| Popis: |
The carrying capacity of water resources is key to the sustainable development of arid and semi-arid regions. There are important challenges related to the detection of discontinuous and crooked water bodies in the vast Mongolian Plateau, despite the availability of remote sensing technology which has the advantage of facilitating water observations over large areas and timelines. Given the high cost and low coverage of high-resolution images and the low resolution of images with high coverage, this study proposes a pixel-based convolutional neural network (CNN) method for the application of water extracted from the region of interest (ROI) to medium-resolution Landsat images. The pixel-based CNN method combines the texture and spectral features of the ground object by connecting the center pixels of the images to the surrounding pixels. ROI is used instead of full-label datasets, reduce the difficulty of building labels in low-to-medium-resolution images. Taking the Tuul River in Mongolia as a case, the pixel-based CNN method, the normalized difference water index threshold (NDWI) method, the modified normalized difference water index (MNDWI) threshold method, U-net model in deep learning, and the pixel-based deep neural network (DNN) method were used with medium-resolution Landsat 8 images with ROI labels. The pixel-based CNN method shows better water extraction results for the cloud, cloud shadows, and building areas, compared with other methods. The method proposed in this study had the highest verification accuracy (92.07%). It also has the advantages of fewer training parameters and shorter training time. The training accuracies of the pixel-based CNN, pixel-based DNN, and U-net were 99.90%, 96.98%, and 93.70%, respectively. All training models and calling methods were uploaded to GitHub (https://github.com/CaryLee17/Pixel-based-CNN). |
Full Text from ScienceDirect