Land-use and land-cover classification using Sentinel-2 data and machine-learning algorithms: operational method and its implementation for a mountainous area of Nepal

In the context of land-use and land-cover (LULC) classification, there is a lack of leverage of the recent increase of the ease of access to satellite imagery data, cloud computing platforms, and classification techniques. We present both the development of an operational method for LULC classification that considers these progresses and the implementation of this operational method for a mountainous area of Nepal. The operational method allows the comparison of three LULC maps, each derived with a different classification technique [classification and regression tree (CART), max entropy (MaxEnt), and random forest (RF)] applied to Sentinel-2 data on the Google Earth Engine platform. The results derived with the RF technique have the highest overall accuracy coefficient (92%). The probabilities that the RF technique produces a more accurate LULC map than the MaxEnt (95%) and CART (61%) techniques are based on Kappa statistics. Results of general linear models suggest that some LULC types have higher producer’s and user’s accuracies at a statistically significant level. The operational method can help the producers of LULC maps conduct future work on areas in developing countries, as such contributing to addressing various issues that involve land use.