A novel approach: Coupling prior knowledge and deep learning methods for large-scale plastic greenhouse extraction using Sentinel-1/2 data

Abstract

Plastic greenhouses (PGs) are integral to modern agricultural practices, enhancing crop yields but also raising environmental concerns. Consequently, comprehending their widespread distribution is essential. Although deep learning has been extensively used for land use/cover classification and extraction with satellite data, the large number of labels limits its application due to the time-consuming and labor-intensive nature of manual labeling. This study introduces a novel approach coupling Prior Knowledge and Deep Learning methods for PG Mapping (PGMPK+DL) using Sentinel-1/2 data. We use an automatic labeling strategy guided by prior knowledge from Sentinel-2 optical data to construct PG labels in six small regions. Moreover, to overcome the cloud contamination issue of optical data, the potential of Sentinel-1 time-series SAR data for PG extraction is analyzed. Deep learning methods are further utilized to capture more abstract and generalized temporal and spatial features from time-series radar data to accommodate complex scenes for large-scale PG extraction. The U-Net model emerges as superior from rigorous comparisons among FCN, SegNet, U-Net, DeepLabV3 + and U-Net3 + deep learning models. Finally, the U-Net model harnessed prior knowledge-based PG labels and Sentinel-1 time-series SAR data to generate a precise map depicting PG distribution across Shandong province, China. Remarkably, it accurately identifies approximately 238,000 ha of PG areas. This PGMPK+DL approach presents a groundbreaking solution for label construction, enabling the achievement of large-scale PG extraction. Beyond enhancing PG extraction, it also holds broader implications for advancing deep learning applications within remote sensing.

Keywords

• Plastic greenhouse
• Prior knowledge
• Time-series
• Sentinel-1
• Semantic segmentation