| Tea is a special economic crop which is widely distributed in tropical and subtropical regions,with high economic value and ecological benefits.As a tea producing country with a long history of tea planting and tea culture,the development of Chinese agriculture and rural economy is largely affected by the tea industry.Rapid and accurate acquisition of the fine spatial distribution of tea plantations and tea yield can provide important technical support for the monitoring and management of tea plantations,thus promoting the sustainable development of the tea industry.However,it has always been a technical problem.Therefore,taking Xinchang County,Zhejiang Province as an example,this paper constructed an effective method to extract tea planting areas,and then monitored the distribution of tea plantations in a long time series,and finally predicts the fine spatial distribution of tea yield.The main work and achievements of this paper are as follows:(1)First of all,according to the different characteristics of tea plantations in different periods,an R-CNN deep learning method that comprehensively considers spatial and temporal information was constructed to identify and extract tea plantations from temporal remote sensing images,and a comparative experiment was carried out using SVM,RF,CNN and RNN methods,and the performance was evaluated by F1 score and Io U evaluation index.The results showed that the R-CNN method had high applicability in the tea plantations extraction task.Its F1 score and Io U value on the test dataset were 0.885 and 0.793,respectively,which were 0.03-0.111 and 0.046-0.161 higher than the other methods.At the same time,visualization of the prediction results showed that this method can not only effectively distinguish the tea plantations from other ground objects,but also reduce noise and increase smoothness of the boundary of tea plantations.Its advantages of fastiness and accuracy can provide certain technical support for the fine mapping task of tea plantations.(2)Then,the multi-source satellite data is synthesized into the annual multi-temporal remote sensing images,and the R-CNN method was used to extract the distribution of tea plantations in Xinchang County from 2000 to 2021.At the same time,the result was compared with the statistical yearbook data,and the distribution characteristics of tea plantations in Xinchang County were further analyzed.The results showed that the method had good prediction accuracy,and the relative error of the extraction results for each year was mainly within 10%,which is of great significance for carrying out tasks such as reasonable planning and monitoring changes on tea plantations.In addition,the distribution index of tea plantations in Xinchang County is the high in the gentle slope mountainous area of 300-500m,while the distribution index in the high altitude or steep slope areas is the lowest.(3)Finally,combining the fine distribution data of tea plantations in Xinchang County with various source features such as meteorological,vegetation index,terrain,and soil data,eight machine learning models were evaluate and compared to carry out the research on tea yield prediction,and finally the fine spatial distribution of tea yield in Xinchang County was predicted.The results showed that it is effective to use multi-source characteristic data to predict tea yield,among which XGB model has the best performance,with the R~2 score on the test dataset reaching 0.880.It can not only accurately capture the nonlinear relationship between tea yield and various influencing factors,but also has high robustness.In addition,the fine distribution of tea yield could reflect the yield situation in various tea plantations to some extent.This study can provide a reference for government decision-making and tea farmers to prepare in advance to cope with the changes in production,and promote the development of tea industry. |
PDF Full Text Request