Use of machine learning-based classification algorithms in the monitoring of Land Use and Land Cover practices in a hilly terrain.

Abstract

The current high rate of urbanization in developing countries and its consequences, like traffic congestion, slum development, scarcity of resources, and urban heat islands, raise a need for better Land Use Land Cover (LULC) classification mapping for improved planning. This study mainly deals with two objectives: 1) to explore the applicability of machine learning-based techniques, especially the Random forest (RF) algorithm and Support Vector Machine (SVM) algorithm as the potential classifiers for LULC mapping under different scenarios, and 2) to prepare a better LULC classification model for mountain terrain by using different indices with combination of spectral bands. Due to differences in topography, shadows, spectral confusion from overlapping spectral signatures of different land cover types, and a lack of access for ground verification, classification in mountainous terrain is difficult task compared to plain terrain classification. An enhanced LULC classification model has been designed using two popular machine learning (ML) classifier algorithms, SVM and RF, explicitly for mountainous terrains by taking into consideration of a study area of Gopeshwer town in the Chamoli district of Uttarakhand state, India. Online-based cloud platform Google Earth Engine (GEE) was used for overall processing. Four classification models were built using Sentinel 2B satellite imagery with 20m and 10m resolutions. Two of these models (Model 'i' based on RF algorithm and Model 'ii' based on SVM algorithm) were designed using spectral bands of visible and infrared wavelengths, and the other two (Model 'iii' based on RF algorithm and Model 'iv' based on SVM algorithm) with the addition of indices with spectral bands. The accuracy assessment was done using the confusion matrix based on the output results. Obtained result highlights that the overall accuracy for model 'i' and model 'ii' were 82% and 86% respectively, whereas these were 87.17% and 87.2% for model 'iii' and model 'iv' respectively. Finally, the study compared the performance of each model based on different accuracy metrics for better LULC mapping. It proposes an improved LULC classification model for mountainous terrains, which can contribute to better land management and planning in the study area.