An optical mechanism-based deep learning approach for deriving water trophic state of China's lakes from Landsat images.

• A lake TSI inversion method based on optical mechanism was proposed. • a t-w (440) was confirmed to be the optimal optical characterization parameter of TSI. • DNN model was superior to other models in the inversion of both TSI and a t-w (440). • Trophic status of lakes across China was assessed using the proposed approach. • TSI of 961 lakes in China in 2020 ranged from 6 to 96, with a mean value of 48. Widespread eutrophication has been considered as the most serious environment problems in the world. Given the critical roles of lakes in human society and serious negative effects of water eutrophication on lake ecosystems, it is thus fundamentally important to monitor and assess water trophic status of lakes. However, a reliable model for accurately estimating the trophic state index (TSI) of lakes across a large-scale region is still lacking due to their high complexity. Here, we proposed an optical mechanism-based deep learning approach to remotely estimate TSI of lakes based on Landsat images. The approach consists of two steps: (1) determining the optical indicators of TSI and modeling the relationship between them, and (2) developing an approach for remotely deriving the determined optical indicator from Landsat images. With a large number of in situ datasets measured from lakes (2804 samples from 88 lakes) across China with various optical properties, we trained and validated three machine learning methods including deep neural network (DNN), k-nearest neighbors (KNN) and random forest (RF) to model TSI with the optical indicators and TSI and derive the determined optical indicator from Landsat images. The results showed that (1) the total absorption coefficients of optically active constituents at 440 nm (a t-w (440)) performs best in characterizing TSI, and (2) DNN outperforms other models in the inversion of both TSI and a t-w (440). Overall, our proposed optical mechanism-based deep learning approach demonstrated a robust and satisfactory performance in assessing TSI using Landsat images (root mean squared error (RMSE) = 5.95, mean absolute error (MAE) = 4.81). This highlights its merit as a nationally-adopted method in lake water TSI estimation, enabling the convenience of the acquisition of water eutrophic information in large scale, thereby assisting us in managing lake ecology. Therefore, we assessed water TSI of 961 lakes (>10 km2) across China using the proposed approach. The resulting a t-w (440) and TSI ranged from 0.01 m−1 to 31.42 m−1 and from 6 to 96, respectively. Of all these studied lakes, 96 lakes (11.40 %) were oligotrophic, 338 lakes were mesotrophic (40.14 %), 360 lakes were eutrophic (42.76 %), and 48 were hypertrophic (5.70 %) in 2020. [ABSTRACT FROM AUTHOR]