Deep learning reveals monthly river and stream surface area on the Tibetan Plateau using Sentinel-2 imagery

Description: The deep learning method, RWSD (River Width estimation from Sentinel-2 images based on Deep learning), utilizes a combination of classification and regression models to estimate river widths, particularly for smaller rivers (e.g., width < 100 meters). The approach first classifies Sentinel-2 TOA images into lakes vs. non-lakes and active river channels vs. non-active channels, which include ice, snow, and land areas. Subsequently, the regression model estimates the wetted river width at the cross-section closest to the image center. This method has been developed for global applicability with a particular emphasis on alpine environments. Files and Descriptions: 1. TP_Lake.csv: This CSV file contains identified lakes in the Tibetan Plateau. 2. TP_River_Monthly_Statistics.csv: Monthly statistics for river data on the Tibetan Plateau, including estimations like active channel percentage and width for different river orders. 3. S2RiverWidth.py: The Python script that contains the main code for our river width estimation model. This script includes the functions necessary for preprocessing Sentinel-2 images and predicting river widths. 4. best_model_width.pth: The pre-trained deep learning model weights used for river width estimation. This model is a ResNeXt model fine-tuned on our dataset. 5. SampleS2Img.tif: A sample Sentinel-2 TOA image in .tif format, used for testing the river width estimation model. Model Input Requirements:The model requires a Sentinel-2 TOA image in .tif format as input. The image should be scaled by a factor of 10,000. The `get_model_input` function automatically crops the image to 224x224 pixels around the center to fit the model's input requirements.