Land use land cover representation through supervised machine learning methods: sensitivity on simulation of urban thunderstorms in the east coast of India

This study assesses the sensitivity of Land Use Land Cover (LULC) representation on the evolution of mesoscale convective systems over Bhubaneswar, a rapidly growing city (~ 77% growth in the last two decades) in India. In this study, three types of LULC maps have been prepared using supervised machine learning (ML) methods such as Classification and Regression Trees (CART), Naive Bayes (NB), and Support Vector Machine (SVM) on Google Earth Engine (GEE) platform using Landsat 8 for 2014. A high accuracy score (87%) and kappa coefficient (84%) revealed the best performance of CART in generating the LULC map. The Weather Research and Forecasting (WRF) model at 6 and 2 km horizontal resolution is forced with these LULC maps. Model results highlight that the CART experiment exhibits relatively less bias in 2 m relative humidity (~ – 10% to – 5%), 2 m temperature (~ 2.5 °C to ~ 0 °C), and 10 m wind speed (– 1 to ~ 1.8 m s−1) up to peak stage of the thunderstorms. The CART performs better with less rainfall error (~ – 16 mm) than CNTL (~ – 33 mm), NB (~ – 37 mm), and SVM (~ – 38 mm) and is supported by the quantitative statistical analysis, viz. less false alarm ratio, critical success index for different thresholds. LULC class-wise analysis indicates a higher variation in surface and lower atmospheric parameters over urban, shrubland, and cropland while less variation over barren, forest, and water. Thus, the study highlights the credibility of ML models in representing LULC information to input the high-resolution models.