Simultaneous multi-crop land suitability prediction from remote sensing data using semi-supervised learning

This study presents an artificial neural network based model that can simultaneously estimate land suitability for barley, peas, spring wheat, canola, oats, and soy in Canada leading to more accurate predictions than single-crop models. The novelties in the modelling method include using an indicator function which allows for a multivariate model to be trained, and a semi-supervised learning approach which allows for training with unlabelled data. The model performs well on land not used in the training set, as demonstrated by both K-fold cross-validation and a visual comparison of crop inventory to predicted land suitability in northern Alberta. The predicted suitability of crops correspond with a region's growing season length; this is in line with literature. Northern Canada is almost completely unused for agriculture, but this may change in the coming decades due to the climate becoming more favorable for agriculture. The model presented in this work can allow for a precise cost benefit analysis regarding environmental damage and economic benefits of cultivating new lands in Canada.