Application of Satellite Data and Data Mining Algorithms in Estimating Coverage Percent (Case study: Nadoushan Rangelands, Ardakan Plain, Yazd, Iran)

Assessing and monitoring rangelands in arid regions are important and essential tasks in order to manage the desired regions. Nowadays, satellite images are used as an approximately economical and fast way to study the vegetation in a variety of scales. This research aims to estimate the coverage percent using the digital data given by ETM+ Landsat satellite. In late May and early June 2018, the vegetation was measured in Ardakan plain, Yazd province, Iran. Information was obtained by 320 plots in 40 transects and also, the satellite images in terms of sampling time were downloaded and processed in USGS website. 16 indices involving NDVI, NIR, MSI, SS, IR1, MIRV1, NVI, TVI, RAI, SAVI, LWC, PD322, PD321, PD312, PD311 and IR2 were estimated. Through estimating the indices and extracting the values in order to conduct index-based predictions, six data mining models of Artificial Neural Network (ANN), the K Nearest Neighbor (KNN), Gaussian Process (GP), Linear Regression (LR), Support Vector Machine (SVM) and Decision Tree (DT M5) have been applied. Model assessment results indicated high vegetation estimate efficiency based on the indices but the model KNN with Root Mean Square Error (RMSE= 2.520) and Coefficient of determination (R2= 0.94) and (RMSE= 2.872 and R2= 0.96) had the highest accuracy in the training and data sets, respectively. As well, to determine the weight and importance of parameters, and to estimate the coverage percent, the weighing process were conducted based on support vector machine. Weighing results indicated that the KNN model and the Simple Subtraction (SS) index had higher weight and importance in terms of vegetation percent.