摘要
建设项目属于复杂语义场景,其自动识别是水土流失动态监测和水土保持监管工作的技术难题.高分辨率遥感影像中的建设项目缺乏统一的语义概念定义,其场景包含多种人工和自然地物,场景内部高度非结构化、图像特征差异显著.笔者提出一种基于细部地物组合检测的建设项目场景识别方法:首先在制作用于目标检测的建设项目及其细部地物数据集的基础上,选择高信息量细部地物用于目标检测;然后采用Faster RCNN算法分别检测建设项目及高信息量细部地物,并采用预测结果框合并和细部地物组合修正的方法,来共同提高建设项目识别置信度,改进检测结果.实验结果表明,该方法制作的武汉市建设项目数据集的精度评价指标均优于其他对比方法,其平均精度值和平衡F1分数分别达到0.773和0.417.该方法对于复杂语义场景下的建设项目能够获得较好的识别结果,可应用于建设项目水土保持全覆盖监管.
Abstract
[Background]Construction projects belong to complex semantic scenes,and their automatic recognition is a technical challenge for dynamic monitoring of soil erosion and supervision of soil and water conservation.The construction projects in high-resolution remote sensing images lack a unified semantic concept definition,and their scenes contain a variety of artificial and natural features,with highly unstructured and significantly different image features inside the scenes.Therefore,it is necessary to study the target detection method for Construction projects.[Methods]We proposed a target detection method and theoretical system for complex semantic scenes of construction projects.GF-1 remote sensing image of 2 m resolution was used for annotation.Firstly,based on the construction projects data and its detailed ground object dataset for target detection,we selected high-information detailed ground objects for target detection according to the information content.Then,the Faster RCNN algorithm was used to detect construction projects and high-information detail ground objects separately,and the prediction result frame merging with detail ground objects combination correction was used to jointly increase the confidence of construction projects identification and optimize the detection results.[Results]Wuhan construction projects dataset is built,including construction project,bare land(rock),cover,construction road,prefabricated house,construction structure and built building,which amount is 752,763,154,82,372,292,and 278,information content is 18.81,20.96,9.93,44.82,28.77,and 8.22,respectively.Comparing this method with three other methods under the same experimental conditions,including Faster RCNN,Yolo,and variation of this method.The experimental results show that the accuracy evaluation indexes of the method on the produced Wuhan construction projects dataset are better than other comparison methods,and its AP(average precision)value and F1 score reach 0.773 and 0.417,respectively.The AP values of the three other methods were 0.755,0.693 and 0.754,and the F1 scores were 0.415,0.361,and 0.405,respectively.Compared with the other three methods,all of them have a certain degree of improvement.This method can effectively reduce the rate of wrong detection and improve the coincidence of correct detection results.[Conclusions]Better recognition results for construction projects in complex semantic scenes can be gained by this method.By the application of this method we can accurately and effectively identify the construction project,and by comparing it with the water and soil conservation program that has been reported,it can determine whether the construction project is built before approval and disturbed beyond the approved boundary,so as to achieve full coverage supervision of the construction project.
维普
万方数据