Comparison of Machine Learning Classifiers for Land Cover Changes using Google Earth Engine

Increasing urbanization causes a variety of environmental changes, not only in regional but in the global scale, especially in developing countries. The developing countries like Lao PDR have very few data on environment and urban management and face many difficulties due to poor data management. Nowadays, using the integration of geographic information system (GIS) and remote sensing provides us with an effective result in determining the land use and land cover changes as well as providing valuable information needed for planning and researching. A new approach has been introduced for detecting and monitoring urban expansion by using the Big Data platform via the Google Earth Engine (GEE) cloud computing. Also, the machine learning algorithms have been recommended in this research, such as Random Forest (RF), Classification and Regression Tree (CART), Support Vector Machine (SVM) and Minimum Distance (MD). In this paper, we conduct GEE combining with multiple sources of satellite optical images time-series from three main satellites, Landsat 5 and Landsat 8 and Sentinel 2. For our dataset collection, we manually classify our dataset based on the pixel-based approach and object-based approach which were divided into training sample (70%) and testing sample (30%). Accuracy was assessed through metrics derived from a confusion/error matrix of the classifier. From the results, RF and CARF outperformed other classifiers and exhibited overall accuracy both training data and testing sample whereas SVM performed mediocrely and Minimum distance displayed least performance.