A Comparative Analysis of Plant Canopy Detection Performance in a Variable-Rate Spraying System Using Deep Learning Models

The variable rate spraying system was developed by recognizing and detecting the plant canopy characteristics with the help of deep learning models employed on different hardware models. The developed spraying system's detection performance was analyzed using deep learning models, viz. Single Shot Detector version 2 (SSDv2), You Only Look Once version 5 (YOLOv5), and EfficientDet deployed on three different hardware models, viz. Raspberry Pi 4 processor, V831 single-board computer, and an OAK-D depth camera at three speeds of operation (2 km/h, 3 km/h, and 4 km/h) with three plant-to-plant spacings (1.5 m, 3.0 m, and 4.5 m). Among the hardware models selected for the study, the Raspberry Pi 4 worked well with all the deep learning models at 2 km/h speed of operation only, whereas the V831 does not support YOLOv5 and EfficientDet; it supports only SSDv2 at all speeds of operation, and OAK-D outperformed the V831 and Raspberry Pi 4 over a wide range of speeds and plant-to-plant spacing in terms of overall accuracy, precision, recall, F1 score, FPS, and response time. The highest average overall accuracy of detection was found 99% with OAK-D using SSDv2 at 2 km/h speed of operation and 1.5 m plant-to-plant spacing. Experimental results recommended that the sprayer with the OAK-D depth camera was suitable to work with all the speeds of operation, with all the plant-to-plant spacings, and with all the deep learning models selected for the study in comparison with the Raspberry Pi 4 and V831.