HyperEst: Context-aware self-supervised pretraining for hyperspectral and multispectral water quality estimation

Abstract

Accurate retrieval of water quality parameters (WQPs) from remote sensing imagery is essential for large-scale aquatic monitoring. However, existing models often suffer from limited generalizability across diverse optical water types, reliance on scarce labeled data, and sensitivity to input variability from different sensors. This paper proposes HyperEst, a self-supervised learning (SSL) framework designed to overcome these challenges. The core of our framework is a novel universal context-aware autoencoder (UCAA), which is pretrained on vast unlabeled hyperspectral imagery using a unique single-pixel reconstruction strategy and a multi-scale diffusion loss to promote robust spectral-spatial feature learning. The pretrained UCAA acts as a strong prior, enhancing generalization to unseen areas and reducing outlier errors. Furthermore, we introduce the weighted mean improvement (WMI), a metric designed for a balanced performance assessment across multiple WQPs, including Chlorophyll-a, total suspended solids, and colored dissolved organic matter. Once fine-tuned, HyperEst achieves state-of-the-art performance. Our model improves $log{R}^2$ by 1.87%–3.75% and reduces prediction bias by 26.24%–65.47% in terms of WMI. The experimental results highlight the scalability and robustness of HyperEst, representing a step forward in developing globally applicable models for water quality estimation from spaceborne observations.