Large-scale automatic extraction of agricultural greenhouses based on high-resolution remote sensing and deep learning technologies.

Abstract

Widely used agricultural greenhouses are critical in the development of facility agriculture because of not only their huge capacity in food and vegetable supplies, but also their environmental and climatic effects. Therefore, it is important to obtain the spatial distribution of agricultural greenhouses for agricultural production, policy making, and even environmental protection. Remote sensing technologies have been widely used in greenhouse extraction mainly in small or local regions, while large-scale and high-resolution (~ 1-m) greenhouse extraction is still lacking. In this study, agricultural greenhouses in an important agricultural province (Shandong, China) are extracted by the combination of high-resolution remote sensing images from Google Earth and deep learning algorithm with high accuracy (94.04% for mean intersection over union over test set). The results demonstrated that the agricultural greenhouses cover an area of 1755.3 km2, accounting for 1.11% of the total province and 2.31% of total cultivated land. The spatial density map of agricultural greenhouses also suggested that the facility agriculture in Shandong has obviously regional aggregation characteristics, which is vulnerable in both environment and economy. The results of this study are useful and meaningful for future agriculture planning and environmental management.