A Multiclass Deep Learning Approach for LULC Classification of Multispectral Satellite Images

Abstract

In general, a visual interpretation technique is adopted for mapping of Land Use / Land Cover (LULC) using temporal satellite data. Although highly accurate, the process is tedious, time consuming and requires a significant amount of domain knowledge. This limitation introduces a scope for partial automation to reduce manual effort involved in interpretation, while maintaining baseline accuracy. The research explores a novel multi-class training approach using a Deep Learning (DL) model to generate major LULC classes. Five spectral bands, namely Blue, Green, Red, Near-Infrared (NIR) and Short wave Infrared (SWIR) from the Sentinel-2A satellite, covering Mandya, Karnataka, India was used to train the model. An existing LULC map of the region was used as an input for automatically generating labeled training samples and a modified SegNet was implemented for classification. Four major LULC classes of interest - water bodies, forest lands, croplands, built-up were classified with an average F1 score of 0.84. The trained model applied to other regions has shown encouraging results which makes this an effective method to explore the generation of LULC maps.