LULC mapping accuracy enhancement through multispectral UAV imagery with nDSM integration

Abstract

The requirement of land use and land cover (LULC) maps as a base in large variety of applications make necessary to improve low-cost and high accuracy production methods. Unmanned Aerial Vehicles (UAVs) present cost-effective alternatives for generating LULC maps when compared to traditional methods and stand out with advanced multispectral sensing technologies. This study aims to assess the multi-class LULC mapping performance of multispectral UAVs and enhance it through the integration of auxiliary data sources, in 11-classes study area. Specifically, high-accuracy normalized digital surface model (nDSM) was generated and incorporated into the classification process to enhance overall mapping accuracy. In addition, three different datasets were created with the various combinations of 68 features consisting of texture, spectral and geometric features of the segments. Object-based classification was performed with the Random Forest (RF) machine learning algorithm for all datasets, and dataset 3 (D3), consists of spectral bands + indexes + texture + geometry + nDSM, exhibited the most successful performance with an overall accuracy of 94.16%. The results clearly demonstrated that MS UAV data has high performance in LULC mapping, and NDSM increased the classification accuracy by 5%.