Simultaneous extracting area and quantity of agricultural greenhouses in large scale with deep learning method and high-resolution remote sensing images

Qingpeng Wang; Wei Chen; Hongzhao Tang; Xubin Pan; Haimeng Zhao; Bin Yang; Honggeng Zhang; Wenzhu Gu

doi:10.1016/j.scitotenv.2023.162229

Simultaneous extracting area and quantity of agricultural greenhouses in large scale with deep learning method and high-resolution remote sensing images

Sci Total Environ. 2023 May 10:872:162229. doi: 10.1016/j.scitotenv.2023.162229. Epub 2023 Feb 13.

Authors

Qingpeng Wang¹, Wei Chen², Hongzhao Tang³, Xubin Pan⁴, Haimeng Zhao⁵, Bin Yang⁶, Honggeng Zhang¹, Wenzhu Gu¹

Affiliations

¹ College of Geoscience and Surveying Engineering, China University of Mining & Technology, Beijing 100083, China.
² College of Geoscience and Surveying Engineering, China University of Mining & Technology, Beijing 100083, China. Electronic address: chenw@cumtb.edu.cn.
³ Land Satellite Remote Sensing Application Center, Ministry of Natural Resources, Beijing 100048, China.
⁴ Institute of Plant Inspection and Quarantine, Chinese Academy of Inspection and Quarantine, Beijing 100176, China.
⁵ Guangxi Colleges and Universities Key Laboratory of Unmanned Aerial Vehicle (uav) Remote Sensing, Guilin University of Aerospace Technology, Guilin 541004, China.
⁶ College of Electrical and Information Engineering, Hunan University, Changsha 410082, China.

PMID: 36791864
DOI: 10.1016/j.scitotenv.2023.162229

Abstract

Greenhouses are an important part of modern facility-based agriculture. While creating well-being for human society, greenhouses also bring negative impacts such as air pollution, soil pollution, and water pollution. Therefore, it is of great significance to obtain information such as the area and quantity of greenhouses. It is still a challenging task to find a low-cost, high-efficiency, and easy-to-use method for the dual extraction of greenhouse area and quantity on a large scale. In this study, relatively easy-to-obtain high-resolution Google Earth remote sensing images are used as the experimental data source, and an area and quantity simultaneous extraction framework (AQSEF) is constructed to extract both the area and quantity of greenhouses. The AQSEF uses UNet and YOLO v5 series networks as core operators to complete model training and prediction, and main components such as SWP, OSW&NMS and GCA complete data postprocessing. To evaluate the feasibility of our method, we take Beijing, China, as the research area and select multiple accuracy evaluation indicators in the two branches for accuracy verification. The results show that the mIoU, OA, Kappa, Recall and Precision with the best performance model in the area extraction branch can reach 0.931, 0.987, 0.867, 0.91 and 0.914, respectively. Additionally, the Recall, Precision, AP@0.5 and mAP@0.5: 0.95 values of the best performance model are 0.781, 0.891, 0.812 and 0.509, respectively, in the extraction of the quantity of greenhouses. Finally, in Beijing, the area covered by greenhouses is approximately 85.443 km², and the quantity of greenhouses is approximately 155,464. With the proposed method, the time consumed for area extraction and quantity extraction is 6.73 h and 12.97 h, respectively. The experimental results show that AQSEF helps to overcome the spatiotemporal diversity of greenhouses and quickly and accurately map a high-spatial-resolution greenhouse distribution product within the research area.

Keywords: Agricultural greenhouses; Area and quantity; Automatic extraction; Convolutional neural network; Google earth remote sensing images.