River Channel Microgeomorphic Feature Extraction and Potential Sandstorm Source Identification Method Based on a Convolutional Autoencoder Model

River channel's microgeomorphic features are crucial for identifying potential sandstorm sources and studying sediment source-sink processes. Current deep learning methods are predominantly applied to visible objects, rendering them unsuitable for latent objects with unstable spatiotemporal distributions, such as potential sandstorm sources. These latent objects require the analysis of their internal structures using unsupervised methods. Convolutional kernels in convolutional neural networks capture local spatial structures, and their size is essential for accurately analyzing internal structures. A key challenge is determining the appropriate scale for identifying the latent objects. This model uses high-resolution remote sensing imagery and employs a convolutional autoencoder to extract common features from river channels. Determining the optimal convolution kernel size enables the automatic and efficient identification of the morphological boundaries of latent objects, extracts the spatiotemporal common features of river microtopography, and reconstructs the microtopography background. Anomaly detection methods are employed to identify regions with spatial structural anomalies, recognizing areas that are potential and dynamic sandstorm sources. It addresses the challenges of spatiotemporal feature extraction in complex geographical environments, and the identification of potential wind-blown sand sources in river channels. The approach was applied to the Yarlung Zangbo River from Qushui to Zedang using 2013–2018 Landsat 8 remote sensing images. The results show that this method can effectively identify river microtopographic features and potential sandstorm sources.