Speaker Bio & Abstract
Graduate School of Life and Environmental
Sciences University of Tsukuba Japan
AbstractPrabath Priyankara | Faculty of Agriculture, University of Ruhuna & Graduate School of Life and Environmental Sciences, University of Tsukuba
Nuwan Senevirathne | Graduate School of Life and Environmental Sciences, University of Tsukuba
Takehiro Morimoto | Faculty of Life and Environmental Sciences, University of Tsukuba Accuracy and efficiency are the key factors to consider during a land-use/cover modeling. There are many approaches available for land-use/cover modeling with the development of geospatial processing and modeling methods alongside the development of computer science. Land-use/cover modeling using conventional statistical methods are still popular, but it will take more time due to high human interaction to build a highly accurate land-use/cover model. With the development of faster machines with many cores of the central processing unit (CPU) and graphical processing unit (GPU), artificial intelligence (AI) allows computers to reduce human behaviors applied in land-use/cover modeling. Machine learning (ML) is a subset of AI where we can directly apply to improve accuracy and efficiency of land-use/cover models. Therefore, the principal objective of this study is to evaluate the accuracy of land-use/cover models based on Support Vector Machines (SVM), Decision Tree (DT) and Random Forest (RF) ML methods using free and open source software (FOSS) tools. Kandy district of Sri Lanka was used as the study area for this case study. Two Landsat 8 OLI/TIRS image, acquired on 13 January 2017 and 28 October 2017 with cloud coverage less than 10% those were acquired during the dry and wet seasons in order to account for seasonality or vegetation phenology in the classification and corrected for atmospheric effects. Five land use/cover classes (built-up, vegetation, agriculture, bare land, and water) were adopted for this study. Reference datasets were developed for classifier training and classification accuracy assessment with 100 samples for each land-use/cover class. The SVM, DT, and RF ML techniques were used for land-use/cover modeling by using R, FOSS software tool. The results showed that RF classifier had the highest classification accuracy of 79% (with Kappa statistic of 75%). Moreover, other two methods, SVM had a classification accuracy of 75% (with Kappa statistic of 71%), and DT had a classification accuracy of 64% (with Kappa statistic of 59%). RF classifier significantly increased the classification accuracy of both the SVM and DT classifiers by 5% and 15% respectively. Therefore, RF classifier can be recommended for land-use/cover modeling for mountain rich areas. Also, the developed R script can be re-used for further land-use/cover modeling very easily with few modifications on R, FOSS software tool.