Remote Sensing Extraction Of Plastic Greenhouse Information In Weifang City Based On Deep Learning And Estimation Of Its Impact On Soil

Plastic greenhouses were representative facilities for modern agriculture and were of great significance for the development of new types of agriculture.Greenhouses effectively improved crop yields and economic benefits by changing the planting environment.Therefore,monitoring and mapping of greenhouses was essential to promote agricultural industry adjustment and promote rural revitalization.Most of the existing greenhouse extraction models distinguish greenhouses from other ground objects by spectral indices,which were subject to images of spectral response changes due to different factors such as seasonal materials,resulting in limited extraction accuracy.The deep learning model required prior knowledge to train the optimal parameter model,and a large and accurate plastic greenhouse data set was particularly important.When extracting greenhouses through deep learning models,usually only a single piece of information could be obtained,and the number and area of greenhouses could not be extracted at the same time.And due to the dense distribution of plastic greenhouses,the problem of overlapping horizontal detection frames was brought about.Weifang City was a typical greenhouse vegetable growing area in China and the largest vegetable production and distribution center in northern China.Long-term greenhouse cultivation threatened the health of the local soil,which in turn harms crop yields.In this context,quantitative analysis of the contribution rate of greenhouse distribution in Weifang City to the spatial differentiation of soil factors had important practical significance for local sustainable agricultural development and environmental protection.This study first produced a public dataset of plastic greenhouse remote sensing image training to meet the needs of greenhouse extraction.Then,a dense scene oriented frame detection model was built.Through the establishment of oriented frame,it could fit the outline of the greenhouse,and achieved the purpose of simultaneously extracting the number and area of greenhouses.Then,taking Weifang City,a typical greenhouse planting area,as an example,the remote sensing extraction and mapping of the greenhouse data in the research area in 2021 was carried out.On this basis,the contribution rate of greenhouse distribution in Weifang City to the spatial differentiation of soil factors was calculated using kernel density estimation and geographic detectors,and the relationship between greenhouses and soil was explored by selecting greenhouses of different construction scales and soil depths.The main conclusions are as follows:(1)A dataset of remote sensing images of plastic greenhouses was produced.The dataset was selected from four typical provinces of China’s greenhouse distribution,and produced using three-channel remote sensing images.At the same time,the remote sensing images came from multi-source data,three different sensors: Google Earth,Gaofen-2,and Jilin-1.All the annotation boxes in the dataset were marked with the five-parameter method and used angle regression.The final dataset generated a total of 2101 image tiles,including a total of 23914 labeled greenhouses.And set37.9% dense scenes and rich background scenes for the dataset.Experiments showed that the produced dataset could effectively train the extraction ability of the target detection model for greenhouses,especially the performance in dense scenes.It could serve a variety of tasks such as greenhouse monitoring mapping and intensive detection method development.(2)Extraction of plastic greenhouses in Weifang City.A dense scene oriented frame detection framework was built.The framework improved the ability to build feature pyramids by designing a novel residual structure and introducing channel and spatial attention;enhances the robustness of the framework by dynamically balancing hard samples and positive and negative samples;used Varifocal loss nested Kullback-Leibler Divergence loss function to achieve accurate detection of oriented boxes.The framework was compared with the four recent object detection models of BBAvectors,Faster RCNN,Retinanet,and YOLOX.The average precision of the five models was ranked as: dense scene oriented bounding box detection framework(87.66%)> BBAvectors(86.05%)> YOLOX(83.08%)> Faster RCNN(68.87%)>Retinanet(66.64%).The proposed model had the highest extraction accuracy,the best extraction effect,and the smallest performance loss in dense scenes.In addition,the dense scene oriented bounding box detection framework had strong extraction capabilities and excellent adaptability in the field of multi-object,large-scale and even text extraction.Therefore,this study used the dense scene oriented bounding box detection framework to extract greenhouses from remote sensing images in the study area.(3)Analysis of greenhouse distribution characteristics in Weifang City based on optimal hyperparameter model and kernel density estimation.The dense scene oriented bounding box detection framework extracted a total of 507,100 greenhouses in the remote sensing images of Weifang City,and the greenhouse coverage area reached 535.96 square kilometers.Among them,228,100 greenhouses had a construction area of more than 1,000 square meters.The distribution density of greenhouses in Weifang City gradually weakened from the northwest to the southeast,and the greenhouses were mainly concentrated around Shouguang,which was in the plain area.The distribution characteristics of multi-point aggregation in the southern hilly area.Large and medium-scale greenhouses were mainly constructed in the northwest of the study area,while the construction of small-scale greenhouses was mainly located in the east of the study area.And overall it showed the characteristics of being far away from the urban area and distributed along the river.(4)Based on geographic detectors,the contribution rate of greenhouse distribution in Weifang City to the spatial differentiation of soil available water content,soil bulk density,soil organic carbon content,soil p H,soil carbonate content,and soil sulfate content was calculated.The distribution of greenhouses in Weifang City had the highest contribution rate of 76.02% to the spatial differentiation of soil available water content,ranking first.For the remaining five factors of the top soil,the ranking of the contribution rate of greenhouse distribution was soil bulk density,p H,sulfate,organic carbon content and carbonate.For the remaining five factors of bottom soil,in the ranking of the contribution rate of greenhouse distribution,the organic carbon content exceeds the soil bulk density to became the first,and the p H and sulfate ranked the third and fourth respectively.Greenhouse distribution had inconsistent effects on top and bottom soils.In the meantime,the contribution rate of large greenhouses to the spatial differentiation of soil factors was higher than that of small and medium greenhouses.The research results could provide data support for Weifang’s future agricultural greenhouse planning and environmental protection.