Abstract

Assessing land suitability is key to agriculture management, as it helps identify the capability of different lands to cultivate different crops. Here, we present DeepS^3, a multilayer perceptron framework that simultaneously predicts the land suitability for multiple crops based on satellite imagery, crop-specific farm location data, and district level crop production census data. Our method incorporates semi-supervised and multi-task learning. Semi-supervised learning accommodates data from different spatial resolutions. Multi-task learning enables a multivariate model in the presence of unobserved responses. We exploit the multi-crop response during training and capture the interdependencies among different crops, thereby facilitating extrapolation. This framework easily generalizes to other spatial problems that incorporate data obtained at diverse spatial resolutions. Applying DeepS^3 to predict agriculture land suitability for Canada under climate change projects diminishing suitability for canola, peas, wheat, and soy in the Prairie Provinces, mainly driven by increased heat stress.