Deep learning system for the detection of abiotic plant diseases in agriculture

Abstract

A computer system for the detection of abiotic plant diseases, comprising a deep learning model trained using transfer learning on preprocessed leaf images to classify environmental stress conditions.

Description

Application area of ​​the invention

[0001] The invention relates to agricultural diagnostic technologies that utilize deep learning systems for the precise detection, classification and monitoring of abiotic plant diseases caused by environmental and physiological stress. Background of the invention

[0002] Abiotic plant diseases, caused by non-biological stressors such as nutrient deficiencies, drought, salinity, chemical toxicity, and extreme temperatures, pose significant challenges to modern agriculture. Traditional diagnostic methods rely on manual inspection and expert interpretation of visible symptoms. These methods are often inconsistent, labor-intensive, and prone to error. With the increasing focus on precision agriculture and automation, there is a growing need for intelligent systems capable of detecting stress-related plant diseases with high accuracy. While deep learning has shown promising results in the detection of biotic diseases, existing solutions lack sufficient classification of abiotic stressors or are not scalable for real-time applications.Furthermore, many current systems neglect adaptive model updates or the centralized management necessary for large-scale agricultural applications. Therefore, an improved system is needed that leverages advanced deep learning architectures, efficient preprocessing techniques, and a dynamic client-server environment to accurately identify abiotic plant diseases and support sustainable agricultural decisions. Summary of the invention:

[0003] The invention provides a deep learning-based system for the automated detection and classification of abiotic plant diseases using transfer learning. The system employs pre-trained convolutional neural network architectures such as InceptionV4, VGG-16, DenseNet-121, ResNet-50, and MobileNet, which are fine-tuned for identifying symptoms caused by environmental stressors. Using a dataset of 17,860 leaf images, the system applies normalization, resizing, and augmentation techniques to improve the model's generalizability and reliability under varying conditions. Hyperparameters such as learning rate and dropout rate are optimized to increase prediction accuracy and reduce overfitting.

[0004] The invention also includes a client-server architecture that enables dynamic model updates, centralized training, and distributed diagnostic functions. Client devices capture leaf images, while the server manages model optimization and provides updated detection models. The system delivers diagnostic results in real time, thus minimizing manual evaluation. It supports precision agriculture through consistent, reliable, and scalable stress detection to improve plant health and productivity. Detailed description

[0005] The invention provides a deep learning system for detecting abiotic plant diseases using digital images of plant leaves. The system processes images from mobile devices, agricultural imaging equipment, drones, or field sensors, thus enabling flexible deployment. Each image is pre-processed, with pixel values ​​normalized, the size adjusted to a standardized resolution, and the data augmented to improve robustness and reduce sensitivity to variable environmental conditions.

[0006] The system's central diagnostic unit uses transfer learning to adapt widely used convolutional neural network (CNN) architectures such as InceptionV4, VGG-16, DenseNet-121, ResNet-50, and MobileNet for classifying abiotic plant stress. The system loads these pre-trained models with frozen base layers to preserve fundamental trait extraction properties, while fine-tuning the classification layers to identify specific types of abiotic stress. This approach reduces training time and improves performance with limited, specialized datasets.

[0007] The training dataset comprises 17,860 leaf images categorized according to various abiotic stress conditions. The dataset is divided into training and test sets to ensure consistent evaluation of model performance. The system optimizes model parameters by adjusting the learning rate and regularizing dropouts to prevent overfitting and maintain stable convergence.

[0008] Feature extraction is performed using folding layers that detect leaf patterns, color distortions, edge changes, and stress-induced structural variations. Classification layers output stress categories with confidence levels, enabling a reliable interpretation of plant health. Misclassified images are logged to iteratively improve model performance in subsequent training cycles.

[0009] The system uses performance metrics such as accuracy, precision, hit rate, and loss value for model validation. Based on the comparison results, the system selects the highest-performing architecture for real-time inference. Updated models are stored on a central server and form part of the client-server architecture.

[0010] Client devices communicate with the server to receive updated model parameters. This ensures consistent diagnostic accuracy across distributed agricultural sites. The server hosts the training environment and processes newly acquired images to refine the model. Through regular synchronization, client applications transmit images and receive classification results.

[0011] The invention supports edge computing functions by using resource-efficient models such as MobileNet for offline or low-connectivity environments. This enables local processing without constant network access and is therefore suitable for remote agricultural regions.

[0012] A diagnostic interface provides results such as the predicted stress type, its probability, and recommended countermeasures. The system can be integrated into existing farm management software to track trends in plant health and issue early warnings of increasing stress patterns.

[0013] Overall, the invention offers a comprehensive system that automates the detection of abiotic stress using artificial intelligence, reduces dependence on manual inspections, improves the reliability of the diagnosis, and supports long-term precision agriculture.

Claims

[1] A computer system for the detection of abiotic plant diseases, comprising a deep learning model trained using transfer learning on preprocessed leaf images to classify environmental stress conditions. [2] System according to claim 1, wherein the deep learning model comprises one or more pre-trained architectures of Convolutional Neural Networks (CNNs) selected from InceptionV4, VGG-16, DenseNet-121, ResNet-50 and MobileNet. [3] System according to claim 1, wherein the preprocessing of the leaf images comprises normalization, resizing and augmentation to improve the robustness of the model and the classification accuracy. [4] System according to any of the preceding claims, comprising a client-server architecture configured to update and deploy optimized deep learning models for real-time detection of abiotic plant diseases on distributed agricultural equipment.