河流边界条件是河流地貌演化、流域水沙模拟、河流物质通量计算、河流生境评估等研究的基础。河流断面形态能够反映河流边界，然而在缺资料的间歇性河流地区，较难通过传统方法获取河流连续的断面形态，因此通常将河流断面概化为简单形状。概化断面的常用形状为一般梯形，但在提取缺资料地区或大尺度流域的连续断面时，难以获取床面宽度，致使概化结果严重失真。该研究以多源卫星影像为数据源，优化并行随机森林-人工神经网络(PRF-ANN)算法，并基于土壤水分降尺度方法构建了考虑枯水期河床特征的河流表面信息提取算法——GPU并行随机森林-卷积神经网络(GRF-CNN)算法。以位于黄土高原的典型间歇性河流无定河流域为研究区域，验证算法效果。结果表明：建立的流域土壤水分降尺度反演方法，反演结果与原位数据相比，R²为0.81、RMSE小于0.059 m³/m³，优于其他产品；GRF-CNN提取平滩期河流表面信息及枯水期河流床面信息的精度分别为95.7%和90.3%，较目前深度学习算法提升超10%；基于水文站实测数据，该研究提取的概化断面面积较PRF-ANN算法精度提升27.3%。该研究显示了考虑河流特征的GRF-CNN算法提取细小河流的突出能力，展现了可解释机器学习在提取河流信息方面的优势。该研究可为基于多源卫星遥感观测提取间歇性河流断面形态提供关键参数，为流域河流边界重构、缺资料地区数字孪生流域建设提供技术及数据支撑。

Objective: Intermittent rivers, distinguished by sporadic flow cessation, account for more than half of the global river network and are particularly prevalent in semiarid and arid zones, such as the northwest region of China. These watercourses are pivotal in hydrological modeling, sediment transport simulations, and evaluation of riverine ecosystems. Nevertheless, precise extraction of riverbed and cross-sectional data for intermittent rivers during low-flow periods poses a formidable challenge, largely due to the paucity of data and the constraints inherent in conventional measurement techniques. The current methodologies, including digital elevation models (DEM) and empirical equations, fail to deliver continuous and accurate data on river morphology. This study tackles these challenges by introducing a graphics processing unit (GPU)-accelerated random forest (GRF)-convolutional neural network (CNN) algorithm, referred to as the GRF-CNN algorithm, that harnesses multisource remote sensing information and integrates soil moisture downscaling to delineate riverbed characteristics under bankfull and dry states. Methods: This research centers on the Wuding River Basin, a typical intermittent river system located on the Loess Plateau of China. This study develops and validates the GRF-CNN algorithm via integration and optimization of advanced techniques for river morphology extraction. Multisource remote sensing datasets, such as thermal infrared data from the Landsat-8 and ZY1-02E satellites and soil moisture data from the soil moisture active passive (SMAP) satellite and the Sentinel-1 synthetic aperture radar, are used for comprehensive surface and subsurface analyses. Building on the parallel random forest-artificial neural network (PRF-CNN) framework, the GRF-CNN algorithm incorporates CNN modules to enhance the downscaling of thermal infrared data and improve soil moisture inversion. A hybrid statistical-physical model, combining the water cloud model, Dubois model, and advanced integral equation model, along with the deep supervised neural network, is employed to increase the spatial resolution of soil moisture data to 10 m. Specific soil moisture thresholds for different river levels (4th-7th order) are established to accurately extract riverbeds and cross-sectional boundaries during dry seasons. The algorithm's performance is confirmed via hydrological station measurements and high-resolution satellite images. Results: The results demonstrate that the downscaled soil moisture data achieve an R² of 0.81 and a root mean square error of < 0.059 m³/m³, outperforming conventional soil moisture products. The GRF-CNN algorithm achieves 95.7% accuracy for bankfull river surfaces and 90.3% for dry season riverbeds, with a 27.3% improvement in accuracy of generalized cross-sectional bottom width extraction versus the PRF-ANN algorithm. Furthermore, GRF-CNN surpasses widely recognized machine learning models, including DiCNN-4, Bi-LSTM, and Transformer-based methods, by a margin of at least 10% in terms of Kappa coefficient, overall accuracy, classification accuracy, F₁ score, and recall, particularly in the extraction of intermittent river features. Conclusions: This study underscores the potential of integrating multisource remote sensing data with sophisticated machine learning algorithms to tackle the challenges associated with the extraction of river morphology in intermittent river systems. The GRF-CNN algorithm offers a scalable and precise solution for reconstructing river boundaries and estimating cross-sectional parameters in data-scarce regions. Moreover, it illustrates the benefits of incorporating soil moisture dynamics into riverbed morphology extraction processes. We believe the results have far-reaching implications for hydrological modeling, sediment transport studies, and developing digital twin watersheds for resource management and ecological restoration in arid and semiarid areas. By addressing the gap in river morphology data extraction, this study provides substantial technical and data support for increasing the accuracy of hydrological simulations and advancing sustainable watershed management strategies.