Description:Semi-Supervised Transfer Learning Framework with Adaptive CNN Fine-Tuning for Resource-Constrained Plant Disease Diagnosis from Leaf Imagery

2.Problem Statement
Plant diseases are a serious menace to the world agricultural output, food security and livelihoods of millions of farmers. The need to detect and classify these diseases early before crop protection is of great importance especially when these diseases exhibit on the surface of leaves (e.g. fungal, bacterial, or viral diseases). Conventional ways of diagnosis like manual analysis (by agronomists or pathologists) are costly, consume a lot of time, are time-consuming and are susceptible to human error. Additionally, inappropriate diagnosis by specialists in rural and under-developed agricultural areas is usually in-accessible.
The recent progress in computer vision and deep learning made it possible to design the automatic system of plant disease detection which relies on Convolutional Neural Networks (CNNs). Even under controlled conditions, these models have proven to be promising but on deployment in real environments, it has been noted that there are some severe challenges to it:
Lack of Data/High Cost of Labelling Data: The majority of deep learning models need plenty of high-quality labelled image data to train them. Available or limited are, however, the annotated datasets of region specific crops and rare disease type. Such datasets are not quick and easy to manually annotate, needing special expertise on the subject material.
Unreliable Generalization and Domain Shift: Pre-trained models cannot be used reliably when exposing the model to new field conditions because of changes in lighting, background noise, leaf orientation and camera quality. Such a shift in domain leads to low precision and constrains the possibility of use in a variety of settings.
Deployment resource limitation: The majority of contemporary CNN-based models are very computationally demanding and cannot run on a low-power device, like a smartphone, drone, or edge-Internet of Things (IoT) system in a remote agricultural area.
Absence of explicability: The existing models usually act like a black box and provide insufficient knowledge about the decision-making process. This non-explainability is a barrier to adoption of such systems by non-technical users such as farmers and introduces a threshold of trust, accountability and usability to suggest decision support.
To fill these gaps, it is imperative to develop an adaptive, semi-supervised, and resource-effective learning scheme that can: (a) learn over the existing agricultural data, whereas relying on transfer learning, (b) making use of unlabelled field data, via semi-supervised learning metrics, (c) produce highly accurate predictions with limited resource utilization even on lightweight devices, and (d) provide explanatory outputs, to assist the users in comprehending and believing in the results provided by the system.
The inventive solution is to address these shortcomings by proposing a new CNN-based solution with a transfer learning, pseudo-labelling and model compression approach--tailored at the identification of plant diseases against images of leaves in data-limited, resource-constrained conditions in real time.

3. EXISTING SOLUTIONS
Over the past years, deep learning methods (mostly Convolutional Neural Network, or CNN) have become popular within the agriculture and plant pathology community as a mean of automizing plant leaf disease detection and classification. Most of them are based on supervised learning, where experts label datasets of images of diseases, and train CNNs to perform decisions related to the identification of fungal infections, bacterial blight or viral discoloration in different crops.

Some of the best contributions include GeoTagger that trains image classification models that use highly curated datasets like PlantVillage which contains photographs of healthy and diseased leaves of various crop species. Transfer learning Via standard CNNs: Common CNN architectures such as VGG16, ResNet, DenseNet and MobileNet have been re-repurposed to produce very high classification accuracies in academic and controlled environments. The latter frequently employs complete supervision and is based on images that are balanced, noise-free and well illuminated. Although they were much better in their results in benchmark tests they have significant drawbacks when being used in the real world conditions.

In line with current methods of detecting plant diseases, transfer learning and semi-supervised learning are mainly used in overcoming the shortage of data. Vision Transformers (ViT) with dual-stage transfer learning (ImageNet pre-trained and then fine-tuned on botanical data, such as PlantCLEF2022) have up to 86.29% accuracy in low-data setting (20-shot learning) which is much higher than that of a regular CNN. In a similar manner, weight-optimized CNNs have now closed domain gaps by pre-training solely on COCO and category-specific augs (e.g., colour/rotation/noise roles) which outperform these baseline networks by 5-8% mAP on rare-disease detection.

In semi-supervised settings, pseudo-label architectures (e.g. student-teacher structures, such as FCOS/Faster-RCNN ensembles) are able to assign pseudo-labels to unlabelled examples and still reach relatively high 76% mAP using merely 10% of labelled data. Efficient models such as DeepPlantNet (25 convolutional + 3 FC layers) have a 98.49 accuracy on 8-class diseases with the usage of 80 percent fewer parameters than ResNet variants. Dual methods (e.g. Masked Autoencoders (MAE) incorporated into PlantCLEF2022) record 37 percent accuracy improvements on 20-shots learning as a result of self-supervised pre-training and domain adaptation.

To account to the lack of labelled data, some researchers have resorted to transfer learning in which models were training on other large databases (e.g., ImageNet) and adapted toward plant disease recognition. Some others have experimented with using unlabelled data through methods of semi-supervised learning, like pseudo-labelling and consistency regularization. Nonetheless, these endeavours are mostly crop-specific, and not readily transferable to newer model crops, varieties, and photo conditions.

Moreover, not many models have been designed to be deployed in resource-friendly settings. Mobile-friendly networks (MobileNet and SqueezeNet) have been investigated to use in the edge, but these imply a trade-off between latency of inference and accuracy of detection. The most published frameworks have not attracted the use of compression and quantization techniques.
Furthermore, the majority of the current disease classifiers of the plant do not possess any interpretability capabilities. Without having embedded Explainable AI (XAI) methods, there is no trust, or validation of the model output to enable a larger systemic use in the workflows of making decisions, especially by end-users of the model, i.e., the farmers or the agronomists.

Overall, despite the previous work that has established the viability of deep learning to detect plant diseases, most solutions have a tendency of encountering the following limitations:
• Reliance on massive amounts of marked data
• Failure to generalize using different domains or field settings
• Incompatibility with mobile/ low power hardware
• Little to no support of model explainability
• Lack of an integration with the semi-supervised pipelines of changing end-field data
Such shortcomings are the reason why a flexible, interpretable, and data-efficient framework with the potential to enable real-time diagnosis in different agricultural environments is required, something that the current invention delivers to a considerable extent.
Preamble
This invention concerns a new computer programmed system and technique of identifying and classifying plant leaves diseases correctly and precisely due to a powerful model of Convolutional Neural Network (CNN) that is well-tuned. To improve the performance of the models, especially in limited labelled scenarios, the system uses both the transfer learning and the semi-supervised learning methods. The incorporation of such novel machine learning methods can make the invention a powerful, scalable and data-efficient method of plant disease detection at an early stage, which will help in timely action and reduced costs and increased output in the agricultural sector. The invention can be implemented in real farming settings and equally used in the field on different crops and disease species.

6. Methodology
A fine-tuned CNN-based methodology for plant leaf disease detection and classification entails a series of combined steps, which are aimed to achieve the best accuracy and robustness, particularly in case of limited labelled data. One starts by obtaining the images of the leaves, then they undergo preprocessing procedures like resizing, normalization, and contrast enhancement (e.g. CLAHE) to standardize the input and make features stand out. To artificially increase the size of the dataset and resemble real-world variation, structured data augmentation is used and includes category specific manipulations of brightness and colour change for the chlorosis, rotation change of orientation, and noise; to simulate field effects.

Figure 1. flow diagram

After preparing the data, a pre-trained CNN model, i.e., MobileNetV2, VGG, or ResNet, is used based on which the base layers are frozen first so that the generic visual properties of large data are kept, such as the ImageNet. Additional custom top layers are added and trained on the dataset of plant diseases and part of the deeper layers are finely tuned to suit to the specific nature of leaf photographs. The pooling and convolutional layers carry out the task of feature extraction and the fully connected layers learn how to relate these features to disease categories.

Figure 2. Methodology Proposed
In order to further improve the performance with few labelled data, a semi-supervised learning protocol is proposed. In this case, a teacher-student structure: the teacher model trained on a small, labelled set provides pseudo-labels to unlabelled images. Then the lightweight student model is trained on labelled data as well as confidently pseudo-labelled data, and the errors are weighted with an uncertainty to reduce the error propagation. This unsupervised process makes use of the availability of unlabelled data in an agrarian setting; a fact which enhances generalization and robustness.

Table 1: Key Methodology Components
Step Description Techniques/Parameters
Image Acquisition Collect leaf images (field/lab sources) RGB, 224x224 pixels
Preprocessing Standardize images, enhance contrast Resize, Normalize, CLAHE
Structured Augmentation Expand dataset, simulate real conditions Brightness, rotation, noise, colour shifts
Transfer Learning Use pre-trained CNN as feature extractor MobileNetV2, VGG, ResNet; freeze base layers
Fine-Tuning Adapt model to plant disease dataset Unfreeze deeper layers, custom dense layers
Semi-Supervised Learning Leverage unlabelled data with teacher-student pseudo-labelling Confidence thresholding, uncertainty weighting
Evaluation Assess accuracy, robustness, and efficiency Cross-validation, F1-score, inference time

The last model is cross-tested on the in-distribution and out-of-distribution test sets as well as real-world field images, to determine overall robustness to conditions. The validity of the effectiveness of the model is achieved by measuring performance metrics of accuracy, macro F1-score, per-class recall, and inference time. Such an approach has the benefit of not only providing high levels of detection accuracy, which has been reported in the literature to be above 95 per cent, it also is flexible between plant species and types of disease and is efficient in terms of computation to be deployed at the edge, and practically useful to farmers and agronomists.

7. Result (Include tables, Graphs and etc..)
The effectiveness of the plant leaf disease detection system with regard to early diagnosis in agriculture use is demonstrated in its results, which makes it practical. The model was then applied to an example of curated plant leaf images, e.g. the PlantVillage dataset, via transfer learning and the usage of a convolutional neural network (CNN) model that had been previously trained and developed on a convolutional neural network (CNN) model converged on (MobileNetV2). The data was split into training and validation, and the whole process of preprocessing was carried out (including resizing, normalization, and structured data augmentation to generate variability in the real world). To prevent overfitting and ensure optimal convergence, the model was trained with Adam optimizer and early stopping mechanisms and learning rate decay option.

After training the achieved accuracy of the model was high, and the accuracy during the validation process was frequently higher than 85% and lower than 95%, which depends on the diversity and quality of the training data. Precision, recall, and F1-score were also high, averaging at 83-92 percent, which reveals how the model achieved proportionate productiveness at diagnosing diseased and healthy leaves successfully. The analysis of the confusion matrix showed that the model was competent to differentiate among the several disease classes as well as healthy ones, and a minimal percentage of them were misclassifications, frequently inside visually similar diseases. Such a healthy performance was also accompanied by training history plots, where accuracy as well as loss plots remained steadily improving between the epochs, and the heatmap of a confusion matrix, which signified the high discriminative power of the model between classes.

To confirm the practical value of the developed model, the basic web application was created in Flask, where the user could send the photo of the plant leaf and get immediate predictions concerning the type of the disease and the level of its confidence. It is mobile and user friendly, and may be used by farmers and by agricultural extension worker in the field. The backend prediction pipeline takes the image uploaded and runs the same preprocessing steps as that done during training and returns the most likely disease class with a percentage of confidence. This is the real time feedback mechanism that is important in responding in time and management of the disease.

On a state-of-the-art GPU (tested with an NVIDIA Tesla T4) training the model required around 2-3 min per epoch, whereas on a CPU-only system and hardware only supporting 32-bit floating point arithmetic, this duration exceeded 0.5 h per epoch. The practical usage of the model on the edge, e.g. on smartphones or even low-end computers, is made possible by the smaller size and penchant, decorous architecture (blame the MobileNetV2), which makes it additionally serviceable on a resource-limited system.

Table 2: Metrics used to evaluate plant leaf disease detection
Model Accuracy Precision Recall F1-Score Overall Sensitivity
CNN 0.71 0.72 0.61 0.6 0.98
VGG16 0.12 0.12 0.16 0.1 0.99
ResNet50 0.35 0.36 0.15 0.15 0.99
InceptionV3 0.93 0.94 0.92 0.93 0.99
DenseNet121 0.92 0.91 0.9 0.91 0.99
DeepPlantNet 0.98 0.98 0.98 0.98 0.99
EfficientNet-B2 0.93 0.93 0.9 0.9 0.99
YOLOv4 0.99 0.99 0.99 0.99 0.99

Figure 3. Plant Disease Detection Metrics

The table 2 and figure 3 is the comparative review of eight deep learning models in recognition of the plant leaf disease, based on five performance metrics: Accuracy, Precision, Recall, F1-Score, and Overall Sensitivity. YOLOv4 and DeepPlantNet show the most general success having almost perfect scores in all points meaning that there is perfect precision and reliability in the classification of the disease. InceptionV3 and EfficientNet-B2 rank next and have strongly balanced performance with high points in all categories. DenseNet121 is also quite robust with a fair accuracy with a little decrease in recall and F1-Score when compared to CNN. Meanwhile, VGG16 and ResNet50 perform poorly with a considerably large margin, which implies their inefficiency in the environment. It is worthwhile to mention that the entire models will have a good Overall Sensitivity, illustrating their comparable responsiveness in determining the presence of a disease with respect to differences in classifications accuracy.

Table 3: Model Size and Inference Time Comparison
Model Parameters (Millions) Model Size (MB) Inference Time (ms/image) Suitable for Edge Deployment
CNN 2.1 8.4 18 Yes
VGG16 138 528 78 No
ResNet50 25.6 98 55 Partial
InceptionV3 23.5 91 49 Partial
DenseNet121 8 33 35 Yes
DeepPlantNet 5.3 21 30 Yes
EfficientNet-B2 8.1 35 32 Yes
YOLOv4 64.4 245 24 Yes (Lite Version)

The table 3 shows the computational requirements of the different deep learning models at the level of the number of parameters, the memory usage, and average inference time per image. Such models as VGG16 and YOLOv4 although their classification accuracy is high, are not optimal in terms of sizes of their parameters, as well as model weight, and could not be used with real-time on low-power gadgets. By contrast, other models including CNN, DeepPlantNet, and EfficientNet-B2 have a positive performance to efficiency trade-off while still having competitive inference speeds and model sizes which would suit mobile or edge-based agricultural monitoring systems where resources are poor.

Table 4: Data Efficiency and Learning Adaptability
Model Pretrained Support Fine-Tuning Required Supports Semi-Supervised Learning Robustness to Domain Shift
CNN No Moderate Limited Low
VGG16 Yes (ImageNet) High No Low
ResNet50 Yes (ImageNet) Moderate Partial Medium
InceptionV3 Yes (ImageNet) Moderate Partial Medium
DenseNet121 Yes (ImageNet) Low Yes High
DeepPlantNet Yes (AgriSet) Minimal Yes High
EfficientNet-B2 Yes (ImageNet) Low Yes High
YOLOv4 Yes (COCO) Low Yes (with augmentation) High

This table 4 highlights the potential of each of the models to be adopted to new data landscapes and the work with small labelled datasets. The majority of modern architecture like EfficientNet-B2, DenseNet121, and DeepPlantNet are adaptive to transfer learning and semi-supervised fine-tuning, so they are more versatile and applicable during the real-world agricultural processes. The support of data augmentation strategies among YOLOv4 makes it stand out among the likes of older architectures, such as VGG16, and limits their usage due to low adaptability and lack of flexibility in the pretraining scheme. The newer models have the capacity to cross the gap of environmental changes, which is first important aspect in plant pathology where lighting, background, and leaf variation are so unpredictable.

Table 5: Explainability and Diagnostic Transparency
Model Built-in Explainability Visual Interpretability (e.g., Grad-CAM) Suitable for Expert Review End-User Trust Potential
CNN Minimal Yes Medium Low
VGG16 No Yes Low Low
ResNet50 No Yes Medium Medium
InceptionV3 Partial Yes High Medium
DenseNet121 Yes Yes High High
DeepPlantNet Yes (Custom) Yes High High
EfficientNet-B2 Yes Yes High High
YOLOv4 Yes (Bounding Boxes) Moderate (Localization Maps) Medium High

Table 5 emphasis is given to the quality of exposition of decision-making process in each model to users. Interpretability is important because the AI used in agriculture has to be comprehensible by farmers and agronomists. Such models as DenseNet121, DeepPlantNet, or EfficientNet-B2 have integrated visual explanations helpers like Grad-CAM, visual representation of decisions made by a model. YOLOv4 offers transparency since it can show bounds boxes yet still remains moderate compared to the transparency of traditional methods such as VGG16 and ResNet50 that provides very little transparency capabilities. It should embed interpretability functionalities that develop users confidence and facilitate the incorporation of AI tools in agricultural extension and decision-making processes.

8. Discussion
The present invention allows presenting a groundbreaking method of diagnosing plant diseases based on a semi-supervised transfer learning paradigm that is specifically designed with the concept of real-time capabilities that can support its usability in resource-limited settings. Conventional deep learning models have a high straightforwardness requirement of broad libelling datasets and high computing requirements rendering them to be hard to use in areas of agriculture where there is penury of computing programming. This invention reduces these constraints in a two-pronged manner that involves using transfer learning to leverage the availability of already trained abundant visual datasets, and the semi-supervised learning models of pseudo-labelling and consistency regularization to leverage unlabelled field data. This highly minimizes the annotation labour intensity but develops generalisation to the various types of leaves, light condition and the type of diseases in a particular leaf.

Among the characteristic features of this system, its model optimization pipeline to mobile deployment should be mentioned. Pruning and quantization of the fine-tuned Convolutional Neural Network (CNN) aims at simplifying the model and maintaining enough accuracy of detection. That is why such a framework is exceedingly appropriate to be utilized in the handheld devices and smartphones and edge-based agricultural observation grounds. Also, the use of real-time interpretability module, such as the Grad-CAM or attention heatmap, allows one to obtain comprehensible and interpretable insights regarding the results of disease classification. These characteristics are priceless in terms of gaining trust and helping farmers or agronomists to confirm automated diagnosis when there is little technical assistance or connection.

In addition to its direct technical applications the invention can imply a great potential with regards to sustainable agriculture and early intervention systems. It enables farmers to make informed decisions in monitoring the health of their plants through enabling low-cost, accurate monitoring of their plants, thereby enabling farmers to decrease crop loss, maximize pesticide use, and decrease their environmental footprint. It also forms the basis of an extensible smart-farming infrastructure since the approach allows smooth interoperability with other IoT-powered farming innovations. Simply stated, the invention fills in the divide between high-risk machine learning research and practice in agronomy, opening a way to inclusive, open, and explainable AI in global food systems.

9. Conclusion
The proposed invention is a universal and innovative way out of the old issues of diagnosis of plant diseases in contemporary agriculture. The incorporation of semi-supervised learning together with transfer learning into compact, fine-tuned Convolutional Neural Networks enables the system to maintain high accuracy even when there is scarce data available, and thus the system, in particular, can be deployed and used in real-world scenario where on-hand labeled data is limited, and there are environmental conditions that vary drastically. It makes strategic use of unlabeled field pictures to improve the generalization of the model and allow it to classify the diseases that appear in excellent success despite the different leaf textures, light conditions, and a variety of geographic areas.

More importantly, the invention is not limited to the theoretical innovation level- it is highly practical. Model compression methods (e.g. pruning and quantization) have also been included to make the solution deployable on low cost resource limited edge devices like the smartphones and portable sensors in the hands of farmers and field agents. These explainability capabilities are embedded, such that users can interpret predictions in visual ways, e.g. by generating heatmaps of them, or overlaying attention as a heatmap, which is vital to establish trust, adoption, and decision facility at a grassroots level. The intelligent diagnostics and fast on-site intervention focus of this user-friendly design creates a feeling of empowerment and an ability to get closer to the problem.

Finally, the invention helps to fill a very important gap between advanced studies in deep learning and detailed, practical concerns of the agricultural system. It makes smart diagnostics more democratized, increases the level of early detection and has a positive contribution to global objectives, such as food security, sustainable agriculture and smart-agriculture innovation. The system is a brand-new and effective development due to its scalability, transparency, and real-time accuracy, applicable in the sphere of AI-based plant pathology..
, Claims:Claims
1. A semi-supervised learning pipeline consisting of adjustable fine-tuner of a convolutional neural network (CNN) that is used to classify the images in an image classification framework that involves classification of plant leaf images with the purpose of plant disease identification.
2. The structure of claim 1, where the CNN is pre-trained through a transfer learning procedure on a huge-scale natural image dataset or agricultural dataset of a certain domain.
3. The structure of claim 1, where the phase of semi-supervised learning is based on the use of confidence-driven pseudo-libelling and consistency regularization to enhance the model recall with the unmarked instances in the field.
4. The architecture of claim 1, in which the CNN model is condensed through the process of model pruning and weight quantization to make it ready to operate in real-time on edge gadgets with limited computing resources.
5. The architecture of claim 1, where an explainability module is provided to interpret the predicted regions of disease to the non-expert end user with the help of interpreting with attention maps, Grad-CAM or saliency visualisation approaches.
6. In the architecture of the first claim, a data augmentation pipeline is added to model different imaging conditions of the leaves, e.g. lighting, rotations or occlusions.
7. A technique of real-time plant disease classification and consists in the following steps:
a. Converting leaf imagery which comes to it via a field-based imaging system.
b. DE normalizing, crops the image by normalizing and Cropping,
c. Labelling the image with the pre-trained semi-supervised CNN-based model.
d. Writing the forecasted disease type and explaining visually over the top.
8. The process of claim 7, whereby the result of the classification and graphical explanation is shown on a portable or handheld device to provide real-time support to the farmer.