Research On 3D Reconstruction Of Karst Cliffs And Automatic Classification Of Plants Based On Unmanned Aerial Vehicle

The cliff ecosystem represents one of the least disturbed natural environments,characterized by its inaccessibility and extreme habitats,which harbor numerous ancient and unique plant species.Due to the challenging conditions found in cliff habitats,such as high elevation,steep slopes,and limited human accessibility,conducting surveys in these areas poses significant difficulties.While comprehensive and systematic information on cliff vegetation cover is currently lacking,obtaining such data is of paramount importance to environmentalists.Traditional survey methods,such as large-area normalized difference vegetation index statistics and small-area quadratic sampling surveys,are not inadequate for cliffs with nearly vertical formations.Therefore,this paper proposes a semi-automatic 3D reconstruction method for karst cliff systems,specifically targeting the karst cliffs of in Guilin.By generating a 3D model,this approach enables the estimation of cliff vegetation cover,facilitates vegetation cover assessment,and enables taxonomic positioning studies of plant species.The objective of this paper is to explore innovative and comprehensive survey methods for karst cliffs,providing a valuable reference for the survey plans concerning"inaccessible" and "inhabitable" cliffs.The primary research objectives of this study are as follows.(1)The combination of UAV tilt photogrammetry and UAV close photography was used to collect about 15,000 high-resolution UAV images of 20 cliffs in the karst region of Guilin,China,and to reconstruct high-precision 3D models of 20 cliffs.The data acquisition replaces the traditional single manual photography with automatic surround photography.The accuracy of modeling results is analyzed by the positioning accuracy of photo positions and texture structure.The results show that the positioning accuracy in X-direction is MAE=0.000001,RMSE=0.496377m;the positioning accuracy in Ydirection is MAE=0.000008,RMSE=0.775091;the positioning accuracy in Z-direction is MAE=0.000069,RMSE=0.946324;the texture of the cliff wall 3D model constructed by this method is clear and the 3D model of the cliff constructed by this method has clear texture and high precision,and meets the requirements of photogrammetry.(2)The vegetation cover of different elevation gradients(bottom of the cliff(gradient 1 to gradient 3),middle of the cliff(gradient 4 to gradient 7),and top of the cliff(gradient 8 to gradient 10))and four slope directions were calculated based on the maximum threshold method and the excess green vegetation index using the dense point cloud generated from the high-precision 3D model of the cliff wall.The extracted 3D model of the cliff face and the vegetation characteristics of the UAV images were also compared and analyzed.The results showed that the mean values of vegetation cover at the bottom,middle and top of the 80m cliff were 0.85,0.65 and 0.77,respectively;the mean values of vegetation cover at the bottom,middle and top of the 100m cliff were 0.84,0.71 and 0.82,respectively;the mean values of vegetation cover at the bottom,middle and top of the 120m cliff were 0.93,0.76 and 0.84,respectively;the mean values of vegetation cover at the bottom,middle and top of the 140m cliff were 0.93,0.76 and 0.84,respectively.The vegetation cover is less sensitive to the change of slope direction.Based on the vegetation feature extraction,we can find that the 3D model vegetation feature extraction results are better than the UAV image feature extraction results.(3)The Inception network of GoogLeNet deep learning framework was used as the training set among 1608 cliff face proximity drone images collected by the group to automatically classify the top ten dominant tree species from drone images of 20 modeled cliff faces.The results show that the training model achieves an average accuracy of 86.8%,an average recognition recall of 87%,and an average F1 score of 87%for plant classification.The loss-epoch curve of network fixed weight strategy training,with the increase of iterations,the model training loss decreases tends to 0.5,and the test accuracy keeps improving tends to 0.8.Overall,the model training effect is good and can be used with cliff vegetation classification and monitoring.In summary,this study proposed an efficient method for investigating the cliff vegetation cover of a single peak cluster in karst landscape with the karst cliffs in Guilin,Guangxi,China as the experimental object,and automatically classified and located the cliff plants by using 3D model and UAV close-up photography data,which improved the efficiency of plant identification and provided reference for cliff vegetation monitoring,and the research results showed the feasibility of the scheme.