Mapping land cover using a developed U-Net model with weighted cross entropy

Abstract

High-quality and high-productivity land cover data as a critical proxy are urgently needed for various communities. Although efforts in mapping projects had been made, various architectures to answer the challenges of Big Data of remote sensing are still needed. Therefore, this paper developed an improved U-Net model with weighted cross entropy (WCE) to map land covers. The accuracy was assessed by confusion matrix, and compared with the other two alternative cross-entropy loss functions. The comparisons highlighted an issue that unbalanced sample space of training dataset is a major cause of lower mapping accuracy. Also, the higher accuracy of U-Net with WCE implied its ability to handle the issue. This paper suggests an alternative solution for mapping land cover to address the challenges.

Keywords:

• Deep learning
• U-Net
• semantic segmentation
• remote sensing image
• land cover map

Acknowledgments

The authors would like to thank the anonymous reviewers for their tremendous time, helpful suggestions and insightful comments that greatly improved our manuscript.

Disclosure statement

No conflict of interest was reported by the authors.

Additional information

Funding

This research was funded by National Natural Science Foundation of China, grant number 41901344. Partial funding was provided by Natural Science Basic Research Program of Shaanxi Province of China, grant number 2020JQ-592 and Special Scientific Research Project of Education Department of Shaanxi Province, grant number 19JK0837.