Comparative Study of Multiple Supervised Classification Methods to Map Land Use in Local and Regional Scales (Case Study:Kan and Karaj Rivers Basin)

Land management leading to sustainable development requires reliable and update data on land cover/use and mapping its changes at various spatial and temporal scales. In this aspect, water resources management also needs to assess land use and its changes across the basin to maintain water quality for a variety of uses. Thus, the primary goal of this study is to evaluate the effectiveness of various spectral-based supervised classification methods of Operational Land Imager (OLI) data for mapping land use across the Kan and Karaj Rivers basin. At the Anderson Level 1 and 2, the basin’s land use was mapped in five and nine classes, respectively using a broad range of different supervised classification methods, including Parallelepiped, Minimum Distance, Mahalanobis Distance, Maximum Likelihood, Spectral Angle, Binary Encoding, Spectral Information Divergence, Neural Net and Support Vector Machine. All classification methods were verified using the Google Earth images and accurate ground control points, in which the Maximum Likelihood method of both levels with Kappa coefficient of 0.77 and 0.76 and overall accuracy of 84.94 and 80.70 percent, identified as the optimum method to map the land use at the local and regional scales respectively. In addition, following the named method, the Neural Net, Support Vector Machine and the Mahalanobis Distance methods also showed acceptable accuracy indicating that like the choice of classification method, precision in procedures and accuracy assessment of land use classification map is very important and could affect the results.
Extended Abstract
1-Introduction
 Sustainable land management requires reliable and update data on land cover/use and mapping its changes at various spatial and temporal scales. In this aspect, water resources management also needs to assess land cover/use changes across the basin to maintain water quality for a variety of uses. Optical remote sensing plays a vital role in defining land cover/use and monitoring interaction between nature and human activities across various levels from local to regional scales. Furthermore, over the past decades, along with the development of computer science and information technology, different methods of land use mapping has been developed using remotely sensed data that provide cost-effective and accurate means to derive land cover/use resources information. Out of all factors that influence the complex process of land cover/use mapping, appropriate classification system and approaches affect the land cover/use outcomes. Thus, the primary goal of this study is to evaluate the effectiveness of various spectral-based supervised classification methods of Operational Land Imager (OLI) data for mapping land cover/use across the Kan and Karaj Rivers basin.
2- Materials and methods: The Kan and Karaj Rivers basin is located in Tehran province and South Alborz Protected Area in Alborz province. Having a perfect and cloud-free coverage, three Operational Land Imager images were used to map the land cover/use of the Kan and Karaj Rivers basin. They were extraxted from the United States Geological Survey in 2013 and the 2 to 7 bands with 30m spatial resolution were combined. Then, at the Anderson Level 1 and 2, the basin’s land cover/use was mapped in five classes including built-up land, farmland, rangeland, bare land and water bodies and nine including built-up land, crop land, orchard, good-quality, moderate-quality and low-quality rangelands, bare land and rock areas and water bodies using a broad range of different supervised classification methods, including Parallelepiped, Minimum Distance, Mahalanobis Distance, Maximum Likelihood, Spectral Angle, Binary Encoding, Spectral Information Divergence, Artificial Neural Net and Support Vector Machine using ENVI software. Two separate categories of points from Google Earth, as training and ground control points, were taken for image classification and the classification accuracy assessment, respectively. In addition, these points for each land cover/use class were obtained in accordance with land cover/use map (1:250,000) from National Cartographic Center. All classification methods were verified using the Google Earth images and accurate ground control points. Besides, confusion matrices including Kappa coefficient, producer’s accuracy, user’s accuracy and overall accuracy were applied to make accurate assessment of supervised classification. Furthermore, at the Anderson Level 1 and 2, the area of land cover/use classes of three methods was obtained. They have acceptable Kappa coefficient
3- Results and Discussion: The findings from accuracy assessment of classification at the Anderson Level 1 and 2 showed that the Maximum Likelihood method with Kappa coefficient of 0.77 and 0.76 and overall accuracy of 84.94 and 80.70 percent were identified as the optimum method to map the land cover/use of the Kan and Karaj Rivers basin at the local and regional scales respectively. In addition, following the aforesaid method, at the Anderson Level 1, Artificial Neural Networks and Mahalanobis Distance methods and at the Anderson Level 2, the Artificial Neural Networks and Support Vector Machine methods also showed acceptable Kappa coefficient in land cover/use mapping. Based on the findings, it could be concluded that the Artificial Neural Network method had better performance in the classification of land cover/use with different and high classes than other methods. Moreover, when the training points are low, the Support Vector Machine method had better performance and achieved to high accuracy than other methods. At the Anderson Level 1, the most and least differences between the land cover/use classes areas were related to rangeland and water bodies classes, respectively whereas at the Anderson Level 2, the most and least differences between the land cover/use classes areas were related to crop land and water bodies classes, respectively.
4- Conclusion: Based on the findings, Maximum Likelihood and Artificial Neural Network classification methods were more appropriate respectively, and had relatively stable performance than the other methods. So, apart of the method, acceptable land cover/use map only can be extracted by increasing the number of training points and accurately determining the location of training points in two Anderson levels. In summary, the results of this study indicated that like the choice of classification method, precision in procedures and accuracy assessment of land cover/use classification map are very important being able to affect the results.