Crop classification for UAV visible imagery using deep semantic segmentation methods

Abstract

Unmanned aerial vehicle (UAV) has become a mainstream data collection platform in precision agriculture. For more accessible UAV–visible imagery, the high spatial resolution brings the rich geometric texture features triggered large differences in same crop image's features. We proposed an encoder–decoder's fully convolutional neural network combined with a visible band difference vegetation index (VDVI) to perform deep semantic segmentation of crop image features. This model ensures the accuracy and the generalization ability, while reducing parameters and the operation cost. A case study of crop classification was conducted in Chengdu, China, where classified four crops, namely, maize, rice, balsam pear, and Loropetalum chinese, it was shown more effective results. In addition, this study explores a fine crop classification method based on visible light features, which is feasible with low equipment cost, and has a prospect of application in crop survey based on UAV low altitude remote sensing.

Keywords:

• Crop
• UAV visible images
• deep semantic segmentation
• encoder–decoder

Disclosure statement

The authors declare no conflict of interest.

Additional information

Funding

This research was supported by the National Natural Science Foundation of China (41701499); RWDL2021-ZD003; ZHJJ2021-ZD001; the National Key Research and Development Program of China under Grant 2017YFB0503601; the National Natural Science Foundation of China under Grant 41671448. The Second Tibetan Plateau Scientific Expedition and Research Program (STEP), China (No.2019QZKK0301); the program of Census of Forest Germplasm Resources in Chenghua District, Chengdu (80303-AHL038); the Second National Survey of Key Protected Wild Plant Resources–Special Survey of Orchidaceae in Sichuan Province(80303-AZZ003); Shanghai Engineering Research Center Marine Renewable Energy (Grant No. 19DZ2254800).