TheLR531v1 – A deep learning multi-branch CNN architecture for day-night automatic segmentation of horticultural crops

Abstract

There is an increasing need for crop management practices to improve the day-night segmentation efficiency of crops grown in the open-field or in the greenhouse, in soil or in hydroponic environments, or both. We propose an asymmetric TheLR531v1 network with a multi-branch structure with powerful feature representation to recover plants, plant leaf or canopy (horticultural crops) from the background in real-time. The multi-branch structure of the network successfully extracts various image features from the input images and classifies them into leaf and background classes. We evaluated the network performance by using over 55,992 augmented images (tomato, eggplant, and lettuce) and obtained 94.55% training and 95.89% validation accuracy. The total parameters of TheLR531v1 are only 4.5 M, which is significantly lower than other state-of-the-art models. But realise that in addition, the proposed network achieves an average BF score of 89.00%, an IoU of 87.00%, and a GA of 96.00% to distinguish plants or plant canopy pixels from background pixels. Overall, TheLR531v1 can replace the laborious manual image segmentation tasks, analyze variability in stress conditions (temperature), and visualize stress related changes in plants, leaves or canopy (pixel area), which can improve crop management practices and yield.