A maize canopy phenotypic traits detection method based on improved Mask2YOLO cascade deep learning network

Abstract

High-throughput detection of phenotypic traits of the maize canopy is significant for cultivar selection, cultivation, and management. Addressing the difficulty of rapidly recognizing maize canopy organs from field-acquired images, this study proposed a maize canopy phenotypic trait detection method based on an improved cascade deep learning network (DTMask2YOLO). The first stage network of DTMask2YOLO was composed of the DT-Mask2Former with DT-Decoder, and the second stage was the DSM-YOLOv8 embedded with dynamic snake convolution (DSConv) and a mixed local channel attention (MLCA) module. First, maize plants under natural conditions were collected as the dataset for this study. Second, the first-stage network was used to separate maize canopy regions from field-acquired natural backgrounds, followed by the second-stage network for canopy organ recognition. Lastly, to validate the application effect of the DTMask2YOLO model, a two-dimensional coordinate system was established and the phenotypic trait detection algorithms were used to detect phenotypic traits such as plant height, canopy width, tassel length, leaf length, leaf width, and stem-leaf angle. Analysis of computed results was conducted by establishing linear correlation models. The results indicated that the DT-Mask2Former canopy segmentation model achieved mIoU and mDice of 90.59% and 95.10%. The DSM-YOLOv8 canopy organ recognition model attained a mAP of 0.909 and an FPS of 65.9, which were 2% and 1.7% higher than the baseline model. The phenotypic calculation method proposed in this study also demonstrated high reliability, with respective R 2 values for plant height, canopy width, tassel length, leaf length, leaf width, and stem-leaf angle being 0.8659, 0.8965, 0.791, 0.7509, 0.7014, and 0.7789. This paper proposed a rapid maize canopy organ recognition model and a phenotypic calculation method, offering reliable technical support for non-destructive and quantitative detection of maize phenotypic traits and providing a technical foundation for the selection of superior cultivar.