XSPECT: An Explainable SHAP-Based Spectral Band Reduction Algorithm for Land Use and Land Cover Multispectral Image Classification
IEEE Access · 2026 · 14 · 125822-125835
DOI: 10.1109/access.2026.3723991 ↗Kurz gefasst
Fachartikel in der Zeitschrift IEEE Access, 2026. Verfasst von Anastasios Temenos und 5 weiteren Autor:innen. Die Arbeit nennt 44 Literaturangaben. Der Volltext ist frei zugänglich (Open Access, CC BY 4.0). Gefördert von: European Union Funded Project NETHELIX “Intelligent Digital Toolbox Toward More Sustainable and Safer Extraction of Mineral Resources” under the Horizon 2020 P….
Abstract (Originalsprache)
A SHAP-based spectral band reduction algorithm, XSPECT, for land use land cover (LULC) remote sensing (RS) applications is introduced. XSPECT operates as a post-hoc deep learning (DL) model evaluation algorithm, relying on the computation of SHAP values to quantify the contribution of each spectral band to the classification output. Based on them, XSPECT performs class-wise band ranking, aggregation of their importance across classes via occurrence scores, identification and selection of the subset of bands contributing most significantly to classification performance, all governed by a user-defined parameter. This characterizes XSPECT as an explainability-guided approach that reduces the need for exhaustive band-permutation experiments and manual band-selection procedures. Experimental results on the EuroSAT-Multispectral and So2Sat LCZ42 datasets utilizing the XSPECT algorithm and a Vision Transformer DL model demonstrate that using fewer bands for training, it can achieve classification accuracy up to 97.62% (EuroSAT-M) and 97.07% (So2Sat LCZ42), comparable to and in some cases surpassing those obtained using all available bands (97.46% and 96.83%, respectively). Furthermore, even under strong dimensionality reduction, the classification performance remains above 96% for EuroSAT-M (4 bands) and 95% for So2Sat LCZ42 (5 bands), showcasing its effectiveness. Per-class performance analysis underscores XSPECT’s ability to identify key bands specific to each class, making it particularly well-suited for domain-specific RS applications. Comparison with popular and widely used band reduction techniques, including Principal Component Analysis (PCA) and Minimum Redundancy Maximum Relevance (mRMR), demonstrates XSPECT’s consistency across both datasets and effectiveness, particularly under aggressive band reduction, where it better preserves class-specific discriminative information.
Förderung
- European Union Funded Project NETHELIX “Intelligent Digital Toolbox Toward More Sustainable and Safer Extraction of Mineral Resources” under the Horizon 2020 P…
Zitieren
APA
Temenos, A., Temenos, N., Zafeiropoulos, C., Rallis, I., Doulamis, A., & Doulamis, N. (2026). XSPECT: An Explainable SHAP-Based Spectral Band Reduction Algorithm for Land Use and Land Cover Multispectral Image Classification. IEEE Access, 14, 125822-125835. https://doi.org/10.1109/access.2026.3723991
Vancouver
Temenos A, Temenos N, Zafeiropoulos C, Rallis I, Doulamis A, Doulamis N. XSPECT: An Explainable SHAP-Based Spectral Band Reduction Algorithm for Land Use and Land Cover Multispectral Image Classification. IEEE Access. 2026;14:125822-125835. doi:10.1109/access.2026.3723991
BibTeX
@article{Temenos2026,
author = {Temenos, Anastasios and Temenos, Nikos and Zafeiropoulos, Charalampos and Rallis, Ioannis and Doulamis, Anastasios and Doulamis, Nikolaos},
title = {XSPECT: An Explainable SHAP-Based Spectral Band Reduction Algorithm for Land Use and Land Cover Multispectral Image Classification},
journal = {IEEE Access},
year = {2026},
volume = {14},
pages = {125822--125835},
doi = {10.1109/access.2026.3723991}
}
EUVIMED