Holistic Response Lifting for Weakly Supervised Land-Cover Classification

This article proposes a holistic response lifting (HoLift) scheme to tackle the issue of region ambiguity in class activation maps (CAMs) for weakly supervised remote sensing land-cover classification (WSLC). In general, the ambiguously responding zones in CAMs are annotated as the background in the pseudolabel for object-centric natural images. However, it is infeasible to postprocess the region ambiguity in CAMs at the stage of computing pseudolabels for WSLC. As different land covers are seamlessly connected without background definition in satellite images, treating ambiguous responses as the background will destroy the stuff-level continuity in land-cover pseudolabels. Therefore, the ambiguous zones must be directly eliminated from land-cover CAMs rather than classified as the background in pseudolabels. To this end, our HoLift scheme is designed for the following: 1) force seamless area responses in CAMs with a holistic continuity lifting loss and 2) reduce response ambiguity in border areas via a cross-image global consistency constraint. Specifically, the pixel-level holistic continuity lifting loss is applied to multiclass CAMs to promote high total responses, reducing the ambiguous areas that lead to discontinuous pseudolabels. Meanwhile, the cross-image global consistency constraint facilitates the learning of comprehensive land-cover features by optimizing the distribution of regional features across images, thereby improving classification reliability near land-cover boundaries. With HoLift, we obtain CAMs free of large-area ambiguous transition zones. Extensive experiments reveal that our HoLift scheme significantly improves recall of CAMs by at least 16.08% and achieves state-of-the-art performance on two challenging datasets. Our code is available at https://github.com/Geo-Tell/HoLift