Application of convolutional neural network for monitoring the condition of strawberries

Abstract

   The article proposes a method for improving the accuracy of diagnosing calcium deficiency in strawberry plants, suggests the use of machine learning algorithms, such as convolutional neural networks (CNN), which makes it possible to train a model on a data set for qualitative detection of signs of calcium deficiency in the leaves. A dataset of images of healthy leaves and leaves with signs of calcium deficiency was collected, the method of artificially increasing the volume of the training sample (image augmentation) was applied, by horizontal and vertical reflection of objects in the images, rotation by a given angle and random addition of «noise». To train a convolutional neural network, an algorithm for obtaining RGB images using a robotic platform is proposed. A modern model of the YOLOv7 neural network was used as a means of detecting the signs of calcium deficiency in the leaves of strawberry in the images. The configuration of the YOLOv7 machine learning algorithm for recognizing areas of damage to strawberry leaves caused by calcium deficiency has been determined. To train the YOLOv7 model, the Transfer learning method was used. To assess the quality of the object recognition algorithms, the metric mAP (mean average precision) – 0.454 was used, the metric F1-score (F-measure) – 0.53, the average absolute error (Mean Absolute Percentage Error, MAPE) of the analyzed model of the YOLOv7 neural network was calculated. The analysis of the results showed that the YOLOv7 model recognized the «Calciuemdeficiency» class, with a MAPE index equal to 7.52 %. The analysis of the research results showed that timely monitoring of the condition of garden strawberries on an industrial plantation carried out using a wheeled robotic platform with the use of the YOLOv7 convolutional neural network for processing the data obtained will allow to determine calcium deficiency in the leaves of strawberry plants with high accuracy up to 94.43 % at the early stages of pathology development.