New strategy to improve estimation of diffuse attenuation coefficient for highly turbid inland waters

The diffuse attenuation coefficient, K_d(λ), is an important water optical property. Detection of K_d(λ) by means of remote sensing can provide significant assistance in understanding water environment conditions and many biogeochemical processes. Even when existing algorithms exhibit good performance in clear open ocean and turbid coastal waters, accurate quantification of highly turbid inland water bodies can still be a challenge due to their bio-optical complexity. In this study, we examined the performance of two typical pre-existing K_d(490) models in inland water bodies from Lake Taihu, Lake Chaohu, and the Three Gorges Reservoir in China. On the basis of water optical classification, new K_d(490) models were developed for these waters by means of the support vector machine approach. The obtained results showed that the two pre-existing K_d(490) models presented relatively large errors by comparison with the new models, with mean absolute percentage error (MAPE) values above ~30%. More importantly, among the new models, type-specific models generally outperformed the aggregated model. For water classified as Type 1 + Type 2, the type-specific model produced validation errors with MAPE = 16.8% and RMSE = 0.98 m^?1. For water classified as Type 3, the MAPE and RMSE of the type-specific model were found to be 18.8% and 1.85 m^?1, respectively. The findings in this study demonstrate that water classification (prior to algorithm development) is needed for the development of excellent K_d(490) retrieval algorithms, and the type-specific models thus developed are an important supplement to existing K_d(490) retrieval models for highly turbid inland waters.