Prediction of rice yield using sensors mounted on unmanned aerial vehicle

Accurate estimation of rice yield helps farmers in optimizing the risk and agronomic management practices. Conventional methods for assessing yield are time-consuming, labour-intensive and costly. Multispectral imageries captured by sensors mounted on unmanned aerial vehicles (UAVs) offer a cost-effective and efficient alternative. This study aimed to develop rice yield predictive machine learning models using vegetation indices (VIs) derived from multispectral imagery. The experiment was conducted in eastern region of India, with different rice cultivars and nitrogen levels to create the yield variability. Ten VIs were generated using imagery captured during panicle initiation stage of rice. Using variance inflation factor (VIF) technique, four VIs were selected for model building. Three models: random forest (RF), support vector machine (SVM) and artificial neural network (ANN) were built using rice yield as the target variable and vegetation indices (VIs) as predictors. The accuracy of SVM in predicting rice yield was higher as compared to RF and ANN. The predicted yield (t/ha) in the SVM model ranged from 3.73 to 5.45 (R² = 0.62), while it ranged from 3.83 to 5.00 for RF model (R² = 0.57) and 3.46 to 5.91 for ANN model (R² = 0.54). Notably, the normalized difference vegetation index and transformed chlorophyll absorption reflectance index were identified as significant contributors for model building. These findings demonstrate that the proposed approach can improve the prediction accuracy of rice grain yield.