| Based on the Google Earth Engine cloud computing platform,we obtained Landsat TM/ETM+/OLI data with lower cloud cover less than 15% from 25 years from 1987 to 2019,and used the machine learning method-random forest classification method to extracted the long-term series land use/land cover information of Huangshui river watershed.According to the characteristics of the long-term series data research method,the preliminary classification results obtained by using the spatio-temporal consistency test were respectively processed for urban consistency and cultivated land consistency to reduce the classification logic Wrong.Finally,the 25-year land use/land cover data of the research area with higher accuracy was obtained through the accuracy assessment.The main conclusions of this paper are as follows:(1)The Google Earth Engine platform is an excellent platform for massive remote sensing data research.It can realize the whole process of work from data collection,data preprocessing,data processing to result output,which greatly improves the efficiency of scientific research.With the assistance of the Google Earth Engine platform,taking this article as an example,the task volume of a conventional stand-alone remote sensing platform for about 1 year can be compressed to about 3months,which improves the work efficiency by about 60%,making the resources scarce Individual scientific researchers can also complete scientific research tasks that previously required teamwork.(2)In supervised classification learning,the effect of the number,quality and spatial distribution of samples on the accuracy of classification results is greater than the improvement of the accuracy of the feature vectors introduced later.Taking the classification results in 2018 as an example,when performing feature optimization,sample selection and location optimization,texture feature and optimal window selection,climatic factor selection and optimal texture combination selection were carried out.The results show that only in the first step,sample selection and locationoptimization,the accuracy improved from about 70% to 81.33%,with an increase of up to 7.32%,and after the last three optimization steps,the final classification accuracy in 2018 was obtained It is 82.93%,and the increase is only 1.6%,which is not much higher than the former.(3)Based on the GEE cloud platform,through image acquisition,feature optimization and random forest classification,the initial data of land use/land cover classification from 1987 to 2019 was obtained.The overall accuracy was between72.25% and 90.20%,and the Kappa coefficient was between 0.68 and 0.89.Affected by the image quality,there will be obvious logical conflicts and errors in the same features in the classification results of the earlier years in consecutive years,and post-classification processing is required to correct the errors.(4)The research method of land use/land cover classification for long time series with high time resolution is different from the research method of interval years,and it is necessary to carry out logical detection of the years before and after.On the basis of drawing on the ideas and algorithms of the previous time and space consistency check of urban impervious surface,for the Huangshui River Basin,in addition to the city time and space consistency test,the spatial and temporal consistency processing algorithm of cultivated land is further proposed.Mainly through time filtering and logic judgment,the town and cultivated land are classified and processed successively,and their accuracy is verified.The results show that the classification results processed by the time-space consistency algorithm are more in line with the development logic of normal things.The overall accuracy is between 74.72%-87.36%,and the Kappa coefficient is between 0.71-0.86.Except for individual years,the overall accuracy is affected by the accuracy of the algorithm.The accuracy has decreased,and the remaining years have increased by 0.41%-3.44%. |
PDF Full Text Request