Comparison of simulated hyperspectral HyspIRI and multispectral Landsat 8 and Sentinel-2 imagery for multi-seasonal, regional land-cover mapping

This study used simulated hyperspectral (HyspIRI) and multispectral (Landsat 8 OLI, Sentinel-2 MSI) satellite imagery to compare regional land-cover mapping capabilities (San Francisco Bay Area, California) within an analytical framework that included consistent reference data and classification rules. Imagery had the pixel resolution (30 m) and extent (30,000 km2) of a Landsat scene, with multi-seasonal (spring, summer, fall) acquisitions from year 2013. Primary study objectives were to assess differences in map accuracy related to Multiple Endmember Spectral Mixture Analysis (MESMA) and Random Forests (RF) classifiers, spectral resolution (hyperspectral vs. multispectral), and temporal resolution (multi-seasonal vs. summer). The RF classifier generally outperformed MESMA by 1.1 to 9.0% overall accuracy, with the exception of summer HyspIRI reflectance data. There were no clear patterns in accuracy when comparing HyspIRI and simulated multispectral reflectance bands with RF and MESMA classifiers. With summer data, HyspIRI had significantly higher accuracy for MESMA (+ 5.5 to + 8.7%) and significantly lower accuracy for RF (− 9.7 to − 16.4%). There were no significant differences in accuracy when using multi-seasonal HyspIRI and multispectral data with RF or MESMA (< 1.2% differences). There were highly significant improvements in overall accuracy (1.7 to 20.9%) with multi-season over summer-only images for all sensors, sample scales and classifiers. This result indicates that repeat image acquisitions from satellite sensors are important for land-cover classification, irrespective of sensor spectral resolution. A companion study (Clark & Kilham, 2016) that used RF with hyperspectral metrics derived from HyspIRI reflectance bands (from indices, derivatives, and absorption fitting) showed significant improvements in overall accuracy relative to map classifications in this study, which all used independent reflectance bands. These findings point to the need of additional land-cover mapping research with machine learning and hyperspectral data that span the spatial and temporal scales afforded by a satellite.