CAREER: Mapping Anthropocene Geomorphology with Deep Learning, Big Data Spatial Analytics, and LiDAR

地理空间数据（例如，航空和卫星图像、数字高程数据和天气观测）正在以惊人的速度生成。例如，美国地质调查局的Landsat地球观测计划每天收集约1TB的数据，他们的3D高程计划（3DEP）正在努力为整个国家生成第一个高细节的高程数据集，计划于2023年完成。虽然这些“大数据”项目提供了许多机会，但目前很难以有效的方式从中提取可采取行动的信息，以支持科学研究和知情决策。虽然人工智能和机器学习的最新进展在分析此类数据方面显示出巨大的前景，但仍需要进一步研究和开发这些技术，以应用于数字测绘任务，例如检测山体滑坡，监测资源开采和记录景观变化。这项研究将开发最先进的基于“深度学习”的技术，利用地理空间数据（包括高程模型和历史地图）获取有关人类对景观改造的宝贵信息，从根本上推进地貌测绘科学。除了支持和培训西弗吉尼亚大学的研究生外，这项工作还将通过开发培训和教学材料来吸引未来的高中STEM教师，在职教师和高中学生，这些材料将有助于下一代数据科学家，地理空间专业人员，该项目将推进地理空间数据分析和先进计算方法的应用，从在广泛区域的地理空间数据，以进一步了解自然景观和人为景观的变化。它专门探讨了基于卷积神经网络（CNN）的语义和实例分割深度学习方法，这些方法能够对空间上下文信息进行建模，用于从地理空间数据中提取地貌特征和历史挖掘，包括历史地形图，光探测和测距（LiDAR）点云以及其他地形表示（即，山体阴影和其他地形衍生物）。最终，该项目将有助于使用越来越丰富的高质量数字地形数据来实现地貌制图的深度学习，最终目标是在区域到全球范围内生成准确的数据集，以便对地貌灾害以及自然和人为景观变化进行记录，量化和建模。该项目由地貌学和土地利用动力学以及刺激竞争性研究的既定计划（EPSCoR）共同资助。该奖项反映了NSF的法定使命，并通过使用基金会的知识价值和更广泛的影响审查标准进行评估，被认为值得支持。

Geospatial data (e.g., aerial and satellite imagery, digital elevation data, and weather observations) are being generated at an astounding rate. One example, the US Geological Survey’s Landsat Earth observation program, collects about a terabyte of data daily, and their 3D Elevation Program (3DEP) is working toward generating the first-ever high-detail elevation dataset for the entire country, which is scheduled to be completed by 2023. While these “big data” projects present many opportunities, it is currently very difficult to extract actionable information from them in an efficient manner that supports scientific research and informed decision making. While recent advances in artificial intelligence and machine learning show great promise in analyzing such data, there is a need to further research and develop these techniques for application to digital mapping tasks, such as detecting landslides, monitoring resource extraction, and documenting landscape change. This research will develop state-of-the-art “deep learning”-based techniques to derive valuable information on human modifications to the landscape using geospatial data, including elevation models and historic maps, to fundamentally advance geomorphic mapping science. In addition to supporting and training graduate students at West Virginia University, the work will engage future high school STEM teachers, in-service teachers, and high school students by developing training and instructional materials that will help enable the next generation of data scientists, geospatial professionals, and coders.This project will advance the application of geospatial data analytics and advanced computational methods to extract high spatial resolution information from geospatial data over wide regions to further understanding of natural landscapes and anthropogenic landscape change. It specifically explores semantic and instance segmentation deep learning methods based on convolutional neural networks (CNNs), which are able to model spatial context information, for extracting geomorphic features and historic mining from geospatial data, including historic topographic maps, light-detection and ranging (LiDAR) point clouds, and additional terrain representations (i.e., hillshades and other topographic derivatives). Ultimately, this project will contribute to operationalizing deep learning for geomorphic mapping using the increasing abundance of quality digital terrain data with the eventual goal of generating accurate datasets at regional to global extents that will allow for documentation, quantification, and modeling of geomorphic hazards and natural and anthropogenic landscape change. This project is jointly funded by Geomorphology and Land-use Dynamics and the Established Program to Stimulate Competitive Research (EPSCoR).This award reflects NSF's statutory mission and has been deemed worthy of support through evaluation using the Foundation's intellectual merit and broader impacts review criteria.