Accuracy Comparison of Land Cover Mapping Using the Object- Oriented Image Classification with Machine Learning Algorithms

Land cover mapping provides basic information for advanced science such as ecological management, biodiversity conservation, forest planning and so on. In remote sensing research, the process of creating an accurate land cover map is an important subject. Recently, there has been growing research interest in the object-oriented image classification techniques. The object-oriented image classification consists of multidimensional features including object features and thus requires multi-dimensional image classification approaches. For example, a linear model such as the maximum likelihood method of pixel-based classification cannot characterize the patterns or relations of multi-dimensional data. In multi-dimensional image classification, data mining and ensemble learning have been shown to increase accuracy and flexibility. This study examined the use of the object-oriented image classification by the multiple machine learning algorithms for land cover mapping. We applied four classifiers: Classification and regression tree (CART), Decision tree with Boosting, Decision tree with Bagging, and Random Forest. The study area was Sado Island in Niigata Prefecture, Japan. Pan-sharpened SPOT/HRG imagery (June 2007) was used and classified into the following eight classes: broad-leaved deciduous forest, Japanese cedar, Japanese red pine, bamboo forest, paddy field, urban area, road, and bare land. We prepared four data sets with the object based features including textural information. The number of features is increased from data set I through IV. As the result, CART was unsuitable for multi-dimensional classification. Random Forest and Decision tree with Boosting showed high classification accuracies. Furthermore, in the data set with the limited features, Decision tree with Boosting was the accurate classifier. Finally, we propose two machine learning algorithms to every datasets. Random Forest is effective in the case of the multi-dimensional image classification such as data set II, III, and IV. Decision tree with Boosting is effective in the case of the image classification with the limited features such as data set I.