On the generalization ability of probabilistic neural networks for hyperspectral remote sensing of absorption properties across optically complex waters

Werther, M, Burggraaff, O, Gurlin, D, Saranathan, AM, Balasubramanian, SV, Giardino, C, Braga, F, Bresciani, M, Pellegrino, A, Pinardi, M, Simis, SGH, Lehmann, MK, Kangro, K, Alikas, K, Ficek, D and Odermatt, D 2025 On the generalization ability of probabilistic neural networks for hyperspectral remote sensing of absorption properties across optically complex waters. Remote Sensing of Environment, 328. 114820. 10.1016/j.rse.2025.114820

Preview

Text 1-s2.0-S003442572500224X-main.pdf - Published Version Available under License Creative Commons Attribution. Download (10MB) | Preview

Abstract/Summary

sensing observations. Yet, their generalization ability when applied to new water bodies, beyond those they were trained on, is not well understood. We present a novel approach for assessing model generalization across various scenarios, including interpolation within in situ observation datasets, extrapolation beyond the training scope, and application to hyperspectral observations from the PRecursore IperSpettrale della Missione Applicativa (PRISMA) satellite involving atmospheric correction. We evaluate five probabilistic neural networks (PNNs), including novel architectures like recurrent neural networks, for their ability to estimate absorption at 443 and 675 nm from hyperspectral reflectance. The median symmetric accuracy (MdSA) worsens from ≥25% in interpolation scenarios to ≥50% in extrapolation scenarios, and reaches ≥80% when applied to PRISMA satellite imagery. Across all scenarios, models produce uncertainty estimates exceeding 40%, often reflecting systematic underconfidence. PNNs show better calibration during extrapolation, suggesting an intrinsic awareness of retrieval constraints. To address this miscalibration, we introduce an uncertainty recalibration method that only withholds 10% of the training dataset, but improves model calibration in 86% of PRISMA evaluations with minimal accuracy trade-offs. Resulting well-calibrated uncertainty estimates enable reliable uncertainty propagation for downstream applications. IOP retrieval uncertainty is predominantly aleatoric (inherent to the observations). Therefore, increasing the number of measurements from the same distribution or selecting a different neural network architecture trained on the same dataset does not enhance model accuracy. Our findings indicate that we have reached a predictability limit in retrieving IOPs using purely data-driven approaches. We therefore advocate embedding physical principles of IOPs into model architectures, creating physics-informed neural networks capable of surpassing current limitations.

Item Type:	Publication - Article
Additional Keywords:	Generalization Neural networks Inherent optical properties Hyperspectral remote sensing Optically complex waters
Divisions:	Plymouth Marine Laboratory > Science Areas > Earth Observation Science and Applications
Depositing User:	S Hawkins
Date made live:	03 Jul 2025 09:15
Last Modified:	03 Jul 2025 09:15
URI:	https://plymsea.ac.uk/id/eprint/10438

Actions (login required)

View Item