Forecasting land use in urban Himalayas: a remote sensing-guided machine learning approach

Abstract

Rapid urbanization in the Himalayan region due to advances in transportation, tourism and industry impacts natural resources causing its depletion. The problem gets aggravated in the urban regions located in the eco-fragile Himalayas. Srinagar being one of the largest urban hubs in the region with exceptionally high population growth rate, the complexities in monitoring land use and associated changes due to conventional methods further exacerbates these issues. Rapid and accurate mapping of land use land cover (LULC) is essentially required for effective green space management in urban landscapes. In this study an assessment of LULC using machine learning based classifiers was done. The present study assessed LULC using Sentinel-2 data through unsupervised K-means algorithm and supervised machine learning algorithms (Artificial Neural Network-ANN, Support Vector Machine- SVM, Random Forest-RF and Decision Tree-DT). Ground truth points collected through extensive field visits and high resolution Google Earth Pro were used for model generation/mapping (70%) and validation (30%). Map validation revealed that SVM (96.60%) had the highest overall accuracy followed by RF (95.86%), DT (95.33%), ANN (88.7%) and K-means (64.51%). F-Scores varied between classifiers on account of precision and recall for different classes. High values for F depicting performance of classification models were observed for all supervised classifiers except ANN which couldn’t effectively classify wastelands (F = 58.73%), SVM performed exceptionally well for agriculture and grassland (94.02%), Habitation (96.02%) and wasteland (96.42%). DT excelled in mapping vegetation (99.41%). Waterbody was classified accurately by all the classifiers (F = 100%) except ANN (99.73%). However, Snow and Agriculture Fallow were depicted well by ANN with F Score of 99.20% and 96.39% respectively.