Enhancing masked autoencoders with kolmogorov-arnold networks and metric learning for robust hyperspectral analysis

The accelerating impacts of climate change and human activities on marine and coastal ecosystems demand advanced monitoring strategies that are both accurate and sustainable. Hyperspectral imaging (HSI) offers unparalleled capability to characterize water constituents, benthic habitats, and coastal vegetation, yet its application in aquatic environments is challenged by spectral redundancy, spatial heterogeneity, and severe domain shift. To address these limitations, we propose a KAN-based Masked Autoencoder (KAN-MAE) framework that integrates artificial intelligence and machine learning innovations into ocean remote sensing. The model employs decoupled spectral and spatial encoders with adaptive spectral masking, cross-attention fusion, and KAN-enhanced Transformer blocks for robust representation learning. During pre-training, Brownian distance covariance (BDC) regularization preserves nonlinear inter-band dependencies, while fine-tuning leverages a lightweight KAN classifier optimized with cross-entropy and Gaussian-kernel triplet loss to enhance discriminability under scarce labels. Experiments on Gaofen-5 coastal imagery and benchmark HSI datasets demonstrate improved classification accuracy, resilience to noise, and strong generalization across domains. By advancing AI-driven self-supervised learning for ocean remote sensing, this work supports operational monitoring of coastal ecosystems, sustainable fisheries, and blue carbon management, aligning with the UN Sustainable Development Goals on climate action (SDG 13) and life below water (SDG 14).