Superpixel-aware credible dual-expert learning for land cover mapping using historical land cover product

Abstract

One of the key solutions to the challenge of collecting training labels for high-resolution remote sensing images is to leverage prior information from historical land cover products, which includes knowledge derived from both same- and low-resolution land cover products (relative to the targeted images). However, employing these products as training labels directly fails to yield encouraging results in the pixel-level training process due to the widespread existence of complex noise labels. These noise labels can be categorized into scale-response noise labels, resulting from resolution discrepancies, and model-cognitive noise labels, caused by misclassifications from historical classification models or temporal changes. To address these noise labels, we propose employing superpixels as training units to mitigate scale-response and small-scale model-cognitive noise labels. The large-scale model-cognitive noise labels might then be adaptively optimized during the training process by integrating multi-source knowledge. Accordingly, we design a superpixel-aware credible dual-expert weakly supervised learning (SCDWSL) approach for high-resolution land cover mapping. Our method utilizes the multi-scale contextual information perception capabilities of superpixels and integrates credible assessment from dual-expert knowledge framework to hierarchically tackle various noise labels. To validate the effectiveness of SCDWSL, we conduct experiments using the WorldCover with a resolution of 10-m as labels. First, we evaluate its capacity to handle both scale-response and model-cognitive noise using the National Agricultural Imagery Program dataset and GaoFen-2 image (1-m resolution). Secondly, we assess its performance on addressing model-cognitive noise alone using Sentinel-2 data. Extensive experiments demonstrate that SCDWSL outperforms existing weakly supervised methods across three datasets, highlighting its unique advantages and applicability on large-scale land cover mapping.