COMPARISON OF THE SUPERVISED AND UNSUPERVISED CLASSIFICATION METHODS OF LAND COVER CHANGE DETERMINATION IN EASTERN IRAN BASED ON REMOTE SENSING DATA |
Paper ID : 1129-SMPR-FULL |
Authors: |
Rasoul Kharazmi *1, mansour karkon varnosfaderani2, Mohammad Reza Rahdari3, Evgeny Panidi4, Viktoriya Vasilievna5, Eugeny Mitrofanov6, Dimitry Lubmin6 1بلوار قدس، شهرک شهید قندی، ک. 15، پ. 10، ک.پ: 7617756843 2Moscow State University of Land Management 3University of Tehran 4Pilotov st. 18/4 app. 48 5Moscow state university of geodesyand cartography 6Moscow State University of Geodesy and Cartography |
Abstract: |
Different image classification and segmentation techniques are used in almost all of the thematic studies that implement the use of remote sensing data. Currently, we have a huge amount of satellite imagery available in near real time mode for monitoring purposes and extremely a huge data amount available as an archive for retrospective analysis and study of environmental interdependencies. This situation leads to the necessity of implementation of the highly automated machine learning techniques, which could be applicable for operations with big data (or at least big-size datasets) and could produce the valuable, comparable and representative results of data analysis. Unfortunately, it is unfeasible to build fully automated and autonomous analysis techniques, first of all, due to the high dynamics and high complexity of environment in whole and to the understudiedness in details of many environmental processes. The general way in remote sensing data analysis is to build the customizable algorithms and techniques, which makes it possible to configure and train the computational model to consider the study context and actual situation at the study area. In our case study, we use this approach to build an analysis tool for investigation of the land cover change at the eastern Iran, where the dry lands are presented in significant amount and exist a high risk of desertification. As a starting point for our study, we set the need of automation of the land cover types allocation and land cover change detection processes. Our study has a regional scale, and the core objective of described stage of the study is evaluation of accuracy of the vegetation indices based unsupervised classification method for monitoring of the change in land cover at the Sistan Basin in Iran. Two multispectral Landsat satellite imagery scenes (of years 1998 and 2013) were selected. The data were pre-processed and the false colour images of the study area were produced. To determine the benchmark land cover maps of the study area, we provided the supervised classification with training samples for both scenes, basing on our background of studies on this area. In order to adopt unsupervised classification technique to the study context, the Soil Adjusted Vegetation Index (SAVI) maps for both scenes were produced. These maps were used as the additional input data for unsupervised classification. As a result, the unsupervised classification calculations produced the three basic classes of land cover (which are vegetation cover, bare land and water bodies) with the Kappa statistics of 0.85 for both scenes were obtained. Our investigations show that adopted SAVI based unsupervised land cover classification technique produces the acceptable accuracy on this case study area, when determining basic land cover classes and land cover change detection. The intensive decrease of vegetated areas, loss of water, and arid lands increase lead towards desertification during the study period. Additionally we conclude that revision of related policies together with application of more efficient techniques of resource management are required to prevent migration of inhabitants of the region, taking into account that the study area is a border region. |
Keywords: |
Sistan Basin, Vegetation Index, Remote Sensing, Spectral Reflectance, Land Degradation, Classification Methods |
Status : Paper Accepted (Poster Presentation) |