Field-Scale Estimation and Comparison of the Sugarcane Yield from Remote Sensing Data: A Machine Learning Approach

The present work describes about estimation of crop yield of sugarcane crop from medium resolution LANDSAT 8 OLI (30 m) imageries by the development of a nonlinear empirical model using classical artificial neural networks. Sugarcane crop attributes were retrieved from high-resolution (30 m) satellite imageries to develop yield prediction models. The feed-forward back-propagation neural network algorithm developed and calibrated using the remote sensing retrieved crop parameters and ground truth data in MATLAB environment. The perceptron was trained with 75 out of the 100 possible inputs upto10,000 epochs with 1–10 hidden neurons. Four performance indices: coefficient of determination (R2), root mean squared error (RMSE), mean absolute error (MAE) and the average ratio of estimated yield to target crop yield (Rratio) were calculated, to achieve optimum neural network. Several runs were performed in determining the optimum number of hidden neurons. The best performance of the models was observed at i + 1 and i + 2 hidden nodes (i = No of input parameters). The range of R2 values of best performed models were between 0.867 and 0.916 for training and same for testing it ranged from 0.829 to 0.991 and Rratio values from 0.997 to 1.006, the normalized RMSE values ranged from 0.066 to 0.150; MAE ranged from 0.034 to 0.119 for training and 0.017–0.184 for testing. The statistical analysis recommends the reliability of the ANN model in sugarcane yield estimation. Multivariate linear regression was also performed for training and testing data separately to test the superiority of the ANN model. The estimated yield was in the range of 60,000 kg/ha–1,30,000 with an average of 73,000 kg/ha.