Arbitrary Hyperspectral Images: A universal interpretation framework

Remote sensing interpretation serves as a critical technological modality for Earth observation and environmental monitoring. Deep learning, with its robust feature-learning capabilities, has revolutionized the automated interpretation of large-scale remote sensing data. However, existing remote sensing interpretation frameworks remain fundamentally constrained by their rigid coupling to sensor-specific spectral distributions and band configurations. This sensor dependency restricts model training and inference to single-sensor paradigms, resulting in the severe underutilization of multi-source Earth observation archives. To address this critical challenge, this paper proposes a Universal Interpretation Framework (UIF) capable of processing arbitrary hyperspectral and multispectral imagery. Based on UIF, we develop a cross-device classification model that supports the mixed training and unified inference of heterogeneous hyper-spectral and multispectral datasets. On the Yellow River Estuary wetland cross-device classification dataset, a single model simultaneously achieves accuracies ranging from 88% to 92% across four datasets with distinct spectral configurations. Furthermore, we extend UIF to construct a cross-device change detection model. On the Yellow River Estuary wetland cross-device change detection dataset, a single model achieves multi-class change detection accuracies exceeding 90% on two distinct change detection tasks involving multiple optical remote sensing platforms. These results demonstrate that UIF effectively breaks the strong coupling between interpretation models and specific sensor configurations, substantially improves the utilization efficiency of heterogeneous remote sensing data, expands the application boundaries of remote sensing interpretation technology, and provides a novel technical paradigm for generalized remote sensing analysis.