Hyperspectral imaging captures a dense stack of spectral bands, along with regular photo-like pixels, allowing it to distinguish minerals and map plants with fine detail. Still, mixing that rich spectrum with the shape and texture seen across space is tough whenever the feature space is vast and the training labels are scarce. In response, we develop a hybrid deep-network system that combines spectral and spatial learning within a two-pronged, or dual-branch, design. Its spectral arm runs a slim 1D CNN that hunts for small but telling shifts in color across just a few wavelengths. Meanwhile, the spatial arm feeds the same scene into a standard 2D CNN that detects edges, blobs, and other local structures. What each branch finds are merged by an adaptive attention layer that weighs the spectral cue against the spatial one on the fly before issuing the final class label. Tests on standard hyperspectral benchmarks demonstrate that our model surpasses traditional CNNs and the latest competitors in both accuracy and generalization to unseen sites. It also maintains high scores when classes are uneven or data is noisy, traits that are crucial for field campaigns and satellite work. Overall, the framework takes a significant step toward extracting all the valuable information from hyperspectral cubes and converting it into trustworthy maps.
Nandy et al. (Sun,) studied this question.