Research On Semantic Segmentation Network UAV Image Greenhouse Extraction Method Based On Fusion Depth Information

In the application of precision agriculture,greenhouses are an important type of agricultural facility.Accurate greenhouses distribution and information acquisition are of great significance for precision agriculture planning,agricultural output estimation and relevant policy formulation.Traditional manual survey is time-consuming and laborious,while UAV remote sensing is a convenient and fast new surface information acquisition tool,which is very suitable for the quick acquisition of greenhouse distribution information.In this paper,a new semantic segmentation network based on depth information is proposed by using remote sensing technology of unmanned aerial vehicles(UAVs)to automatically and accurately extract the distribution information of greenhouses on images.The main work of this paper is as follows:(1)Carried out image collection and sample set generation of greenhouse UAVAccording to the data distribution and morphological characteristics of the greenhouses,the altitude of 400 m,the course overlap rate of 80% and the side overlap rate of the greenhouses were set,and the greenhouses area image of about 2.52 km in jiangning district,nanjing city in August 2019 was collected.A dense 3d point cloud in the experimental area was generated by the multi-vision matching technique,and then the depth information was mapped to the 2d image by affine change registration to recover the depth information in the experimental area.Then,the data of the processed uav images were marked,and a set of technical procedures for generating sample sets and image classification were developed to generate sample sets for network training.(2)A greenhouses extraction method based on deep information extended semantic segmentation network was proposedOn the basis of combing various classical semantic segmentation network structures,the basic network structures of U-net and SegNet are adjusted.The introduction of ASPP module,the removal of BN layer and the introduction of asymmetric convolution were improved,and the ACUnet network was built.Then,ACDUnet was built by taking the depth information as an independent branch and doing parallel operations with the spectral information to realize the expansion of the depth characteristics of the conventional semantic segmentation network.(3)The application test and comparative study of the method on UAV image were carried outThe method was applied to the greenhouse extraction in jiangning district,nanjing city,jiangsu province.The experiment shows that the performance of ACUnet without introducing depth information is better than that of U-net and SegNet.The performance of ACDUnet using depth information through fusion is better than that of ACUnet,and the performance is also better than that of ACUnet?C4 which directly uses depth information as the fourth channel.In general,ACDUnet has high accuracy and high robustness.From data acquisition to network training and model prediction,it does not need too much manual intervention,and the overall process is highly automated.