Mixture Density Networks Improve Inter-Sensor Consistency of Optical Water Type Classification

Le résumé fourni par la source

In aquatic remote sensing, optical water type (OWT) preclassification has been used to map suitable algorithms to optically diverse targets and over optical gradients. However, current methods can yield inconsistent-type assignments across sensors with different spectral capabilities. Probabilistic neural networks (NNs) are increasingly used for constituent retrieval in optically complex inland water bodies because they can deliver higher accuracy than empirical or semianalytical algorithms, while also quantifying estimation uncertainty. However, their interpretability within bio-optical theory remains elusive. To address the limitations brought by these two methodological streams, we train mixture density networks (MDNs) to predict OWT membership distributions. OWTs serve as an optical-biogeochemical interpretable scaffold, alongside associated classification uncertainty, while the classification misalignment between Sentinel-3 Ocean and Land Colour Instrument (OLCI) and Sentinel-2 Multispectral Instrument (MSI) is reduced. Using >29000 co-located OLCI and MSI observations from 59 lakes, MDN ensembles trained to reproduce OLCI-derived OWT membership scores achieved 99% overall accuracy (OA) for OLCI and improved MSI cross-sensor OWT agreement from 58% to 73%, compared to using spectral angle as OWT distance metric. The MDNs enabled identification of aleatoric (upstream bias) and epistemic (lack of representative data) components of the uncertainty envelope, with posthoc uncertainty recalibration used to adjust uncertainty magnitudes to reliable confidence intervals, reducing miscalibration by 93% for OLCI and 80% for MSI, for both 68% and 95% intervals. Application of trained MDN ensembles across optically diverse systems confirmed that they learned generalizable optical relationships, improving OWT classification agreement across sensors, and providing diagnostics for bias identification. This approach preserves the interpretability of OWTs while enabling uncertainty-aware, cross-mission processing, identifying likely sources of classification uncertainty, particularly useful for less capable sensors, which can guide appropriate postprocessing and suitable algorithm selection and blending.

Ce résumé expose les affirmations des auteurs. BNTIC ne l’interprète pas comme une validation indépendante des résultats.