Explainable deep convolutional neural networks for insect pest recognition. (15th October 2022)

• Record Type: Journal Article
• Authors: Coulibaly, Solemane; Kamsu-Foguem, Bernard; Kamissoko, Dantouma; Traore, Daouda
• Is Part Of: Journal of cleaner production. Volume 371(2022)
• Publication Date: 2022-10-15
• Subjects: Explainability of neural networks -- Transfer learning -- Deep learning -- Image classification -- Insect pest detection in crops; Factory and trade waste -- Management -- Periodicals; Manufactures -- Environmental aspects -- Periodicals; Déchets industriels -- Gestion -- Périodiques; Usines -- Aspect de l'environnement -- Périodiques; 628.5
• DOI: 10.1016/j.jclepro.2022.133638
• Languages: English
• ISSNs: 0959-6526

Abstract

Abstract: Fungal infestation of crops is critical to food security as it affects yield and quality of production. Indeed, one element responsible for this situation is insect pests. Early detection of pests based on parcel images is a real challenge in the context of precision agriculture. Nowadays, technical advances in deep neural networks have led to better results in all areas, including crop health management in agriculture. Despite these satisfactory results of deep neural networks in image classification tasks, one of the drawbacks is that it is difficult to decode what the neural networks have learned. The proposed method consists of identifying and locating insect pests in crops using a Convolutional Neural Network (CNN). The localization of insects from the input data is based on explainability methods. For this, explainability highlights the colors and shapes captured by the CNNs using visualization maps. This provides opportunities for human interaction with the learning system for validation of the results provided by the CNN models. In this study, we used over 75, 000 images for 102 different pest categories from the IP102 reference dataset. Various explainability methods are combined to formally interpret insect location. The degree of combination is quantified by the mutual information score. The obtained results allow a better interpretation of the reasoning performed by the deep learning system and identified an optimal number of feature extraction layers. Consequently, we simplified a CNN model by decreasing the number of network parameters by 58.90%. This facilitates their explanation in the field of plant science for the effective application in crop diagnosis. Highlights: Farmers have to cope with insect pests affecting agricultural crop production. Explainable methods to have clear and transparent rules for insect pest recognition. Transfer learning with data augmentation to improve Deep Neural Network training. Explanation methodology of convolutional neural networks in multi-classification. Join explainable methods with image alignment using mutual information measurements.