Bayesian Joint Adaptation Network for Crop Mapping in the Absence of Mapping Year Ground-Truth Samples

Crop mapping is a fundamental step for various higher level agricultural applications, such as crop yield prediction, farm management analysis, and agricultural market regulation. Recent advancements in deep learning models have greatly promoted crop mapping using satellite imagery time series (SITS), enabling frequent and extensive monitoring of croplands. However, a classifier trained on a specific year(s) with crop type labels (i.e., source domain) can exhibit reduced effectiveness when directly applied to a different year(s) without reference data (i.e., target domain) due to the interannual variation in image signals and crop growth dynamics. To address this issue, we propose an unsupervised domain adaptation (UDA) method named Bayesian joint adaptation network (BJAN), which aims to align the joint distributions of input SITS and output crop types across different years, thereby facilitating crop mapping in years without ground-truth samples. In the proposed BJAN method, Bayesian uncertainty is used to detect target data that are outside the support of the source domain. By minimizing the uncertainty on target samples, the model is trained to align the task-specific conditional distributions of source and target domains. Simultaneously, by constraining the feature distributions of source and target domains, the discrepancy of data-related marginal distributions is alleviated. Our experiments on two landcover classification datasets from the U.S. showed that BJAN has effectively aligned source and target domains and outperforms several state-of-the-art domain adaptation methods.