Authors: Yanyan Liu*, University of West Florida, Zhiyong Hu, University of West Florida
Topics: Geographic Information Science and Systems, Environment
Keywords: classification,land use/cover
Session Type: Poster
Start / End Time: 9:55 AM / 11:35 AM
Room: Lincoln 2, Marriott, Exhibition Level
Presentation File: Download
Urban land use/cover classification from remotely sensed imagery is one of the major applications of remote sensing and poses a challenge in resolving land use classes due to inhomogeneous urban landscape and mixed spectral responses. With the amazing innovative advancements in artificial intelligence (AI), machine learning has gained increased popularity in exploiting remote sensing imagery The paper represents an application of a random forest-based supervised classification algorithm in land use/cover classification (LULCC) using Sentinel-2B MSI image data for Pensacola area, Florida. Training areas were selected using visual interpretation and a modified USGS land use/cover level 2 classification scheme. Land use/cover types include water, developed (low and high intensities), barren, forest, grassland, cultivated, sand beach, and wetlands. Input data include 10-band 10 m or 20 m visible, NIR and SWIR SMI band images, derived NDVI and textural layers. The procedure created models and generated land use class predictions using an adapted Breiman’s random forest algorithm. A series of model diagnostics were output and used to tune the model. The accuracy of the resulting LULCC raster map was assessed against temporally matched high-resolution Google Earth Pro historical imagery. A traditional maximum likelihood classifier was also used to compare accuracies. It was found that random forest based classifier achieved significantly higher overall accuracy and kappa value.