An integrated comparative data driven approach for spatiotemporal land use/cover mapping and change detection analysis

Recent progress in remote seining sciences and improving the quality of satellite images in all aspects of spatial, spectral and temporal resolution provided large number of data whit demanded developing automated data driven and machine learning techniques. Thus, machine learning approaches in the scientific community have recived a significant interest,for imgae classification and environmental anlysis. As can be figured out, there are plenty of machine learning algorithms being employed for the image processing tasks. In order to evaluate the efficiency of within this research we intended to apply and compare the efficiency of two best known machine learning algorithm including support vector machine (SVM) and random forest (RF) for time series land use land cover (LULC) monitoring in the vicinity area of Urmia lake in north west of Iran. For this object, we employed time series Landsat satellite images on the platform of Google Earth Engine (GEE). We employed three methods for valiadtaion and accurassy assessment. For this goal, first the validation step performed using overall accuracy, kappa coefficient based on the grand control points collected in field operation as as well Fuzzy Synthetic Evaluation for computing the confidence level of classification in sub category of each data driven approach.In addition, in the second step the the Dumpster Shafer theory (DST) was applied to carry out the spatial uncertainty of obtained LULC maps. Results of accurassy assment and also uncertinity anlysis, pointed out that the SVM algorithm performed classification much efficiently rather than RF algorithm. According to the results of validation through ground control points and spatial uncertainty analysis using the DST, as a best performance the SVM could deliver the LULC classified map with the overall accuracy of 92.57% as well the spatial accuracy of 0.97. While, the best performance of the RF algorithm computed to be 86.20 % in overall accuracy and 0.88 in DST for the spatial uncertainty analysis. As these results from both validation methods confirm, there were extensive LULC change in the study area which essentially contributed to Urmia lake drought and respective environmental degredation.