Estimating water abundance from Landsat imagery integrated with boundary information under the Google Earth Engine framework

Accurate and efficient monitoring of surface water bodies holds critical significance. To address the accuracy limitations of traditional water body extraction methods in processing mixed pixels, this study proposed a Google Earth Engine (GEE)-based method for estimating the water abundance from Landsat imagery. Specifically, the water body boundary information was extracted through stacked neural networks; the spectral and boundary features were jointly extracted using a pseudo-siamese network; the water abundance was finally estimated by integrating the multi-source features. The model was deployed on the GEE platform to enable online prediction, effectively avoiding the transmission and storage limitations commonly encountered in large-scale applications of traditional offline methods. Using the Landsat and GF-2 data from the Jianghan Plain, the proposed method was tested and compared with a linear regression model, a very deep super-resolution (VDSR) model, and a convolutional neural network (CNN) model without boundary information. The results show that compared to the above three models, the proposed method achieved an average reduction of 10.5% in the root mean square error (RMSE) and 14.5% in the mean absolute error (MAE), and an average improvement of 4.7% in the coefficient of determination (R²), while also significantly saving the data storage space and transmission time.