Authors: Zihan Lin*, , Jiaguo Qi, Michigan State University
Topics: Remote Sensing, Land Use and Land Cover Change, Applied Geography
Keywords: Supervised classification, remote sensing, empirical data, samples
Session Type: Virtual Paper
Start / End Time: 11:10 AM / 12:25 PM
Room: Virtual 18
Presentation File: No File Uploaded
Supervised classification usually provides more accurate results in remote sensing image classification. The most fundamental distinction between supervised and unsupervised classification is the involvement of ground truth collections. However, ground-truthing information is not always accessible and available, leading to less optimal classification reliability for many cases. In this research, we established a new geospatial method to allow the auto-generation of relatively solid long-term field samples to facilitate supervised classification when in-field collections are not obtainable. We have successfully applied this method to the Bugun cropland in Kazakhstan and California, USA, to generate 10-m crop-specific land cover maps from 2016 to 2020, and 30-m state-wide land use land cover map from 1984 to 2020. Historical samplings for each site were acquired from two field trips to Bugun and historically surveyed landscape products from the California Department of Conservation. Both classification and accuracy assessment outcomes have confirmed the feasibility and reliability of this new approach. The highest supervised classification accuracy is 97.04% for the Bugun cropland. We also experimented with the number of ground truth samples and confirmed that more sampling could improve the approach performance. This work can be generalized to other studies seeking field data when such information is in short supply. The classification accuracy depends heavily on the number of samples collected for the first image to be classified.