Tipburn disorder detection in strawberry leaves using convolutional neural networks and particle swarm optimization

Tipburn is a disorder that is caused by calcium deficiency in plants and may lead to decrements in crop yield. Therefore, it is important to detect tipburn for an appropriate treatment process. In this work, a sequential convolutional neural network (CNN) architecture was developed for tipburn detection in images of strawberry leaves. The parameters of the CNN architecture as well as the dropout and learning rates were determined through a two-stage search operation based on particle swarm optimization (PSO) algorithm. The resulting model, PSO-CNN, contains five convolutional layers and three fully-connected layers with varying number of filters and hidden units. The model development was performed using an original dataset of strawberry leaf images taken from the field under realistic conditions and has been made publicly available with this study. The performance of the PSO-CNN was compared with the performance of eight different benchmark CNN models, namely, VGG16, VGG19, MobileNetV2, EfficientNet, ResNetV2, NasNetMobile, InceptionV3 and InceptionResNetV2. According to the results obtained by ten independent re-runs of the classification task, PSO-CNN achieved the best average performance by 0.9895, 0.9863, and 0.9936 for accuracy, sensitivity, and specificity values, respectively. In addition, the number of parameters of the PSO-CNN model is smaller than those in the benchmark models. This means that PSO-CNN model requires relatively less amount of computation for an efficient model training and performing prediction on the test images. Finally, further experiments were performed on a multi-class problem to demonstrate the effectiveness of the PSO-CNN for tipburn detection.