| The combination of different resolutions of remote sensing data to monitor in large scale forest area and its changes has significant advantages over traditional ground survey methods.The Democratic Republic of the Congo(DRC)is located in the heart of Africa.It has not only the largest tropical rainforest in Africa,but also several ecoregions such as Miombo woodland.Mastering forest resource data and its changes it is particularly important for the DRC,which is still relatively weak in this regard.In this study,two typical areas were selected as research areas,and two studies are carried out separately:1.A large-scale remote sensing sampling survey based on total probability transfer matrix is proposed.The steps of this method are as follows:(1)Landsat 8 data is used to carry out supervised classification of the entire study are;(2)using high-resolution data(Google Earth images)to perform systematic sampling,followed by visual interpretation of each single sample,whose results are considered as ground truth values;(3)combining each sample visual interpretation data and its corresponding TM classified data to establish a probability transfer matrix;(4)calculate the probability estimation of the area using probability transfer matrices and TM classified results.To compare the results,three kinds of probability sampling estimation methods are designed.(1)is called method 1,which is the method proposed in this research.The probability transfer matrix is calculated based on the area transfer matrix of all plots;(2)is called method 2,which is an existing method for which the probability transfer matrix is the average of the probability transfer matrix of a single plot;(3)is called the method 3,and the simple random sample estimation is performed only using visual interpretation sample.TM classified data and visual interpretation are divided into 7 land classes.There is a total of 25 scenes TM scenes,and a total of 112 plots(3km × 3km)which are visually interpreted.The classification accuracy is 84.52%.The overall probability sampling accuracy of method 1,method 2 and method 3 are 91.94%,92.21% and 86.29%,respectively.The method 1 and the method 2 use the information of TM classified data and visual interpretation data,and their estimation efficiency is high,while simple random sampling only uses the visual interpretation data,and its efficiency is relatively low.Also,method 1 is unbiased,and the result is more reasonable than method 2,although the overall sampling accuracy is slightly lower.2.Large-scale dynamic change monitoring.This study proposes dynamic change monitoring based on MODIS-NDVI data and robust regression methods.Four periods of data from 2001-2005,2005-2009,2009-2013,and 2013-2017 were selected for analysis.The basic steps are as follows:(1)establish a robust regression model between two consecutive periods based on all pixels(NDVI data);(2)determine the threshold according to the variance of the robust regression model,and diagnose the change value;(3)divide into four significant levels(α = 0.1;0.05;0.025 and 0.005)evaluate the number of pixels changed and determine the area of change.For each period,the results show that the area with decrease in vegetation cover(at different significance levels)is larger than the area with increase in vegetation cover;results that were confirmed by high spatial resolution images. |
PDF Full Text Request