Mapping land cover with hyperspectral and multispectral satellites using machine learning and Spectral Mixture Analysis

Abstract

The goal of this study was compare hyperspectral and multispectral imagery for mapping broad land-cover classes at the spatial scale of a satellite image. The study area was the San Francisco Bay Area and was roughly the size of a Landsat scene (30,000 km2). The Random Forests machine learning and Multiple-Endmember Spectral Mixture Analysis (MESMA) classifiers were compared to predictor variables composed of simulated HyspIRI hyperspectral images, simulated Landsat 8 and Sentinel-2 multispectral images, and real Landsat 8 images. The Random Forests machine learning classifier consistently outperformed MESMA and there were significant improvements in overall accuracy with multi-temporal (spring, summer, fall) over summer-only images for all sensors tested. Hyperspectral reflectance data had no difference to less accuracy relative to comparable multispectral datasets. However, HyspIRI hyperspectral metrics that targeted key spectral features, related to chemical and structural properties, yielded significantly improved accuracy over both real and simulated multispectral datasets.