Family-based plant disease characterization using deep neural networks

Multimedia Tools and Applications https://doi.org/10.1007/s11042-025-20835-w Family-based plant disease characterization using deep neural networks Sivasubramaniam Janarthan1 · Selvarajah Thuseethan2 Sutharshan Rajasegarar1 · John Yearwood1 · Received: 5 April 2024 / Revised: 20 March 2025 / Accepted: 8 April 2025 © The Author(s) 2025 Abstract Over the years, researchers have applied various deep learning techniques to automatically recognise plant diseases from both raster and spectral images. The primary focus of the existing studies is developing individual species-specific or disease-specific models, where the former recognises diseases of single crop type and the latter recognises single diseases of single or multiple crop types. Building one global model to recognise diseases of multiple crops has also been widely explored, where a class is treated as a crop-disease combination. While training individual species-specific or disease-specific deep models is labour-intensive, embracing a vast number of crop species and inherent diseases present on this planet makes the model cumbersome. In order to address this problem, a more intuitive and feasible familybased plant disease characterisation approach with botanical reasoning is proposed in this study. This approach demonstrates the feasibility of six state-of-the-art deep neural networks through a set of extensive experiments incorporating six key strategies. The results on a newly built family-based plant disease dataset confirm that the proposed novel approach is convincing to be applied in a plant family-based disease recognition problem. Further, this study creates future opportunities for more intuitive plant disease data collection and benchmark classification model development. Keywords Plant disease recognition · Plant family · Deep learning · Convolutional neural network · Transfer learning B Selvarajah Thuseethan Sivasubramaniam Janarthan Sutharshan Rajasegarar John Yearwood 1 School of Information Technology, Deakin University, Geelong 3125, VIC, Australia 2 Faculty of Science and Technology, Charles Darwin University, Casuarina 0810, NT, Australia 123 Multimedia Tools and Applications 1 Introduction The world population is expected to reach 9.6 billion by 2050, which demands a 70% increase in the world food supply. Plant diseases have become the major threat to agriculture sustainability, and continuous food supply together with other dominating threats, such as plant pests, climate change and recently loomed environmental pollution [1, 2]. For instance, even the less harmful target spot disease caused 10-42% of yield losses in America despite the fact that improved agricultural practices involving cultivar selection, crop succession, and intensive use of fungicides are in place [3]. Sudden yield losses cause not only financial difficulties but also put extra pressure on farmers and ag-allied professionals. Thus, it is necessary to provide automated solutions to accurately identify plant diseases resulting in increased productivity of farmers and allied professionals. Researchers in the past explored various computer vision-based techniques for accurate plant disease recognition from both raster and spectral images. Deep learning-based techniques, such as convolutional neural networks (CNN), have achieved tremendous success in image classification tasks without requiring any explicit feature engineering. A systematically designed deep neural network trained with a task-based objective function using a gradient descent optimization algorithm results in a network that can automatically extract important features and classify them based on the trained objective. The first deep CNN with five convolutional layers and a fully connected layer, namely AlexNet [4], outperformed all the existing classical approaches on the ImageNet [5] dataset. Subsequently, many different architectures have been proposed for the purpose of improving accuracy [6–8] and enhancing computational efficiency [9–12]. As a pitfall, deep networks often require large data for parameter learning and generalization. This challenge is overcome by a technique called transfer learning, where the feature extraction ability of the pre-trained networks is fine-tuned to a new task with small training data [13]. Thus, several studies in the recent past have incorporated deep learning techniques to develop plant disease recognition systems from images [14–16]. The deep learning-based plant disease recognition models are being built in two different approaches: individual models and global models. The individual model approach has primarily focused on the species-specific or disease-specific scheme. In species-specific models, multiple disease categories of a particular crop type are considered for classification. For example, individual models are proposed for the recognition of diseases belonging to crop types, rice [17], and potato [18]. In [19], the deep CNN model trained to identify ten disease types of the rice plant achieved 95.48% accuracy, which is much higher than the existing approaches. In a very recent study, solely the tomato diseases are taken into consideration for improved classification accuracy [20]. The key objective of disease-specific models is to effectively recognize one disease type that can be observed in single or multiple crops. For instance, an improved deep network is used to recognize canker disease from citrus plant [21]. In another work, a MobileNetv2-based YOLOv3 model is proposed for early identification of tomato grey leaf spot disease [22]. Developing an individual model of both kinds is not only a time-consuming task but also requires a large storage capacity. When it comes to the global model approach, one model is trained to classify multiple diseases of multiple crop types where each crop-disease combination is treated as a class [23, 24]. In studies that proposed global models, the PlantVillage [25] is the most widely used dataset as it consists of multiple diseases of a range of crop types. Precisely, this dataset contains 54,309 images for 38 different diseases and healthy classes of 14 different crop types. By treating the crop-disease combinations of the PlantVillage dataset as individual classes, in [26], the GoogleNet [27] achieved 99.35% accuracy. In another work, the same 123 Multimedia Tools and Applications approach is applied to another dataset comprising 87,848 diseased samples for 58 different diseases and healthy classes of 25 different crop types [28]. On this dataset, the VGG [29] network obtained the highest classification accuracy of 99.53%. Subsequently, there are numerous studies have been proposed using the global model strategy to handle the plant disease recognition task. When the vast number of crops and their inherent diseases are considered, it is challenging to provide a workable global model with an expected level of performance. The performance of a global mo (...truncated)