Evaluating the Performance of Crop Yield Forecasting Models Coupled with Feature Selection in Regression Framework

As crop yield is determined by numerous input parameters, it is important to identify the most important variables/parameters and eliminate those that may reduce the accuracy of the prediction models. The feature selection algorithms assist in selecting only those relevant features for the prediction algorithms. Instead of a complete set of features, feature subsets give better results for the same algorithm with less computational time. Feature selection has the potential to play an important role in the agriculture domain, with the crop yield depending on multiple factors such as land use, water management, fertilizer application, other management practices and weather parameters. In the present study, feature selection algorithms such as forward selection, backward selection, random forest (RF) and least absolute shrinkage and selection operator (LASSO) have been applied to three different datasets. Regression forecasting models have been developed with selected features for all the algorithms. The forecasting performance of the proposed models was compared using statistical measures such as root mean square error, mean absolute prediction error and mean absolute deviation. A comparison was made among all the feature selection algorithms. The regression models developed with LASSO, RF and backward selection algorithms were the best for different datasets.