Topological segmentation of shallow waters for remote sensing-based water quality modeling

Abstract

Satellite based monitoring of inland lake water quality is challenged by optically shallow areas, where bottom reflectance introduces bias into reflectance based retrievals. Pansharpening improves spatial resolution of multispectral imagery by fusing multispectral bands with a higher resolution panchromatic band. The effect of pansharpening on chlorophyll-a (Chl-a) retrieval across depth stratified water zones has not been quantified in shallow boreal lakes. This thesis addressed this gap for Lake Saimaa, Finland, using Landsat 8 and 9 multispectral imagery and bathymetric depth data stratified into five mutually exclusive depth zones. A Kernel PLS model optimized with Kernel Flows (KF-PLS) is calibrated from in-situ Chl-a measurements. The model is applied to both the original 30 m multispectral imagery and the Additive Wavelet Luminance Proportional (AWLP) pansharpened imagery at 15 m resolution. AWLP pansharpening introduces spectral distortion predominantly in the near infrared and shortwave infrared bands, which also carry the strongest depth related spectral pattern among all predictor bands. Systematically higher Chl-a predictions are produced from the pansharpened imagery in all five depth zones. The bias is consistent across zones, with no evidence that shallow pixels are disproportionately affected. Standard quality metrics compare the fused product against it’s inputs rather than the original spectral reference and do not predict this bias. Spectral changes induced by pansharpening could systematically affect Chl-a retrieval when a model calibrated at the original resolution is applied to pansharpened imagery. Direct spectral comparison with original bands is necessary in addition to standard quality metrics. The depth zone diagnostic framework creates a basis for depth aware water quality retrieval in shallow boreal lakes.