Machine learning approaches for crop yield prediction with MODIS and weather data

For accurate prediction, many studies have been actively conducted to estimate grain crops using machine learning techniques. However, there are only few studies which compare the accuracies using many kinds of machine learning techniques for different types of crop. This study was conducted to estimate corn and soybean yields in Illinois and Iowa in the U.S. through four kinds of machine learning techniques, including deep learning algorithms. ANN (Artificial Neural Network), CNN (Convolutional Neural Network), SSAE (Stacked-Sparse AutoEncoder), and LSTM (Long-Short Term Memory) were used as prediction models, and total 14 years of MODIS (MODerate resolution Imaging Spectroradiometer) data, climatic data and crop yield statistics were used as input variables with the six different periodic scenarios. The accuracies were compared in terms of %RMSE (percentage Root Mean Square Error) and compared to the baseline prediction model, which is DT (Decision tree). As the results, CNN model was most accurate over all with the lowest average %RMSE errors in corn (9.36%) and SSAE model in soybean (10.05%) respectively. The best periodic scenario for corn yield prediction was May to September, and for soybean was June to August. This study identified suitable scenario and prediction technique for corn and soybean yield, which will be useful for farming activities and agricultural planning.