Enhancing Green Fraction Estimation in Rice and Wheat Crops: A Self-Supervised Deep Learning Semantic Segmentation Approach

Abstract

The green fraction (GF), which is the fraction of green vegetation in a given viewing direction, is closely related to the light interception ability of the crop canopy. Monitoring the dynamics of GF is therefore of great interest for breeders to identify genotypes with high radiation use efficiency. The accuracy of GF estimation depends heavily on the quality of the segmentation dataset and the accuracy of the image segmentation method. To enhance segmentation accuracy while reducing annotation costs, we developed a self-supervised strategy for deep learning semantic segmentation of rice and wheat field images with very contrasting field backgrounds. First, the Digital Plant Phenotyping Platform was used to generate large, perfectly labeled simulated field images for wheat and rice crops, considering diverse canopy structures and a wide range of environmental conditions (sim dataset). We then used the domain adaptation model cycle-consistent generative adversarial network (CycleGAN) to bridge the reality gap between the simulated and real images (real dataset), producing simulation-to-reality images (sim2real dataset). Finally, 3 different semantic segmentation models (U-Net, DeepLabV3+, and SegFormer) were trained using 3 datasets (real, sim, and sim2real datasets). The performance of the 9 training strategies was assessed using real images captured from various sites. The results showed that SegFormer trained using the sim2real dataset achieved the best segmentation performance for both rice and wheat crops (rice: Accuracy = 0.940, F1-score = 0.937; wheat: Accuracy = 0.952, F1-score = 0.935). Likewise, favorable GF estimation results were obtained using the above strategy (rice: R2 = 0.967, RMSE = 0.048; wheat: R2 = 0.984, RMSE = 0.028). Compared with SegFormer trained using a real dataset, the optimal strategy demonstrated greater superiority for wheat images than for rice images. This discrepancy can be partially attributed to the differences in the backgrounds of the rice and wheat fields. The uncertainty analysis indicated that our strategy could be disrupted by the inhomogeneity of pixel brightness and the presence of senescent elements in the images. In summary, our self-supervised strategy addresses the issues of high cost and uncertain annotation accuracy during dataset creation, ultimately enhancing GF estimation accuracy for rice and wheat field images. The best weights we trained in wheat and rice are available: https://github.com/PheniX-Lab/sim2real-seg.