重力侵蚀主要包括崩塌、泻溜、泥流等，是陡峭区域的主要地貌过程之一，也是沟道泥沙的重要来源，其影响因素多，发生不确定性强，导致监测数据极其缺乏，制约了时空特征与机制研究。发展重力侵蚀高效监测方法，开展重力侵蚀识别研究，可为重力侵蚀的深入研究奠定数据基础，是重力侵蚀研究的关键所在。本研究以黄土丘陵沟壑区典型流域桥沟为研究区，结合野外调查、定位观测、无人机激光雷达（ULS）、无人机摄影测量（UAV-P）、地面三维激光扫描（TLS）和卫星遥感等技术，通过2020年-2023年11次研究区雨季（7月-10月）数据采集，获取流域高精度三维地形数据及216组重力侵蚀样本数据（包括重力侵蚀发生位置、发生类型、侵蚀体积及降水、土壤、植被、地形等相关的影响因素）。基于所获取数据集，评估了不同算法和不同源地形数据监测重力侵蚀的精度，在此基础上改进了已有算法，形成重力侵蚀监测方法；详细分析了重力侵蚀的时空特征及其影响因素的定量关系，在此基础上利用深度网络对侵蚀事件进行识别研究。主要研究成果如下：

（1）系统评估了不同算法和数据源监测重力侵蚀的效果，结果表明，ULS数据、UAV-P数据和二者融合数据与DoD、C2M、C2C和M3C2算法的组合均可监测到重力侵蚀。UAV-P数据与C2M算法和ULS数据/融合数据与M3C2算法的组合分别得到侵蚀面和侵蚀体空间分布的最佳结果。融合数据与DoD、C2C、C2M和M3C2算法组合的估算侵蚀体积与实测侵蚀体积显著相关（R² > 0.90，p < 0.01）。ULS数据与DoD算法及融合数据与M3C2算法组合更适用于监测无植被区域或稀疏植被区域的侵蚀，融合数据与C2M算法组合更适用于监测植被茂盛区域的重力侵蚀。融合数据与M3C2算法组合更适用于监测低量级（< 0.1 m³）和中等量级（0.1 m³~0.5 m³）的重力侵蚀，ULS数据/UAV-P数据与DoD算法组合和融合数据与M3C2算法组合更适用于监测高量级（> 0.5 m³）的重力侵蚀。

（2）通过引入动态调整的投影尺度改进了M3C2算法，结果表明，改进方法能够更精准地监测重力侵蚀，减少了监测噪点，提升了变化检测的灵敏度。在侵蚀和沉积体区域空间分布监测中，改进算法结果优于原始算法的占比为92.59 %。改进算法在侵蚀面区域的监测噪点明显减少，且能监测到更多变化。较原始M3C2算法，改进算法的监测结果与实测数据有更高的拟合度（R² = 0.94）、纳什效率系数（NSE = 0.79）和更低的均方根误差（RMSE = 0.28 m³）。此外，改进算法减少了对侵蚀体积的高估，从原始算法的51.85%降至18.52%，其监测结果相对误差最小值、最大值、中位数和平均值较原始算法分别降低0.95 %、177.87 %、15.25 %和28.38 %。

（3）综合分析重力侵蚀的时空特征及其与多种影响因素的定量关系，结果表明，重力侵蚀主要发生在坡度大于70°的沟坡，其中崩塌是流域最主要的重力侵蚀类型，占总样本数的87.9 %。暴雨（降水强度超过50 mm/天）显著增加了重力侵蚀的发生频率、侵蚀面积和侵蚀体积，特别是在土壤容重为1.34 g/cm³时，重力侵蚀达到最高发生频率。此外，重力侵蚀多发生在沟坡无植被和无根系区域，沟坡植被抑制重力侵蚀发生的作用明显。表面光滑且侧剖面为平面的边坡是重力侵蚀发生的典型区域。重力侵蚀的发生频率、侵蚀面积和侵蚀体积随着地形湿度指数（TWI）和距坡顶线的距离增加而降低，随着距沟道距离的增加而增加。多种因素对重力侵蚀的相对影响分析结果表明降水量和抗剪强度分别是促进和抑制重力侵蚀发生的最重要影响因素，而地形对侵蚀面积和侵蚀体积的影响最大。

（4）基于重力侵蚀与影响因素关系分析结果，结合模型理论分析，开发CRCCF重力侵蚀识别模型，该模型结合Cascade Regions with CNN Features、Dual-Convnextv2和特征金字塔网络。结果表明，验证集的总体平均精度为0.55，小尺寸（< 9 m²）、中等尺寸（9 m²~92.16 m²）和大尺寸（> 92.16 m²）重力侵蚀平均精度分别为0.49、0.96和1.00，整体能够正确识别59.70 %的样本，小尺寸、中等尺寸和大尺寸的重力侵蚀正确识别比例分别为53.50 %、99.40 %和100 %。测试集的总体平均精度为0.53，小尺寸、中等尺寸和大尺寸的重力侵蚀平均精度分别为0.47、0.96和0.93，整体能够正确识别56.90 %的样本，小尺寸、中等尺寸和大尺寸的重力侵蚀正确识别比例分别为52.70 %、97.50 %和96.0 %。本研究所采集数据和样本量有限，所建模型精度有限，未来应通过增加样本采集量和优化模型算法来提高精度。

Mass movement mainly occurs in the form of collapses, spalling and mudflow, which is a principal form of geomorphic processes on steep-sloping areas and provides a primary sediment source for gullies and channels. Mass movement, influenced by various nonlinearly correlated factors, has often been seen as a random phenomenon, which presents a vast difficulty for the monitoring of mass movement and thus for the subsequent processes and mechanism studies. It was therefore crucial for process understanding to develop robust monitoring and identification techniques for mass movement, which provided powerful tools for the acquisition of mass movement relevant measurements. In line with this, eleven field surveys were undertaken in a mall typical catchment (Qiaogou) of the hilly and gully Loess Plateau to collect high-resolution terrain data and 216 mass movements relevant sample datasets (including high-precision three-dimensional topographic data of the Qiaogou catchment, location of mass movement, type of mass movement, erosion volume and influencing factors relevant to rainfall, soil, vegetation and topography), using a variety of methods including field investigations, positional observations, unmanned aerial vehicle laser scanning (ULS), unmanned aerial vehicle photogrammetry (UAV-P), terrestrial laser scanning (TLS), and satellite remote sensing. Based on the collected dataset, this study developed an efficient monitoring method for mass movement based on a thorough assessment of the combination of contemporary geomorphic change detection methods and multi-source topographic data. The spatial and temporal dynamics of mass movement along with their relationships with influencing factors was investigated. With reference to this understanding, deep learning networks were then employed to establish an identification method for mass movement. Main results are as follows:

(1) The effectiveness of different algorithms and data sources for monitoring mass movement showed that the combination of ULS data, UAV-P data and their fused data with digital elevation model (DEM) of difference (DoD), cloud-to-mesh distance or cloud-to-model distance (C2M), cloud-to-cloud comparison with closest point technique (C2C) and multiple model to model cloud comparison (M3C2) algorithms were capable of monitoring mass movement. The best spatial distributions of detachment niches and deposition areas were obtained through combining UAV-P data with C2M method and ULS data / fused data with M3C2 method, respectively. Erosion volume derived using the fused data plus DoD, C2C, C2M and M3C2 methods was significantly correlated (R² > 0.90, p < 0.01) with measured erosion volume. ULS data plus DoD method and fused data plus M3C2 method were preferable for bare areas and sparsely vegetated areas, while fused data plus C2M method was suitable for well vegetated areas. The fused data plus M3C2 method was more suitable for monitoring low (< 0.1 m³) and medium (0.1 m³-0.5 m³) magnitude erosion, while ULS data / UAV-P data plus DoD method and the fused data plus M3C2 method were more suitable for monitoring high magnitude (> 0.5 m³) mass movement.

(2) A dynamically adjusted projection scale was proposed to adapt the M3C2 algorithm. Results showed that the adapted algorithm achieved more accurate monitoring results, reduced monitoring noise, as well as enhanced sensitivity of change detection. In terms of spatial distribution for deposition areas, the adapted M3C2 algorithm outperformed the original algorithm, achieving superior monitoring results for 92.59 % of the samples. Additionally, the adapted algorithm considerably reduced noise in the detected detachment niches and achieved a greater number of changes. Performance metrics revealed that the adapted M3C2 algorithm, compared to the original algorithm, achieved a higher goodness-of-fit (R² = 0.94), a higher Nash-Sutcliffe efficiency coefficient (NSE = 0.79), and a lower root-mean-square error (RMSE = 0.28 m³). Another notable enhancement was the reduction for the overestimation of erosion volume, with the adapted algorithm being 18.52 % compared to 51.85 % for the original algorithm. In terms of error metrics, the adapted algorithm also considerably decreased the relative error in monitoring mass movement, with reductions of 0.95%, 177.87%, 15.25%, and 28.38% for the minimum, maximum, median, and mean values, respectively, when compared to the original algorithm.

(3) A comprehensive analysis of the spatial and temporal characteristics of mass movement and their relationship with influencing factors was conducted. Results showed that mass movement predominantly occurred on gully slopes with slope gradient greater than 70°. Collapses were the dominant type of mass movement in the Qiaogou catchment, which accounted for 87.9 % of the total number of samples. Rainstorms (rainfall intensity > 50 mm day^-1) significantly increased the occurrence frequency, erosion area and erosion volume of mass movement. This effect was particularly pronounced at a soil bulk density of 1.34 g/cm³, where erosive activity occurred with the highest frequency. Mass movement occurred most frequently on unvegetated or unrooted gully slopes, where the resisting effect of vegetation on mass movement was significant. Gully slopes with smooth rather than rugged profiles were found to be typical areas of mass movement. The occurrence frequency of mass movement decreased with the elevated topographic wetness index (TWI) and distance to slope top and increased with the distance to channels. For the relative impact of different factors, rainfall and shear strength were key factors facilitating and resisting the onset of mass movement, respectively, while topography exerted the greatest influence on erosion area and volume.

(4) A CRCCF mass movement identification model was developed based on the analysis of the relationship between mass movement and its influencing factors as well as model principles. The established model specifically incorporated Cascade Regions with CNN Features, Dual-Convnextv2, and a feature pyramid network. The results showed that validation datasets demonstrated an overall average precision of 0.55, achieving specific average precisions of 0.49 for small-size mass movement (less than 9 m²), 0.96 for medium-size mass movement (9 m² to 92.16 m²), and 1.00 for large-size mass movements (greater than 92.16 m²). The model successfully identified approximately 59.70 % of the samples across all sizes, with identification accuracies being 53.50 % for small samples, 99.40 % for medium samples, and 100 % for large samples. Results of test dataset showed an overall average precision of 0.529. The precision of the model for small-size, medium-size, and large-size mass movements was 0.47, 0.96, and 0.93, respectively, correctly identifying about 56.90 % of the samples. The rates of accurate identification for small, medium, and large samples were 52.70%, 97.50%, and 96.00%, respectively. The number of data and samples collected in the study was relatively limited, constraining the accuracy of the established identification model. Future efforts should be focused on enhancing data collection and refining modeling approaches to improve the accuracy of the model.