Accurate mapping of rapeseed fields in the initial flowering stage using Sentinel-2 satellite images and convolutional neural networks

Abstract

In high-intensity farming, swiftly and accurately adjusting the proportion of artificially reared pollinators according to the flowering phenology of crops is vital. This adjustment is essential for sustaining crop yields without disrupting the ecological niche of native pollinators. Although advancements in remote sensing provide practical means to gather data on crop coverage and extent, identifying insect-pollinated crops during their initial flowering to guide the deployment of managed pollinators is an underexamined research area. Here, we tested the capability of utilizing Sentinel-2 satellite images combined with a deep learning model to map crop fields during the initial flowering stage of insect-pollinated crops in Zhaosu County, Xinjiang, China. Specifically, we examined rapeseed fields by employing 12 neural network designs to identify images of the initial flowering stage. Different network combinations of three convolutional neural network (CNN) models (U-Net, PSPNet, and DeepLab V3) and four different backbone networks (ResNet-18, ResNet-34, ResNet-50, and ResNet-101) were explored to determine the most effective model for detecting rapeseed fields at the initial flowering stage. A comparison was conducted with Sentinel-2 images obtained at the peak stage of rapeseed flowering. Our results suggest that the use of a deep learning model in combination with Sentinel-2 image data can successfully identify rapeseed fields at the initial flowering stage, thereby offering preliminary insights for the strategic introduction of managed pollinators. The PSPNet model emerged as the superior choice for the identification of rapeseed fields, exhibiting high accuracy in both detection and boundary recognition, with F1 scores of 88.17 % and an intersection over union (IoU) of 53.11 % during initial flowering and F1 scores of 93.33 % with an IoU of 53.49 % during peak flowering. The planting area of rapeseed fields detected by the model in Zhaosu County was 122.16 km2, distributed within an area of 1068 km2. These findings can provide foundational data for informed decisions regarding artificial pollinator supplementation, helping to sustain the health and stability of agricultural ecosystems.

Keywords

• Agroecosystems
• Sentinel-2
• Rapeseed field
• Semantic segmentation
• Remote sensing
• Pollinator