In this paper, a new image filtering method is proposed using nonlinear symbolic functions. In this way, feature extraction from images is achieved using a small number of filters and parameters. Nonlinear symbolic functions are determined by an optimization method and different approaches are proposed to increase the filtering speed.
Nonlinear functions also act as a regression layer and more efficient design is achieved by moving the image to nonlinear space. In sustainable agriculture, artificial intelligence has been recently utilized to pre-detect diseases and determine their severity to perform spraying more accurately in site-specific applications for human health and soil ingredients.
Therefore, a new dataset for detecting the disease root rot severity of Arugula (Eruca vesicaria) was created and labeled for the image classification such that there is a significant decrease in production yield every year due to root rot disease. The proposed CNN model with nonlinear symbolic function filtering was produced %86.92 accuracy for the classification of disease severity whereas conventional 4-layer CNN model and pre-trained large image classification models such as VGG16, VGG19, ResNet50, ResNet101, InceptionV3, DenseNet, EfficientNetV2S, MobileNet, MobileNetV2 and Xception have been fine-tuned and classification accuracy is provided between 57-92% without preprocessing of images.
Between the designed models, MobileNet model with less computational time and high accuracy can be a potential for future application in drone-based spraying systems. However, the proposed model with less number of parameters is also promising for future applications. The application results of image classification facilitate variable-rate spraying for sustainable agricultural production, ultimately protecting both human and soil health.
