A cross-spatiotemporal weakly supervised framework for land cover classification: Generating temporally and spatially consistent land cover maps

Abstract

High-resolution land cover mapping tasks guided by publicly available decameter-level land cover products often suffer from label inaccuracies caused by land cover changes and scale discrepancies resulting from spatiotemporal resolution inconsistencies. To address this issue, this study proposes a cross-spatiotemporal weakly supervised dual-stage classification framework (CTS-WS) that implements temporal and spatial correction strategies to rectify erroneous labels and scale differences, achieving spatiotemporal consistent high-resolution land cover mapping. In the cross-temporal stage, we establish an NDVI screening and uncertainty noise correction mechanism by leveraging the spectral characteristics of high-resolution imagery and the feature fitting capability of convolutional neural networks, effectively eliminating pixels with spectral feature mismatches. The cross-spatial stage proposes a dual-branch parallel network integrating spatial and spectral features, which combines a periodic label screening module with boundary metric loss to learn fine-grained spatial features and refine boundaries. To validate the effectiveness of the proposed method, this study constructed the GF1-CTS dataset by integrating Gaofen-1 satellite imagery with ESA-GLC10 product, and conducted parallel experiments on both the GF1-CTS dataset and a large-scale Chesapeake Bay watershed dataset. Experimental results demonstrate that CTS-WS successfully achieves cross-spatiotemporal resolution land cover mapping from 10 m to 2 m and from 30 m to 1 m, outperforming various mainstream methods and state-of-the-art technologies. This study provides a novel solution for high-resolution remote sensing image land cover mapping across spatiotemporal resolutions.