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

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

Problem Statement
Landslide detection using deep learning on satellite imagery has increasingly relied on large, highly correlated input feature sets, typically combining raw Sentinel-2 and ALOS PALSAR bands with numerous engineered spectral and structural indices. The authors argue that this practice of appending every available band and derived index lacks physical interpretability, inflates computational overhead, and risks degrading model performance due to the Hughes phenomenon (where increasing input dimensionality relative to training data size reduces accuracy). Furthermore, conventional single-band ablation studies fail to capture interaction effects between channels, leaving the specific physical cues driving model predictions poorly understood.

Methodology
The study proposes a systematic, explainable channel-selection framework applied to the Landslide4Sense benchmark dataset, which contains 3,799 multispectral patches across four distinct geographic regions. The methodology involves the following components:

1. Input Space: The candidate pool consists of 30 channels: the 14 native Sentinel-2/ALOS PALSAR bands plus 16 engineered features (including vegetation indices like NDVI, moisture indices like NDMI, soil exposure indices like NBR, and structural features like edge detectors and gradients).
2. Proxy Model: A lightweight U-Net++ architecture with a ResNet-50 backbone is used as the proxy model for all selection experiments.
3. Selection Algorithm: The core of the framework is Sequential Forward Floating Selection (SFFS). Unlike greedy forward selection or single-band drop tests, SFFS iteratively adds the feature that maximizes validation performance and subsequently performs a "floating" backward elimination step. This allows the algorithm to remove previously selected features if they become redundant in the presence of new additions, effectively capturing non-linear interactions and redundancy.
4. Training Protocol: To address extreme class imbalance (where landslide pixels constitute only ~2.5% of the data), the authors employ class-weighted Binary Cross-Entropy combined with Dice loss, class-balanced sampling, and hard-negative mining. Thresholds are optimized via a validation sweep (0.3–0.8) to maximize F1 scores.
5. Attribution Analysis: Following the selection of the optimal subset, Permutation Importance tests are conducted. This involves shuffling individual band values to measure the resulting drop in F1 score, quantifying the relative contribution of each retained channel.

Key Contributions

• Systematic Feature Selection: The paper introduces a principled approach to input design in Earth observation, moving away from the "append-all" strategy toward a data-driven selection of complementary features.
• Compact Subset Identification: The framework successfully identifies a minimal 8-channel subset that preserves or exceeds the performance of configurations using up to 30 channels.
• Physical Interpretability: By isolating the necessary inputs, the study interrogates which spectral and topographic features genuinely drive landslide predictions, distinguishing informative signals from noise.
• Validation of the Hughes Phenomenon: The results provide empirical evidence that adding redundant or highly correlated engineered indices can actively degrade model performance rather than improve it.

Results

• Performance: The SFFS-selected 8-channel subset achieved an F1 score of 78.32%. This outperformed the baseline of 14 raw spectral bands (77.2%), the full 30-channel engineered set (77.3%), and a referenced 23-channel configuration (78.1%).
• Selected Features: The final 8-channel subset includes:
  • Spectral Bands: B4 (Red), B5 (Red Edge 1), B8 (NIR), B11 (SWIR1), and normalized versions of B3 and B12.
  • Topographic/Structural: Slope (B13) and a Grayscale composite (Band 21).
  • Excluded Features: Most engineered spectral indices (NDVI, NDMI, NBR, SAVI, EVI, NDWI, MNDWI) and structural features (gradients, edge detectors) were consistently excluded, suggesting they are redundant with respect to the raw bands or introduce noise.
• Feature Importance: Permutation analysis revealed that B4 (Red) is the dominant contributor to the model's performance, followed by Slope and the Grayscale composite. Other retained bands (B5, B8, B11) showed moderate contributions, while DEM (B14) had a limited impact, likely due to redundancy with Slope.

Significance and Claims
The paper claims that accurate landslide segmentation does not require the full set of available raw and engineered inputs. Instead, a compact, physically interpretable subset is sufficient to achieve state-of-the-art performance. The authors argue that:

1. Efficiency: Reducing input dimensionality from 30 to 8 channels offers significant computational and storage benefits without sacrificing accuracy.
2. Interpretability: Models operating on a small number of physically meaningful channels are easier to analyze and explain, which is critical for operational disaster response.
3. Re-evaluation of Indices: The study challenges the prevailing assumption that pre-computed spectral indices provide complementary information; in this context, they often act as redundant transformations of the original bands.
4. Principled Input Design: The authors advocate for SFFS as a standard approach for input design in remote sensing, contrasting it with the current practice of hoping deep learning models can implicitly learn to ignore irrelevant inputs.

The authors remain modest regarding generalizability, noting that the study is limited to the Landslide4Sense benchmark and that cross-sensor transferability and performance across different architectures require further investigation.