基于多任务迁移学习的震后建筑物检测及受灾程度评估模型

Currently, building detection models based on remote sensed images are largely limited by the size and spatial resolution of the images, as well as the conditions that images are taken. The universality and portability of the models are poor. Moreover, the training methods are mostly single threaded, which ignore the relevant geophysical information. Faced with such shortages, we train a model that simultaneously detects buildings after earthquake and evaluates their damage level using multi-task transfer learning. This project will focus on the remote sensed images with spatial resolution between 0.1m-0.5m, where there occurs an earthquake above a magnitude M5.0. Firstly, we design a global feature that describes whole structure of buildings after earthquake. Secondly, a multi-task model is trained, whose main task is evaluating damage level and secondary task is detecting buildings after earthquake. We further apply the trained model to other high spatial resolution remote sensed images after earthquake. The universality and portability of the remote sensed model can be improved by merging the information from buildings of different damage levels. This project may provide a new idea and method to efficiently and intelligently take emergency response and evaluate disaster loss after earthquake.

由于目前基于遥感数据构建的建筑物提取模型在较大程度上受限于实验影像的尺寸、分辨率和拍摄条件，导致模型的普适性和可移植性较差。而且目前的模型训练方法多采用单线程训练模式,忽略了相关地物信息的影响。本项目针对这些不足，研究利用多任务迁移学习方法训练相关模型，对地震灾后建筑物进行检测及受灾程度评估。项目以发生里氏5.0级以上地震区域的0.1-0.5米分辨率的高分辨率遥感影像为研究对象，首先构建描述震后建筑物的整体结构特征，继而以评估建筑物受灾程度为主要学习任务，检测受灾建筑物为辅助学习任务，用多任务迁移学习思想训练模型，然后将训练好的模型应用在其他地震灾后高分辨率遥感影像中，对建筑物进行检测与受灾程度评估，并通过震后不同程度的受灾建筑物之间的相辅相成，改进模型精度，提高模型检验效果的鲁棒性和适应性，从而达到快速、准确检测评估地震灾害影响的目的。本项目可以为高效的震后应急处理和受灾评估提供新思路。

随着城镇化进程的加快和人口密度的增加，地震造成的建筑物损害已经成为制约城镇发展和居民生命财产安全的主要威胁之一。目前针对震后建筑物提取多集中于背景地物相对简单，建筑物形态较为单一的情况，对复杂背景地物条件下大量震后建筑物提取研究相对较少。本研究采用多任务深度学习语义分割模型，对海地2010年震后高空间分辨率遥感影像的建筑物进行提取和评估其受灾情况。对损坏建筑物提取的平均召回率和准确率分别达到81.14%和84.38%，对完好建筑物提取的平均召回率和准确率分别达到95.29%和92.06%。该方法通过引入多任务模型结构，同时实现了建筑物提取和受损标注。相较于单任务深度学习语义分割模型，多任务模型结构在完好建筑物和损坏建筑物提取方面，可以分别取得2%和6%的精度提升。研究表明，构建的多任务语意分割深度学习模型可以取得比较可靠的完好建筑物和损坏建筑物的提取精度，为地震发生后应急应用提供可能。