area was found to be 37.1%, using random forest classifier that yielded an accuracy above 90% and kappa coefficient of 0.89. The machine learning models were able to achieve R2 and d-index values above 0.9 and RMSE below 0.5 t/ha. XGB was the best model for predicting rice yield using remote sensing and biophysical parameters. The findings of this study have the potential to contribute to the greater goal of sustainable agriculture and food security.