| Over the past few decades,high demand for tea has driven the expansion of tea plantations in tropical and subtropical regions.Tea planting promotes economic development and creates jobs,but the expansion of tea planting has a significant impact on biodiversity,carbon and water cycles,and ecosystem services.Mapping the spatial distribution of tea plantations in time is of great importance to land use management and policy formulation.However,the extraction and mapping of tea plantations in tropical and subtropical regions are faced with great challenges due to the lack of obvious phenological characteristics of tea plantations.Therefore,based on the Landsat ETM + / OLI and DEM as the data source,with the main tea producing areas in yunnan province（Lincang city,Xishuangbanna Dai Autonomous Prefecture and Pu’er city）as the research area,by using the field survey data of the area and the data of various typical ground features manually vectorized on Google Earth,the time series plots of various vegetation indexes of three typical ground features（tea plantations,natural forest and cropland）are plotted on the GEE platform,analysis of three kinds of features and different vegetation index in the tea plantations phenological period separability of tea plantations and other features.Firstly,the ASD spectrometer was used to measure the spectral of unpruned and pruned tea plantations samples,and it was found that the spectral of unpruned tea plantation and after pruning had significant differences in the red band,near infrared band and short-wave infrared band;Through spectral inversion of pruned and unpruned tea plantation samples by Landsat,it was found that there are also significant differences in the above three bands,and the variation trend is consistent with the measurement results of the spectrometer,which proves that Landsat images can be used to extract tea plantations in large regions.In this paper,a total of 3,815samples（training sample: 1125 pieces;verification samples: 2,690 pieces）,including2,810,891 pixels,were obtained by combining field investigation,photos coordinates,secondary survey data and visual interpretation of images.For each kind sample,stratified random sampling was conducted and 30% of each kind sample was selected as training samples for time series analysis.Based on the sample set,a variety of vegetation indices time series related to the above three bands were constructed to analyze the change of tea garden time series from 2015 to 2019（a total of 1508 landscape images）.Finally,NDVI（300-345d）,LSWI（30-90d）and DBNL（60-120d）were determined to be the best indices and corresponding time window for tea plantations extraction.In addition,during the pruning phenology period of tea plantations in the study area determined by time series analysis,the supervised classification（classification regression tree,random forest and support vector machine）was carried out based on 2018 Landsat data mosaicked with a high-quality image.Firstly,the importance of the constructed 18-dimensional spectral features was evaluated.Since the feature dimension was not high and each feature had a certain contribution to tea plantations extraction,the 18-dimensional features were all used for supervised classification.Then,taking Xishuangbanna as the test area,we adjusted the proportion of training and validation samples,trained the random forest model with samples in different proportions,and determined the best proportion of training samples and validations samples by comparing the user accuracy,producer accuracy and F1 score by extracting tea plantations in different proportions.Finally,the distribution map of tea plantations in the study area in2018 were mapped,and the extraction results of tea plantations of multiple classification and binary classification by the three supervised classification methods were verified and compared with the results extracted by the threshold method.The main conclusions are as follows:（1）Based on time series phenological period method for tea plantations extracting,the producer accuracy is 91.56%,which is higher than random forest（+3.62%）and support vector machine（+1.84%）respectively;The user accuracy is 80.65%,which is higher than that of random forest（+20.31%）and support vector machine（+32.74%）.Kappa coefficient（0.85）,higher than random forest（+0.14）and support vector machine（+0.23）,respectively.（2）By adjusting the proportion of training samples and test samples,random forest was used for classification.By comparing the user accuracy,producer accuracy and F1 score from the tea plantations in different proportions,the user accuracy and producer accuracy were the most balanced when the training sample/verification sample was 3:7,while the F1 score was also relatively stable in different proportions.Determining the training/validation as 3:7 as the sample allocation scheme in this paper.（3）In the multi-classification extraction of mountain tea plantations based on three supervised classifications（random forest,support vector machine and classification regression tree）,the producer accuracy of random forest extraction and support vector machine extraction of tea plantations has little difference（PA: 65.19%;66.80%）,the Kappa coefficient was the same（0.87）,but the user accuracy of extracting tea plantations from random forest was much higher than that of extracting tea garden from support vector machine（UA: 58.58%;43.24%）,and the accuracy of the three indexes were all higher than that of the classification regression tree.In general,the extraction accuracy of random forest is better than that of support vector machine and classification regression tree.（4）In the extraction of mountain tea plantations based on binary,the producer accuracy of random forest was slightly lower than that of support vector machine（PA:87.94%;89.72%）.However,the user accuracy of random forest for tea plantations extraction is 12.43% higher than that of support vector machine and 0.09 higher than that of Kappa,which are both higher than that of classification regression tree for tea plantations extraction.Considering the overall performance of the best is still random forest;Moreover,the user accuracy and mapping accuracy of tea plantations extraction based on the supervised classification method are higher than those of the corresponding classifier’s multi-classification method.The four types of ground features,including forest,water body,farmland and impervious layer,were combined into one category,which averaged the differences among the four types of ground features and increased the differences between the tea plantations and the forest and cropland.Therefore,the classification accuracy of tea plantations extraction that binary classification is higher than multi-classificationTaking three major tea-producing areas in Yunnan as the study area,this paper analyzed the tea plantations with low natural phenology for a long time series,and found that the "artificial phenology" caused by the pruned of the tea plantations was an important feature that distinguished from other land features.Then,NDVI,LSWI and DBNL were used to extract the tea plantations.In addition,based on time series analysis of tea plantations phenological period,1 scene high quality image was mosaiced to extract tea plantations by using supervised classification method,and compares the different methods to extract the accuracy of the tea plantations.It is concluded that the extraction accuracy of tea plantations based on phenological period of time series is higher than that of traditional supervised classification.This paper provides a new method for high precision extraction and mapping of tea plantations in complex tropical landscape mountainous area only using open source Landsat data at only 30 m spatial resolution. |
PDF Full Text Request