Research And Application Of Flood And Forest-fire Recognition Technology Based On UAV Aerial Image

With the rapid development of UAV technology,UAVs have the characteristics of lower takeoff limit and more flexible deployment mode,using UAV as a imaging platform of disaster recognition has become an important development trend.Meanwhile,deep learning flourishes in the field of images.It is of great practical significance to study the use of deep learning methods to mine rich information in aerial high-resolution images of UAV in disaster scenarios.At the same time,it also brings severe challenges: UAV aerial images offer top-views,such limited views pose challenges to classify similar appearance targets at the pixel level?Ground objects become much smaller,which poses challenges to fine segmentation of small targets? Complex disaster scenarios,more chaotic features and coexistence of multi classes of ground objects pose challenges to the recognition of image semantic information.Considering the above challenges,This thesis proposes dual attention mechanism based and multi-task learning mechanism based model structures,designing and implementing the relevant application system,conducting the model testing and experiments on Flood Net and Flame dataset.The main research achievements and innovations are as follows:(1)In view of the fact that the traditional DCNN cannot fully mine the global feature information and their feature expression ability is insufficient,this thesis proposes awareness enhanced dual attention module(AEDAM)based on DANet and a segmentation model based on the dual attention mechanism in disaster scenarios.For the spatial attention,the adaptive awareness ability of important feature information in the spatial dimension is strengthened by combining deformable convolution? For the channel attention,the feature information of space and channel dimensions is crossed through the combination of Strip Pooling and MLP to strengthen the awareness ability of the accurate position of the object of interest.(2)In view of the fact that the ambiguous boundary segmentation and missed cut small objects in aerial disaster images,this thesis proposes a multi-task learning model for joint detection and segmentation in disaster scenarios(DSDNet).This thesis constructs a boundary extraction sub-network as a bypass on the basis of the regular segmentation network.This sub-network will learn the boundary information and improve the model’s ability to extract the boundary information of objects.In addition,this thesis also introduces the object detection task branch based on the above structure to further optimize the fine segmentation ability of semantic segmentation task,and eagerly explores the relationship between the object detection task and the semantic segmentation task in the pioneering field of UAV disaster recognition.(3)Based on the above model,this thesis designs and implements a disaster recognition system based on UAV aerial images to conduct practical application research on the model proposed in this thesis.For the segmentation model based on the dual attention mechanism,the experimental results show that,compared with the benchmark model Deep Labv3+,the use of AEDAM can effectively improve m Io U(from 48.32% to 50.92%),and the additional use of pseudolabel method(To make full use of the data,Allowing the model to label unlabeled data as close to the real label as possible)can further increase m Io U to 51.01%? For the multitask learning model,the experimental results show that,compared with the benchmark model Deep Labv3+,its m Io U has increased by 1.98%? At the same time,compared with the single task detection branch,its m AP increased by 1.8%? For the disaster recognition system,it mainly includes visualization interface,disaster recognition module and so on,which is convenient to realize automatic UAV disaster recognition.