Multi-source Unsupervised Domain Adaptation on Corn Yield Prediction

Yuchi Ma, Zhou Zhang

Keywords: yield prediction, remote sensing, deep learning, unsupervised domain adaptation

TL;DR: corn yield prediction using multi-source unsupervised domain adaptation

Abstract: Recently, supervised machine learning methods based on remote sensing observations have achieved satisfactory results in crop yield prediction. However, supervised learning models tend to have poor transferability. Due to domain shifts between observations in different regions, models trained with data from one spatial region (i.e., source domain) often lose their validity when directly applied to another region (i.e., target domain). To address this issue, we proposed a Bayesian Multi-source Maxi-mum Predictor Discrepancy (BMMPD) neural network which is an unsupervised domain adaptation (UDA) approach to improve the model’s transferability for corn yield prediction at the county level. We proposed to maximize the discrepancy between two yield predictors’ out-puts to detect unlabeled target samples that are far from the support of the source domain. A feature extractor then learned to align source and target domains by minimizing the predictor discrepancy. Moreover, we applied Bayesian learning to prevent overfitting. A case study was conducted in the U.S. corn belt to evaluate the proposed BMMPD model. Time-series vegetation indices and weather observations were collected and aggregated to the county level and used as the input predictors. Experiment results demonstrated that the proposed BMMPD has effectively reduced domain shifts and outperformed several state-of-art domain adaptation methods.