Identifying and Handling of Errors Caused by Spectral Ambiguities over Water

Kurzfassung

Challenge All surface waters contain phytoplankton, coloured dissolved organic matter (CDOM) and non-algal particles. The diversity of type, chemical composition, size distribution, and in particular concentration introduces a fundamental numerical problem to data analysis: different combinations of water constituents can lead to similar reflectance spectra. These spectral ambiguities not only depend on the water constituents, but also on the sensor: less bands and increased noise reduce the measured spectral information and therefore increase the ambiguity problem. This problem is unfortunately not easily recognized during data processing, i.e. the concentration maps derived from a multispectral or hyperspectral image can appear plausible even when spectral ambiguities introduce severe errors. Methodology The Water Colour Simulator (WASI) software is available since 20 years for simulating spectral measurements in and above water, and for analysing such measurements by inverse modelling. Its module WASI-2D can process images from multispectral and hyperspectral sensors since 10 years. A new module, WASI-AI, has now been developed that combines physical modelling with Artificial Intelligence (AI) for image processing. Processing randomly selected pixels with two independent methods, inverse modelling and AI, makes it possible to determine the presence of ambiguities for each fit parameter: without ambiguities, the results are highly correlated, while the results become increasingly decorrelated the more severe the ambiguity problem is. Correlation plots of inverse modelling vs. AI results are therefore useful to identify ambiguity problems. WASI has been designed from the beginning for handling such problems by allowing the user to import site-specific optical properties to each model parameter, deciding for all model parameters if they shall be used as fit parameters or kept constant, defining image-specific initial values and ranges for each fit parameter, and giving the sensor bands different weights according to their information content and noise. The correlation plots now allow assessing the usefulness of each measure and optimizing image processing with respect to the ambiguity problem. Results Our goal was distinguishing from space different phytoplankton groups that occur as mixtures at low to moderate concentrations. Low concentration in combination with similar optical properties require a hyperspectral sensor and little noise. From all currently available spaceborne sensors with a spatial resolution suitable for inland waters, DESIS has the highest spectral resolution, therefore we used DESIS images from Lake Constance as test dataset. As expected, the spectral ambiguity problem is quite heavy for this task due to the similarity of the optical properties of phytoplankton groups. By applying the described methodology, we were able to improve the number of distinguishable groups from two to four for one scene, while for other scenes still just two groups could be differentiated, but with reduced uncertainty.