An Explainable Human-AI in the-Loop Framework for Segmentation-Aware Hyperspectral Image Classification

Hyperspectral imaging has emerged as a pivotal technology in domains such as remotesensing, precision agriculture, environmental assessment, and intelligent monitoring systems,owing to its capacity to represent scenes with highly detailed spectral signatures. Despite notableprogress in machine learning and deep learning–based classification methods, practicaldeployment of these models remains challenging. Many existing solutions suffer from limitedtransparency, substantial computational demands, and minimal involvement of human experts. Furthermore, models trained solely on data often underperform when labeled samples are scarceand fail to leverage valuable domain knowledge. In response to these challenges, this studypresents an interpretable, human-in-the-loop framework for segmentation-driven hyperspectralimage classification. The proposed methodology integrates spectral–spatial feature aggregationwith affinity propagation–based segmentation and an efficient Extreme Learning Machineclassifier. To enhance reliability, explainability components and structured expert feedback areembedded directly into the learning pipeline. Unlike traditional automated workflows, theframework enables human validation of uncertain outcomes and supports adaptive modelrefinement. Formal mathematical expressions are provided to characterize feature integration,similarity estimation, classifier learning, and feedback-guided optimization. Comprehensive experiments conducted on standard hyperspectral benchmarksdemonstrate that the proposed framework delivers superior performance compared to conventionalclassification approaches, particularly in scenarios with limited training data. Improvements areobserved not only in classification accuracy but also in robustness and interpretability. Thesefindings underscore the value of unifying segmentation strategies, efficient learning models, andhuman-centered AI concepts to develop dependable hyperspectral classification systems. Theproposed approach offers a scalable and transparent solution suitable for real-world applicationswhere expert oversight and explainability are essential.