Virtual Laser Scanning for Machine Learning Algorithms in Geographic 3D Point Cloud Analysis (VirtuaLearn3D)

地形激光扫描（LS）是一种领先的遥感技术，用于获得地球表面及其物体的详细3D点云表示。虚拟激光扫描（VLS）模拟在计算机环境中再现LS采集的真实世界场景。当真实的实验不可行时，例如由于技术、经济和物流限制，VLS是有用的。机器学习，特别是监督式深度学习的最新进展表明，改进复杂自然对象（例如植被）和场景（例如地貌环境）的地理3D点云分析具有巨大潜力。深度学习算法的成功在很大程度上取决于高质量和适当大量训练数据的可用性。该项目的主要目的是推进虚拟激光扫描的概念，以解决缺乏训练数据的问题，从而为地理点云分析提供强大的机器学习算法。(1)一个关键的目标是找到有效的组合的真实的LS数据与理论上无限数量的模拟LS数据的监督训练。有效的解决方案可以缩小模拟数据与真实的数据之间的差距，并在减少昂贵的输入数据的同时保持较高的分类精度。(2)此外，我们想知道VLS数据生成在多大程度上可以支持迁移学习策略，以便使用预训练模型来迁移到不同地理特征和LS数据类型。由于LiDAR技术在科学和日常生活设备中的可用性急剧增加，因此非常需要VLS-supported transfer learning。(3)将开发和测试VLS模拟中的“动态对象”的新概念，这使得例如包括具有物候变化的植被以及移动对象（例如植物、汽车、人）。这一拟议的方法步骤将推动VLS模拟在机器学习中的大规模使用，并开辟全新的应用领域。该项目将侧重于机载激光扫描（包括。无人机载LS）和基于对象的树种分类和语义城市场景分类的任务，虽然开发的通用概念的相关性不受调查的例子。

Topographic laser scanning (LS) is a leading remote sensing technique to derive detailed 3D point cloud representations of the Earth’s surface and its objects. Virtual laser scanning (VLS) simulations recreate real-world scenarios of LS acquisitions in a computer environment. VLS is useful when real experiments are not feasible, e.g. due to technical, economic and logistic constraints. Recent advances in machine learning, in particular supervised deep learning, indicate a huge potential to improve geographic 3D point cloud analysis of complex natural objects (e.g. vegetation) and scenes (e.g. geomorphological settings). The success of deep learning algorithms strongly depends on the availability of high-quality and appropriately large amounts of training data. The main aim of this project is to advance the concept of virtual laser scanning to tackle the lack of training data to enable powerful machine learning algorithms for geographic point cloud analysis. (1) A key objective is to find effective combinations of real LS data with theoretically unlimited amounts of simulated LS data for supervised training. Effective solutions can close the reality gap from simulated to real data and keep high classification accuracy while reducing costly input data. (2) Furthermore, we would like to find out to what degree VLS data generation can support transfer learning strategies to enable the usage of pre-trained models for transfer to different geographic characteristics and types of LS data. VLS-supported transfer learning is highly demanded due to drastically increasing availability of LiDAR technology in sciences and also on daily-life devices. (3) A new concept of ‘dynamic objects’ in VLS simulations will be developed and tested, which enables e.g. to include vegetation with phenological changes and also moving objects (e.g. plants, cars, people). This proposed methodological step will push large-scale usage of VLS simulations for machine learning and opens up completely new fields of applications. This project will focus on airborne laser scanning (incl. UAV-borne LS) and the tasks of object-based tree species classification and semantic urban scene classification, though the relevance of the developed generic concepts is not limited by the investigated examples.