Research On Cultivated Land Parcel Extraction Method Based On High Spatial Resolution Remote Sensing Image

With the rapid development of Chinese space information technology,the number of high spatial resolution remote sensing images continues to increase,providing strong data support for ground information acquisition in many fields.In the remote sensing of agriculture,artificial interpretation of vector cultivated land plot based on high spatial resolution remote sensing image is time-consuming and laborious.At the same time,conventional image segmentation is prone to over-segmentation or under-segmentation.Based on the investigation and research on the development status of image segmentation technology,combined with the current convolutional neural network in the field of image processing,a series of researches were carried out.The main contents are as follows:1.Using the GF1 remote sensing image as the data source to constructed the image-cultivated land plot edge data set.HED and RCF models suitable for edge detection in conventional images were retrained with this data set and applied for edge detection in high spatial resolution remote sensing images.Based on the analysis of those effective structure and lots of experiments,a fully dilated-RCF model was proposed,which achieved better edge detection.2.An edge-based method for extracting cultivated land plot was constructed.The automatic extraction was completed by performing multiple post-processing on the field edge results detected by the fully dilated-RCF model.This process mainly includes speckle removal based on morphological refinement,noise removal based on eight-neighbor edge tracking,edge closure based on main direction growth,and raster region vectorization based on run length coding.Compared with the image segmentation results from FNEA by e Cognition and Mean Shift,the extracted plots from our method were more regular and homogeneous and got the highest detection accuracy in QS,which over 0.1 and 0.2 respectively.Results showed that CNN was also effective in the processing of remote sensing images,and had great potential,which can effectively promote the development of automatic interpretation of high spatial resolution remote sensing images.