An open-set domain adaptation approach for cross-region crop mapping

View at publisher Summarise with VIRAT
Authors: Yijia Xu, Zhou Zhang
Cited by: 0
References: 52

Summary

No summary yet. VIRAT can read this paper and write a plain-language summary of its aim, methods, key findings and limitations.

Summarise with VIRAT Free to try with your Science Catalogs account.

Abstract

The increasing availability of satellite data and advances in deep learning (DL) have facilitated more accurate crop type mapping using satellite imagery time series (SITS). However, DL models trained in one region often fail to generalize across regions due to differences in geographic and ecological conditions. While unsupervised domain adaptation (UDA) methods have been proposed to mitigate temporal-spectral shifts, they typically assume that the source and target domains share an identical set of crop types. In practice, crop type distributions often differ across regions due to varying environmental conditions and agricultural practices, leading to label space mismatch and limiting the effectiveness of conventional UDA approaches. To address this issue, this study formulated cross-region crop mapping as an open-set domain adaptation (OSDA) problem, aiming to align the shared crop types between regions while identifying unknown samples unique to the target region. Specifically, we proposed OpenMatch, a novel OSDA framework that integrates self-training for shared-class adaptation with adversarial learning to separate and reject unknown classes. The study area covers the Corn Belt and Great Plains of the United States, divided into two regions with different climatic conditions to evaluate cross-region generalization. The dataset comprises Sentinel-2 image time series and reference data from the Crop Sequence Boundaries (CSB) provided by the U.S. Department of Agriculture (USDA) from 2019 to 2022. Experimental results demonstrate that OpenMatch outperforms several state-of-the-art UDA methods in most transfer cases, achieving improvements of 37.29% in OA and 0.1919 in macro F1 score over direct transfer. It effectively mitigates over-rejection of unknown classes, especially when direct transfer perform poorly. These results highlight the potential of OSDA techniques to address label set discrepancies and enhance the scalability of DL-based crop mapping across diverse agricultural regions. The source code is available at: https://github.com/YXu556/OpenMatch.

Fields of Study & Subjects

Remote Sensing in Agriculture Remote-Sensing Image Classification Smart Agriculture and AI

Publication details

Journal / source
International Journal of Remote Sensing
Volume
47
Issue
10
Pages
4304–4326
Publication date
26 March 2026
Language
EN