Acre-Scale Grape Bunch Detection and Predict Grape Harvest Using YOLO Deep Learning Network

To provide the harvesting weight of grapes in real time to the farmer in the form of acres or square feet of land, the proposed system provides an estimate of grape harvest in the form of kilograms. Some of the challenges in the field where crop harvest may vary due to differences in soil, the size of grapes, maturate of the grapes, color, disease, etc. The proposed system uses YoloV5, a real-time object detection system, to detect the grape bunches, and the total bunch count is based on several images collected from the forms. The total area of acres is divided into bins, which are in turn divided into sectors. The grape data are collected from the Dodballapur district; Karnataka, which consist of 4000 grape images and manual annotation was performed on each image to classify the grape bunch and leaves. Pre-trained YoloV5 is used to detect the grapes by extracting visual features from the training images. The visual images are fetched and used to build a detection network using YoloV5. The research presents three-object detection-pre-trained YOLO models such as YoloV3, YoloV4, and YoloV5, which perform detection of grape bunches. Three models are trained on grape bunches and were compared with precision, F1 Score, and recognition recall value, achieving good results with YoloV5 along with counting grapes, which helps in knowing the harvesting status of the bunch 98%. The recall and F1 Scores are 90.2, 91.25, 95.63 and 92.21, 92.68, 93.21 for YoloV3, YoloV4, YoloV5, respectively. The network also calculates the total bunches in each image, and multiplied with average gram of a bunch and provides the final kilograms expected from each acre for the set of images collected for the particular acre.