Deep learning-based remote sensing estimation of water transparency in shallow lakes by combining Landsat 8 and Sentinel 2 images

Abstract

Water transparency is a key indicator of water quality as it reflects the turbidity and eutrophication in lakes and reservoirs. To carry out remote sensing monitoring of water transparency rapidly and intelligently, deep learning technology was used to construct a new retrieval model, namely, point-centered regression convolutional neural network (PSRCNN) suitable for Sentinel 2 and Landsat 8 images. The impact of input feature variables on the accuracy of the inversion model was examined, and the performance of an optimized PSRCNN model was also assessed. This model was applied to remote sensing images of three shallow lakes in the eastern China plain acquired in summer. The PSRCNN model, constructed using five identical bands from Landsat 8 and Sentinel 2 images and 20 band combinations as the input variables, the input window size of 5 × 5 pixels, proves a good predictive ability, with a verification accuracy of R² = 0.85, the root mean squared error (RMSE) = 13.0 cm, and the relative predictive deviation (RPD) = 2.58. After the sensitive spectral analysis of water transparency, the band combinations that had correlation coefficients higher than 0.6 were selected as the new input feature variables to construct an optimized PSRCNN model (PSRCNN_opt) for water transparency. The PSRCNN_opt model has an excellent predictive ability, with a verification accuracy of R² = 0.89, RMSE = 11.48 cm, and RPD =3.0. It outperforms the commonly retrieval models (band ratios, random forest, support vector machine, etc.), with higher accuracy and robustness. Spatial variations in water transparency of three lakes from the retrieval results by PSRCNN_opt model are consistent with the field observations.

Keywords: Convolutional neural network; Landsat 8; Sentinel 2; Shallow lakes; Water transparency.