Sequential Feature Selection for Efficient Landslide Segmentation from Multi-Spectral Data
Researchers have introduced a systematic, explainable channel-selection framework to improve landslide detection from satellite imagery. While deep learning has advanced this field, existing models often rely on large, highly correlated spectral-topographic inputs, leading to computational inefficiency and reduced interpretability due to the Hughes Phenomenon. To address this, the study applies Sequential Forward Floating Selection (SFFS) using a lightweight U-Net++ proxy model on the Landslide4Sense benchmark. This approach combines Sentinel-2 multispectral data, ALOS PALSAR terrain data, and engineered indices. Unlike conventional single-band drop tests, SFFS accounts for feature interactions, identifying a compact 8-channel subset that matches or exceeds the segmentation performance of configurations using up to 30 channels. The process not only optimizes input design but also reveals the physical cues driving model predictions. The authors argue that SFFS offers a principled alternative to the common practice of including all available bands, significantly enhancing both efficiency and physical interpretability in Earth observation applications.
Wire timeline
Sequential Feature Selection for Efficient Landslide Segmentation from Multi-Spectral Data
Researchers have introduced a systematic, explainable channel-selection framework to improve landslide detection from satellite imagery. While deep learning has advanced this field, existing models often rely on large, highly correlated spectral-topographic inputs, leading to computational inefficiency and reduced interpretability due to the Hughes Phenomenon. To address this, the study applies Sequential Forward Floating Selection (SFFS) using a lightweight U-Net++ proxy model on the Landslide4Sense benchmark. This approach combines Sentinel-2 multispectral data, ALOS PALSAR terrain data, and engineered indices. Unlike conventional single-band drop tests, SFFS accounts for feature interactions, identifying a compact 8-channel subset that matches or exceeds the segmentation performance of configurations using up to 30 channels. The process not only optimizes input design but also reveals the physical cues driving model predictions. The authors argue that SFFS offers a principled alternative to the common practice of including all available bands, significantly enhancing both efficiency and physical interpretability in Earth observation applications.
cs.AI updates on arXiv.org