Comparison of Simulated Multispectral Reflectance among Four Sensors in Land Cover Classification

• Published in: Remote sensing (Basel, Switzerland) Vol. 15; no. 9; p. 2373
• Main Authors: Chen, Feng, Zhang, Wenhao, Song, Yuejun, Liu, Lin, Wang, Chenxing
• Format: Journal Article
• Language: English
• Published: Basel MDPI AG 30.04.2023
• Subjects: Accuracy; Case studies; Channels; Classification; Classifiers; Comparative analysis; Data collection; Earth resources technology satellites; Gaofen; Image classification; Land cover; Landsat; Landsat 8; machine learning; multispectral; Reflectance; Remote sensing; Response functions; Satellites; Sensors; Sentinel-2; Short wave radiation; Simulation; Spatial discrimination; Spatial resolution; Spectral sensitivity; Support vector machines; Surface dynamics; WFV; Indiana; China
• ISSN: 2072-4292; 2072-4292
• DOI: 10.3390/rs15092373

Multispectral images accessible free of charge have increased significantly from the acquisitions by the wide-field-of-view (WFV) sensors onboard Gaofen-1/-6 (GF-1/-6), the Operational Land Imager (OLI) onboard Landsat 8 (L8), and the Multi-Spectral Instrument (MSI) onboard Sentinel-2 (S2). These images with medium spatial resolutions are beneficial for land-cover mapping to monitor local to global surface dynamics. Comparative analyses of the four sensors in classification were made under different scenarios with five classifiers, mainly based on the simulated multispectral reflectance from well-processed hyperspectral data. With channel reflectance, differences in classification between the L8 OLI and the S2 MSI were generally dependent on the classifier considered, although the two sensors performed similarly. Meanwhile, without channels over the shortwave infrared region, the GF-1/-6 WFVs showed inferior performances. With channel reflectance, the support vector machine (SVM) with Gaussian kernel generally outperformed other classifiers. With the SVM, on average, the GF-1/-6 WFVs and the L8 OLI had great increases (more than 15%) in overall accuracy relative to using the maximum likelihood classifier (MLC), whereas the overall accuracy improvement was about 13% for the S2 MSI. Both SVM and random forest (RF) had greater overall accuracy, which partially solved the problems of imperfect channel settings. However, under the scenario with a small number of training samples, for the GF-1/-6 WFVs, the MLC showed approximate or even better performance compared to RF. Since several factors possibly influence a classifier’s performance, attention should be paid to a comparison and selection of methods. These findings were based on the simulated multispectral reflectance with focusing on spectral channel (i.e., number of channels, spectral range of the channel, and spectral response function), whereas spatial resolution and radiometric quantization were not considered. Furthermore, a limitation of this paper was largely associated with the limited spatial coverage. More case studies should be carried out with real images over areas with different geographical and environmental backgrounds. To improve the comparability in classification among different sensors, further investigations are definitely required.