A probabilistic deep learning framework for retrieving chlorophyll-a from hyperspectral imagery: Integrating channel attention and mixture density networks

Accurate monitoring of Chlorophyll-a (Chla) is critical for assessing aquatic ecosystem health, yet ecological complexity often leads to ambiguous spectral signatures in satellite data. Traditional deterministic models assume a one-to-one mapping between spectra and pigments, often failing to capture these high-dimensional analytical challenges. In this study, we propose a novel deep learning architecture, the Channel Attention-Mixture Density Network (CA-MDN), to retrieve Chla from the National Aeronautics and Space Administration (NASA) Earth Surface Mineral Dust Source Investigation (EMIT) hyperspectral mission. The CA-MDN integrates an attention mechanism to dynamically select ecologically relevant spectral bands and employs a probabilistic output layer to quantify retrieval uncertainty. Trained and tested against the global GLObal Reflectance community dataset for Imaging and optical sensing of Aquatic environments (GLORIA) in situ dataset, the model achieved a prediction error of 40.52%, significantly outperforming conventional machine learning baselines. Case studies in California waters demonstrate the model's ecological utility, showing how probabilistic modeling can resolve fine-scale variability and flag high-uncertainty regions. This study presents a robust computational framework for leveraging spaceborne imaging spectroscopy in complex coastal and inland environments. • Proposed a Channel Attention-Mixture Density Network (CA-MDN) for Chla retrieval. • Algorithm handles spectral redundancy and quantifies prediction uncertainty. • Outperforms standard ML models on the global GLORIA dataset. • Demonstrated high-resolution (60 m) monitoring utility using NASA EMIT data.