<p>Hyperspectral image (HSI) classification has been core to the environmental monitoring practices such as land-cover mapping, vegetation analysis, and water quality measurement. Nevertheless, the current techniques tend to not leverage both spectral and spatial context effectively, especially when the amount of labeled data is sparse as is the case with real-world environmental surveys. In our paper, AF-REMNet (Attention-based Feature Refinement Environmental Monitoring Network) is proposed, which is a new dual-path network that learns simultaneously discriminative spectral and spatial representations by using complementary encoding pathways. We use (1) spectral dependency encoder that uses gated 1D convolutions to calculate band-wise correlations, (2) multi-scale spatial encoder that applies dilated depthwise separable convolutions to effectively extract the textures, (3) spatial-channel attention modules that refine the features, and (4) gated adaptive fusion mechanism that dynamically weights both spectral and spatial contributions on individual samples. We propose a powerful training scheme with spectral augmentations, band dropout, and consistency regularization to handle domain shifts due to the variation of the environment. Extensive experiments on the WHU-Hi benchmark datasets prove that AF-REMNet is a state-of-the-art at different training set sizes (25–300 samples per class), and overall accuracy gains of 2–5% over current approaches. In addition, we offer deployment optimized versions that can be used in real-time environment monitoring applications.</p>