| Remote sensing technology is the core component of agricultural spatial information technology,and its macroscopic,timeliness and objectivity make it become one of the indispensable important means of agricultural resources monitoring and management.Water bamboo is an aquatic cash crop and an important aquatic vegetable in south China.Accurate monitoring of its planting area is of great significance to large-scale planting layout,information management and logistics distribution arrangement of vegetables.Traditional planting area information acquisition mainly adopts the method of layer by layer statistics,which is time-consuming and costly,and the result is not very accurate due to the interference of human factors,which can not meet the actual needs.Therefore,in order to meet the needs of production and management of water bamboo,it is urgent to develop a method to acquire planting information of water bamboo using remote sensing.At present,most of the researches on crop remote sensing identification are based on field crops,and less on cash crops.It only involves the pixel result level,and the study on the refined crop recognition based on the plot needs to be further explored.In this study,huangyan District of Taizhou city was selected as the main research area to obtain sentine-2 and Gaofen-2 images of water bamboo during its growth period,and analyze the planting characteristics,spectral characteristics,vegetation characteristics and color characteristics of single and multitemporal images of water bamboo,combined with the field sampling of planting samples.On the pixel scale,the optimal identification methods of water bamboo based on pixels were compared and screened out by five supervisory analysis methods,namely maximum likelihood method,minimum distance method,neural network method,support vector machine method and decision tree method.At the object oriented level,the multi-scale segmentation method based on eCogniton software and the analysis method based on U-NET deep learning were used to study a land-oriented remote sensing recognition method for planting information of water bamboo.Finally,based on the two dimensions of pixel scale and block-oriented,the matching analysis method of pixel and block-oriented was studied by analyzing the relationship between pixel and block vector.Finally,a remote sensing recognition method of water bamboo in southern mountainous area was formed.The main research results of this paper are as follows:1.Decision tree classification method is the supervision classification method of optimal water bamboo identification.For different ground objects and remote sensing data sources,each remote sensing classification method has its own applicability and limitations.To solve this problem,remote sensing images covering the whole growth period of water bamboo were used in this study,sampling points of field investigation and visual interpretation were mixed as training samples and validation samples.Based on the analysis of planting characteristics,color characteristics and spectral characteristics of remote sensing images,The maximum likelihood method,minimum distance method,neural network method,support vector machine method and decision tree method were selected to extract water bamboo area in the study area,and the pixel results of double bamboo bamboo in huangyan area based on five supervised classification methods were obtained.The Kappa coefficient and overall classification accuracy of the five methods were compared.The overall classification accuracy was 74.25%,75.76%,73.29%,81.34%and 83.35%,respectively.The Kappa coefficients were 0.7920,0.7872,0.8062,0.8035 and 0.8325,respectively.The results showed that the method of decision tree model was effective in extraction of water bamboo in the study area,and its pixel results were the best overall.2.The cultivation plot identification method of water bamboo based on U-NET deep learning is superior to the multi-scale segmentation method based on eCogniton software.As the classification accuracy of pixel results is poor,and the actual land boundary is mostly the boundary between crops and ridges,roads and trees,pixel results cannot truly reflect the current situation of crop planting.In order to solve this problem,research on information extraction of oAT was carried out.In the process of object-oriented classification,the standards and principles of training samples were determined first,and then standardized training samples were selected as the basis of object-oriented classification.Then,eCogniton software was used to extract information of water bamboo.After several experiments,multi-scale segmentation algorithm was selected for image cutting,and the nearest neighbor classification method was selected for classification method.During image segmentation,segmentation scale parameters were selected as 100,shape parameter 0.3 and exquisite parameter 0.7,and the overall classification accuracy was 65.25%.Kappa coefficient is 0.6920.The training model was established by Tensorflow with 4000 training times and Sigmoid activation function.The gPU-accelerated training was used to obtain the distribution results of planting information of water bamboo.The overall classification accuracy of this method was 82.25%and Kappa coefficient was 0.8372.By comparing the accuracy of the two methods,the distribution accuracy of water bamboo information extracted based on U-NET deep learning technology is the best,which is determined as,the best block-oriented water bamboo information recognition scheme.3.Construct a remote sensing,recognition method for water bamboo based on pixel and plot analysis.To solve the problem that pixel results and field vector results are different in dimension and cannot be analyzed in the same dimension,the collaborative research of pixel and field vector of zizania latifolia was carried out.On the basis of the pixel results of water bamboo shoot and the vector results of water bamboo plot,the cooperative method of pixel proportion was adopted to carry out the research,and the vector of water bamboo plot with pixel proportion greater than 80%was determined to be the actual planting plot of water bamboo shoot.Aiming at the problems of small pixel proportion and interference of boundary mixed pixels,two optimization methods for collaborative processing of pixels and block vectors were proposed.After the collaborative processing of pixels and block vectors,the overall classification accuracy of the three methods including boundary elimination method,quadrate method and the combination of boundary elimination and quadrate method was 93.05%,92.16%and 94.15%.The Kappa coefficients are 0.914,0.932 and 0.926,and the results are all better than the single pixel extraction results and the single object-oriented extraction results in terms of classification accuracy.However,there is little difference between the three methods,and the overall classification accuracy of the combination of boundary elimination method and quarto method is slightly higher than the other two methods.Therefore,the experiment proves that removing the boundary and reprocessing with the quartering method can effectively improve the accuracy of crop identification in proportion analysis. |
PDF Full Text Request