A fully convolutional neural network model combined with a Hough transform to extract crop breeding field plots from UAV images

Abstract

High-throughput phenotypic analysis plays an increasingly important role in crop breeding. In such research, the breeder usually establishes hundreds to thousands of plots, with each plot having its independent genetic breeding sources. The breeding plot extraction of genetic sources is usually outlined manually on RGB UAV imagery, which is time-consuming and subject to human bias. Therefore, a rapid method to extract the breeding plot for each genetic source in high-throughput phenotypic analysis would be very significant. In this paper, we propose a transferable method for extracting breeding plots from UAV RGB imagery. We utilized the fully convolutional neural network model A-UNet with an attention gate. After obtaining binary raster data from deep learning, we introduced post-processing. Subsequently, the raster data after post-processing were converted to vector data to obtain geographical coordinates. Finally, the UAV imagery was masked by the vector data to obtain the extraction results for each plot. The results showed that A-UNet achieved accuracies of over 90 % in precision, recall, and F1 score. Post-processing resulted in a 93 % average IoU in breeding plot extraction in the main study area. The average IOU achieved over 86 % in different spatial resolutions (1.6 cm and 0.4 cm), plot sizes (1 m × 1.5 and 2 m × 5 m), and crop types (rice). In summary, this study developed a method for extracting breeding plots in high-throughput phenotype analysis, which would help to be used as a high-throughput screening technique for accelerating crop breeding.

Keywords

• Unmanned Aerial Vehicle (UAV)
• Deep learning
• Breeding experiment
• RGB images