基于多源地学时空数据分层学习的滑坡易发性动态区划

On the basis of multi-source geological big data mining, landslide susceptibility dynamic mapping is used to dynamic evaluate and partition the distribution of regional landslide susceptibility by the establishment of regional landslide dynamic spatio-temporal evolution model. This project will focus on the landslides in the Three Gorges region. Advanced computational intelligence methods will be utilized to construct the landslide inventory maps, to study the evolutionary pattern switched in multi-stage of collapse, sliding and flow, and to predict the regional landslide susceptibility. In this project, target image segmentation methods will be utilized to recognize the spatial distribution and spatial morphology of landslides. Hierarchical learning will be used to identify the type hierarchy of landslides, and it will facilitate the construction of landslide inventory maps with rich semantic information. Unsupervised learning and semi-supervised learning will be used to adaptively divide different evolution stages of three different landslide types including collapse, sliding and flow. A hierarchical switched system will be constructed to dynamically simulate the evolution of landside deep displacement. To build a dynamic prediction model of regional landslide susceptibility, deep neural networks will be used to fuse timing information and spatial information in two phases. The research of this project will promote the cross-integration between machine learning and engineering geology, as well as promote the research of large-scale geo-data mining and landslide hazard prediction.

滑坡易发性动态区划是在多源地学大数据挖掘的基础上，通过建立区域滑坡动态时空演化模型，对区域滑坡易发性散布情形进行动态地评价和分区。本项目以三峡库区滑坡为研究对象，拟发展先进的计算智能方法对滑坡动态编目图构建、崩滑流多阶段演化模式切换和区域滑坡易发性预测展开研究。项目拟利用目标分割方法识别区域滑坡群空间分布和空间形态，利用分层学习识别滑坡层次结构类型，试图构建具有丰富语义信息的滑坡编目图；利用无监督和半监督学习自适应划分崩滑流三种不同滑坡运动类型下的不同演化阶段，构建分层切换系统动态模拟滑坡深部位移演化；利用深度神经网络分两阶段融合时序信息和空间信息，构建区域滑坡易发性动态预测模型。本项目的研究将促进机器学习技术和工程地质技术的交叉融合，为地学大数据挖掘和滑坡灾害预测研究起到一定的推动作用。

我国滑坡灾害发生频率高、散布密度大，严重威胁人民的生命和财产安全。近年来，随着信息监测技术的不断发展，能够获取到大量的滑坡演化多场信息数据。本项目结合机器学习领域的新进展，发展基于机器学习的滑坡多场信息数据挖掘技术，实现滑坡灾害区域易发性预测，为滑坡防治提供新思路和新方法。项目取得的主要成果如下：1）收集、整理和完善了多个实地监测的重大工程灾变滑坡多场信息数据集，为机器学习算法的规范化训练和性能评估提供了数据基础。2）提出了基于选择性集成的深度学习模型，更有效地挖掘滑坡多场信息非线性表征。3）提出了基于迁移学习的神经网络区间预测模型预训练方法，提升了滑坡单监测点位移区间预测模型的鲁棒性。4）将滑坡单监测点的位移预测推广到整个滑坡面的多监测点位移预测，提出了基于多任务学习的滑坡面多点位移预测模型，实现了对滑坡面整体演化趋势和灾害易发性的预测。本项目按照计划完成了相关研究，发表了多篇相关论文，总体完成了预期要求。