Research On Facility Agricultural Land Identification Technology Based On High Resolution Remote Sensing Image

The quantity and distribution of facility farmland affects the productivity of cultivated land,represents the level of agricultural production,and reflects the development of modern agriculture.It is an urgent need for agricultural production and management departments to obtain information on the spatial distribution of facility farmland in a timely and effective manner.However,due to the small sizes of objects,the large number of types and complex background environments of the facility agricultural land,it is also easily affected by other occlusions such as clouds,fog,shadows,etc.The traditional remote sensing classification methods based on shallow learning are difficult to meet the needs of business production departments.Based on high-resolution remote sensing images,the use of deep learning methods to classify and extract facility agricultural land has important application significance.In this study,the 0.5-meter high-resolution DOM image is used as the data source,and Xiangcheng Town,Lanling County,Linyi City,where the large-scale operation and management of facility agricultural land was relatively good,was selected as the research area.The facility agricultural land in the area was divided into three types: multi-span greenhouse,single arch shed and solar greenhouse,which were grouped by linear labels for sample labeling.The UNet deep learning model was used to complete the extraction of facility agricultural land in the study area.In order to further analyze and compare the effectiveness and robustness of the U-model,based on the object-oriented technology,this study used random forest and nearest neighbor classifier methods to extract the facility agricultural land in the study area,respectively.In view of the various situations that affect the classification accuracy in the classification process,this study focused on the sheltered and non-obstructed areas in the farmland,the sheltered and unobstructed areas in the construction land,and the area proportion of the agricultural land in the village.Accuracy comparison was carried out for the size and other conditions,and the conclusions obtained through the comparative analysis were as follows:(1)Among the two object-oriented classification methods,the overall accuracy of the nearest neighbor classifier was 73.49%,and the overall classification accuracy of the random forest classifier was 78.57%,which is significantly higher than that of the nearest neighbor classifier;the overall classification accuracy of the UNet model was 96.84%,which is significantly higher than the random forest classification accuracy,and its boundary fit is also significantly better than the random forest method.(2)In the unobstructed area of the farmland,the distribution of single arch sheds is relatively scattered.Due to the limitation of scale adaptability in the segmentation process of object-oriented classification,there would be mixed objects of facility farmland and cultivated land,resulted in obvious misclassification;There are areas in the farmland covered by trees and other objects,and the existence of shadows led to serious misclassification and omission of object-oriented classification results,incomplete extraction and poor boundary matching.The UNet model extracts the facility agricultural land well for the above situations.(3)When there is no occlusion in the construction land,due to the phenomenon that the spectral characteristics of buildings are similar to those of facility agricultural land,the objectoriented method misclassified roads and other buildings into facility agricultural land,and the classification results were poor;When occluded,the object-oriented method had a significantly poorer effect on the extraction of facility farmland in the shadow area,and the more severe the shadow,the more serious the leakage phenomenon.(4)When the proportion of facility agricultural land in the village increases,its area also increased,and the concentration was also higher,which would have a significant impact on the classification results.When the UNet model was used to extract various proportions of facility farmland,it could obtain better classification accuracy;the object-oriented classification method was affected by the changes in spatial arrangement and spectral characteristics due to the change of the local size of facility farmland,and its classification effect was far lower than the U-model.It can be concluded that the results obtained by the U-model classification were less fragmented,the boundary of the facility farmland was more consistent with the field,and was less affected by the occlusion of ground objects,changes in spectral characteristics and spatial characteristics,and in different types and areas.In this case,better classification results could be obtained.This paper has 27 figures,16 tables and 61 reference...