Reaserach On Land Cover Classification Based On UAV Images

UAV(unmanned aerial vehicle)technology has become an important surveying and mapping method in recent years with its good maneuverability,and has been widely used in topographic surveying and mapping,remote sensing mapping,and geographic and national conditions monitoring.The high-resolution image data obtained by UAV contain abundant ground object information.However,due to the large amount of data,few bands,and large spectral differences of the images of the same land type,it is more difficult for computer automation processing.The automatic extraction of land cover information based on UAV images faces more challenges.Although there are more and more studies on object-oriented sub-meter image classification,there are still over-segmentation of similar plots in image segmentation,and the classification accuracy of surface coverage needs to be improved.To solve this problem,this paper uses the separation threshold method to select the best feature in the UAV image segmentation algorithm,and improves the FNEA segmentation algorithm.Through the experimental comparison of three machine learning image classification methods,the CNN classification model is used as the basis of high resolution image classification of UAV.A multi-scale CNN model is constructed,and the entropy weight method is used to optimize it,which further improves the classification accuracy of UAV images.The main research contents and results of this paper are as follows:(1)An improved FNEA segmentation method based on feature selection is designed.This method selects the best vegetation index and texture feature in UAV images by improved Seperability and Thresholds method,and uses FNEA algorithm to segment the image together with the original RGB data.The improved method is compared with the original FNEA algorithm and SLIC algorithm for segmentation experiments.The results show that the improved method has higher segmentation accuracy,and the over-segmentation in the same plot is well controlled,which is suitable for the segmentation of UAV images.(2)Three machine learning image classification methods are compared by experiments.Using the UAV image and ground measured data of a village in Hunan Province,the RF,SVM and CNN models were used for classification experiments.The results show that the RF model is better than the SVM model in classification efficiency and classification accuracy.Although the classification efficiency of CNN model is not high,its accuracy is higher than that of RF model,and it does not require artificial construction of feature set.Under the data conditions of this paper,the CNN model has obvious advantages,so the CNN model is used as the basic method of land cover classification of UAV image.(3)The optimization method of multi-scale CNN model based on entropy weight method is studied.In order to further improve the classification accuracy,a multi-scale CNN model is constructed,and a model optimization method based on entropy weight method is studied.Through the experiment,the optimized multi-scale CNN model classification accuracy were higher than those of single scale CNN model before optimization.It is proved that the model optimization method based on entropy weight method is conducive to improving the classification accuracy of images.In this paper,aiming at the segmentation and classification of UAV highresolution images,the FNEA segmentation algorithm is improved by adding the best features,and the CNN model is optimized by the entropy weight method.The experimental results show that the proposed method can improve the quality of land segmentation and the classification accuracy of land cover,which provides technical support for the accurate extraction of land cover information in UAV images.