Sequential Feature Selection for Efficient Landslide Segmentation from Multi-Spectral Data

Researchers present a Sequential Forward Floating Selection (SFFS) framework for identifying the minimal set of satellite imagery channels needed for accurate landslide detection, demonstrating that 8 carefully selected channels match or exceed the performance of models using 30 channels. The work addresses computational efficiency and model interpretability in Earth observation machine learning by moving beyond conventional approaches that simply include all available data.

Machine learning models in Earth observation typically suffer from feature bloat—the practice of feeding every available satellite band and computed index into deep learning architectures with the assumption that models will learn to ignore irrelevant inputs. This research challenges that paradigm by systematically pruning spectral and topographic data for landslide detection. The authors employ Sequential Forward Floating Selection, a more sophisticated alternative to simple drop-tests that captures feature interactions rather than evaluating channels in isolation. Using the Landslide4Sense benchmark combining Sentinel-2 multispectral imagery and ALOS PALSAR terrain data, they construct a lightweight U-Net++ proxy model to iteratively test candidate feature subsets, ultimately identifying an 8-channel configuration that matches larger models' performance. The significance extends beyond efficiency gains. Redundant or correlated inputs obscure which physical phenomena actually drive model predictions, making results difficult to interpret for domain scientists. The Hughes Phenomenon—where excessive features degrade rather than improve classifier performance—represents a real risk in high-dimensional satellite datasets. This work demonstrates that principled feature selection improves both computational overhead and model transparency. For Earth observation practitioners, the framework provides a replicable methodology for input design that prioritizes interpretability and efficiency. The research suggests that current industry practice of appending all available bands may actually handicap model understanding and deployment scalability, particularly relevant as satellite constellations generate increasingly voluminous data streams.

• →Sequential Forward Floating Selection identified an 8-channel subset matching 30-channel model performance in landslide segmentation tasks.
• →Feature interaction effects are critical—conventional single-band drop tests miss dependencies that SFFS captures through iterative selection and pruning.
• →Redundant spectral inputs obscure physical interpretability and trigger the Hughes Phenomenon, actively degrading model performance rather than improving it.
• →Lightweight proxy models enable computationally tractable feature selection for Earth observation without training full production architectures.
• →Principled input design improves both operational efficiency and scientific transparency compared to appending all available satellite bands.