A machine learning-based classification of LANDSAT images to map land use and land cover of India

Abstract

The land use and land cover (LULC) maps are often required by planners and policymakers for effective planning and management interventions at the local, national, regional and global levels. Various attempts have been made to develop LULC maps using field-based surveys and by processing remotely sensed images. These maps can be developed using different tools and methodologies at different scales to achieve different levels of accuracy. With the advent of remote sensing technologies and its application in making LULC maps, attempts have been made to develop such maps with improved accuracy and consistency. The machine learning-based approaches have been attempted to develop LULC maps with varying levels of accuracy using different satellite images. Making LULC maps for a large region such as India covering a total area of ca. 3,287,469 km2 can be a cumbersome process using conventional approaches. Thus the map was developed using remotely sensed images using machine learning algorithm (Mnlogit) on LANDSAT images (2005, 2006, 2007 and 2016) for entire India region. We developed LULC maps of years 2005, 2006, 2007 and 2016 to test the consistency of classification using the trained Mnlogit model using field survey based signatures for corresponding years in respective images. We could achieve reasonably good accuracy varying in the range of 80–86% during all four years. A Kappa statistic - K (hat), was obtained in the range of 0.71–0.81 which indicates reasonably good accuracy. The study can be replicated for other regions using other available satellite remote sensing images to obtain LULC maps. In general, the suggested approach in this study will help planners to obtain LULC maps at different time intervals to study land-use change dynamics in a shorter time and cost-effective way.