Review: Artificial Intelligence and Deep Transfer Learning for Plant Disease Detection and Classification

1. Introduction

The impact of plant diseases on food security and agricultural efficiency is considerable [1,2]. Rapid and accurate detection of these diseases is crucial for implementing effective control measures and reducing crop damage. This article explores the potential of transfer learning, an advanced artificial intelligence (AI) technique [3], to transform the identification and classification of plant diseases.

Deep transfer learning is a process which uses existing models trained on large image datasets to extract relevant features for new tasks. In the field of disease recognition, these pre-trained algorithms can assimilate general visual representations from real-world images, which can then be used to recognize patterns of diseases specifically. A salient example is the work of Zahra et al. [82], whose Integrated Framework of Two-Stream Deep Learning Models: Optimal Information Fusion for Fruits Disease Recognition demonstrates how fusion architecture optimizes feature extraction, corroborating the principles outlined here. This approach offers distinct advantages over conventional supervised learning: (1) it mitigates the dependency on extensive labeled datasets—a critical limitation in plant disease detection [2] and (2) it capitalizes on hierarchical feature representations encoded in pre-trained networks, enhancing both computational efficiency and diagnostic accuracy [3].

Several research studies have highlighted the effectiveness of deep transfer learning in the classification of plant diseases. Building on this paradigm, Naqvi et al. [83] proposed a novel framework in their study Fruit and Vegetable Leaf Disease Recognition based on a Novel Custom Convolutional Neural Network and Shallow Classifier, demonstrating that the strategic fusion of custom convolutional neural networks (CNNs) with lightweight shallow classifiers achieves significant gains in computational efficiency while maintaining diagnostic precision. These innovations are paralleled by complementary applications in pest detection, such as the work of Mkonyi et al. [4], who developed a deep-learning system for early identification of Tuta absoluta infestations in tomato crops, enabling proactive mitigation strategies to safeguard yield. Similarly, Seth et al. [5] used deep learning techniques to achieve accurate classification of tomato diseases. These examples underscore the transformative impact of AI in providing farmers with rapid and reliable disease identification capabilities.

The integration of artificial intelligence (AI) with agriculture signifies the advent of a novel era characterized by precision farming [3]. Deep learning methodologies, such as convolutional neural networks (CNNs), demonstrate proficiency in image recognition tasks by efficaciously extracting features from image data, rendering them conducive to the identification of plant diseases [6,7]. Hybrid models that amalgamate CNNs with techniques such as Long Short-Term Memory (LSTM) networks further augment their capabilities by capitalizing on the strengths inherent in each approach [8].

Using deep transfer learning, researchers[9] have developed AI models capable of accurately detecting and classifying a range of plant diseases across different crops [10]. This rapid and precise detection enables farmers to take timely actions, such as applying targeted fungicides or implementing preventive measures, thereby reducing crop losses and safeguarding food security [2]. Furthermore, deep transfer learning provides an accessible and effective alternative to conventional methods, making it a valuable tool for farmers with limited resources [11,12].

This review examines recent advances in deep transfer learning for plant disease identification and classification. It explores the technical aspects of this approach, evaluates its advantages and limitations, and discusses potential future directions. By harnessing the power of AI, we can revolutionize plant disease management and promote a more sustainable agricultural future.

2. Methodology

The present review focuses on key aspects of research on the application of AI techniques, including machine learning (ML) and deep learning (DL), to plant disease classification. The methodology includes the following steps:

Systematic Literature Review:

A systematic literature review was conducted using keyword searches across Google Scholar, IEEE Xplore, and SpringerLink. Search terms included combinations of "plant disease," "machine learning," "deep learning," "classification," "image recognition," "computer vision," "agriculture," "precision agriculture," and "crop disease detection." Articles, books, and conference proceedings published within the past decade were prioritized to ensure the inclusion of recent advancements.

Selection Criteria:

The inclusion criteria focused on papers that specifically applied ML and DL techniques to agriculture, with clear experimental validation and performance metrics. Studies were considered "acceptable" if they:

• Used ML or DL for plant disease classification.
• Reported on publicly available or well-documented datasets.
• Provided measurable outcomes, such as accuracy, precision, or recall.
• Were peer-reviewed and published in reputable journals or conferences.

Exclusion Criteria Eliminated Studies That:

• Applied ML or DL techniques in non-agricultural domains.
• Lacked experimental validation or reproducibility.
• Did not provide sufficient details about their methodologies or datasets.

Technical Evaluation:

Each selected article was reviewed individually, focusing on:

• Problem Addressed: The specific plant diseases or agricultural issues being tackled.
• Techniques Used: ML or DL algorithms and architectures employed.
• Data Sources: The origin, size, and diversity of the datasets used.
• Performance Metrics: Overall accuracy, robustness, and scalability of the models.

Focus on Performance:

The present review aims to evaluate the effectiveness of machine learning (ML), and deep learning (DL) techniques based on their classification accuracy and other performance indicators. The analysis will identify patterns and trends in successful implementations, thus guiding the selection of optimal methods for plant disease classification. The objective of this analysis is to highlight the most effective and efficient artificial intelligence (AI)-based solutions that can advance agricultural practices.

3. Comprehensive Survey on Classification and Detection Techniques for Plant Leaf Diseases

Due to the increasing demand for food production, research on automatic detection and classification of plant diseases has developed rapidly in recent years. Researchers are harnessing the power of artificial intelligence through various machine learning and deep learning techniques to analyze large amounts of data. These techniques are trained on different datasets of images of healthy and diseased plants, allowing them to detect subtle visual signs of disease. This article reviews some of these promising methods and their capabilities and potential impact on agriculture:

Nithish Kannan et al. proposed a deep convolutional neural network (CNN) based on the ResNet-50 architecture, achieving an accuracy of 97%. Utilising the PyTorch framework, the researchers classified six tomato diseases and enhanced the model's performance through data augmentation and validation using parameters derived from the ResNet-50 model. The study utilised the PlantVillage dataset, which comprised 12,206 images, augmented to 39,204 images, and addressed challenges such as overfitting and limited training data.The model was also able to tackle hardware constraints and complex disease patterns in real-world leaf imagery [13].

Ashok et al. proposed a methodology based on CNN which has a classification accuracy of 98.12%. The model was trained on a dataset comprising 10,000 high-resolution retinal images, addressing challenges such as imbalanced classes and subtle lesion variations caused by lighting inconsistencies and anatomical noise [14].

Magsi et al. focused on identifying Sudden Decline Syndrome (SDS) in date palms at various infection stages using a dataset of 1,200 leaf images. The researchers employed convolutional neural networks (CNNs) alongside a hybrid feature extraction approach. For colour analysis, the images were converted into the HSV colour space to enhance infection detection, while RGB analysis was utilized for precise colour quantification. In terms of texture extraction, the study used the grey-level co-occurrence matrix (GLCM) to capture spatial relationships within the images, and the Scale-Invariant Feature Transform (SIFT) was applied to detect key features and calculate area ratios. This integrated methodology achieved an overall accuracy of 89.4%, with an impressive 99% accuracy for late-stage (Stage 4) detection. This provides valuable insights for effective disease management in date palm cultivation [15].

Peng Jiang et al. used a deep disease detection model based on CNN. The proposed model could detect diseases with high accuracy with real-time input images, obtaining a 78.80% detection rate of mAP [16].

Qimei Wang et al, proposed object detection models using a deep CNN architecture. They achieved the highest rate and best performance of 99.64% mAP by combining Mask R-CNN with ResNet-101[17].

Akshay Kumar et al. proposed CNN-based architectures. In the proposed model, VGGNet performed well and had an accuracy of 99.25% [18].

Mehmet Metin Qzguven et al. propose a faster R-CNN architecture. The proposed model is time-consuming in terms of disease detection rates. A maximum and overall classification accuracy of 95.48% is achieved [19].

Karthik R.et al. proposed a methodology, namely the attention-based residual convolutional neural network, which achieved a classification accuracy of 98% [20].

Thair A. Salih et al. proposed a deep learning model based on a convolutional neural network (CNN) for the detection and classification of diseases affecting tomato plants. The model, which consists of 14 layers, achieved a classification accuracy of 96.43% when trained and tested on a dataset of 6,202 images obtained from the Plant Village dataset. The images were divided into six categories: five types of diseased leaves and one category of healthy leaves [21].

Yang Zhang et al. set out their approach for enhancing the Faster RCNN model with ResNet101 for the purpose of feature extraction. Utilising a dataset comprising 4,178 tomato leaf images, categorised into four distinct disease categories (namely, powdery mildew, blight, leaf mold fungus and ToMV), their approach resulted in an accuracy of 98.54% mAP (mean average precision), following the application of k-means clustering to refine bounding box anchors [22].

Nitish Gangwar et al. conducted a study focusing on the classification of grape leaf diseases, addressing challenges in identifying and categorizing leaves affected by diseases such as black rot, Esca (black measles), and leaf blight.The researchers utilized the InceptionV3 network, fine-tuned for the task by leveraging transfer learning.Specifically, the model acted as a feature extractor, and a logistic regression classifier was applied to achieve disease classification. This approach led to a substantial reduction in training time while attaining a state-of-the-art accuracy of 99.4% on the test dataset.The study's findings underscore the potential of transfer learning in automating disease detection and supporting agricultural practices [23].

Vallabhajosyula et al. Propose a deep ensemble neural network that uses transfer learning to improve disease diagnosis of plant leaves. The network incorporates pre-trained models such as ResNet 50 & 101, InceptionV3, DenseNet 121 & 201, MobileNetV3, and NasNet. The proposed method outperforms the most advanced models available, highlighting its superiority in the detection of plant leaf diseases [24].

Aversano et al. propose an approach that utilises models known as VGGNet and ResNet, incorporating approximately 1,600 images for the purpose of classification into ten distinct classes. The VGGNet model demonstrates an accuracy of 97%, exhibiting commendable precision [25].

Saleem et al. used a comparative analysis on 26 category classifications using various pre-trained deep networks including ResNet-50 and OverFeat, the best model CNN was selected and the performance of the model was further improved by deep learning optimizers, and results showed that the model trained with Adam optimizer achieved the highest of 99.81% [26].

Nawaz M. et al. proposed a powerful deep method called Faster-RCNN based on ResNet-34 to address disease detection and classification of tomato leaves. This approach using the Convolutional Block Attention Module (CBAM) achieves exceptional accuracy and mean average precision (mAP) scores on the PlantVillage Kaggle dataset. The proposed approach aims to replace manual disease detection devices, providing a cost-efficient and automation-compatible solution [27].

Mohit Agarwal et al. Present an article that introduces a new method for accurate detection and classification of diseases of tomato leaves using a convolutional neural network. The strategy includes three layers of convolution with max pooling. The proposed model successfully addresses the critical challenge of identifying diseases in tomato crops, outperforming pre-trained models like VGG16, InceptionV3, and MobileNet by achieving an average accuracy of 91.2% [28].

S. Jeyalakshmi and R. K. Radha proposed an innovative approach to identify and classify diseases of tomatoes. They used an improved automatic GrabCut image segmentation algorithm to efficiently extract healthy and diseased leaf regions. Ensemble learning frameworks include random forests, multi-layer perceptron, and support vector machine (SVM) classifiers. By combining their predictions using a soft voting classifier, they obtained an accuracy of 93.13% in classifying tomato diseases. This dataset includes 1817 images of tomato leaves infected with tomato spotted wilt virus (TSWV) and tomato yellow leaf curl virus (TYLCV) [29].

Kibriya et al. present a methodology for the effective identification of diseases affecting tomato leaves. The approach utilizes GoogLeNet and VGG16 convolutional neural network (CNN) models, attaining noteworthy accuracies of 99.23% and 98%, respectively [30].

Parvez et al. present a deep-learning methodology to identify leaf diseases of tomatoes at an early stage. Using convolutional neural networks, including GoogLeNet and VGG16, the model achieves an impressive 98.39% testing accuracy on a dataset including 6,926 tomato plant photos. The study seeks to boost agricultural output and profitability by equipping farmers with an efficient tool for autonomous disease identification and early prevention [31].

Nagamani H. S. et al. propose a study that uses fuzzy support vector machine (fuzzy-SVM), convolutional neural networks (CNN), and region-based convolutional neural networks (R-CNN). R-CNN-based classifier achieves a remarkable 96.735% accuracy in early diagnosis of disease using advanced approaches such as image scaling, thresholding colors, and the local ternary gradient pattern. The research improves the field of agriculture by presenting a streamlined and automated disease detection method [32].

Al-gaashani et al. propose a new hybrid method that integrates transfer learning with feature concatenation. They used pre-trained MobileNetV2 and NASNetMobile kernels to extract features from tomato leaf images. These functions are concatenated and dimensionally reduced via kernel principal component analysis. Traditional learning algorithms then process the reduced features. Connected features have significantly improved performance reaching an impressive average accuracy of 97%. The researchers assessed the performance of three conventional machine learning classifiers: Random Forests, Support Vector Machines, and Multinomial Logistic Regression. Among the options considered, multinomial logistic regression is the most effective classifier [33].

Lakshmanarao et al. predicted plant diseases by applying a transfer learning technique. The Plant Village dataset, collected from Kaggle, was used. The actual dataset was segmented into three each assigned to different plants. They applied three transfer learning techniques: VGG16, RESNET50, and Inception, achieving accuracies of 98.7%, 98.6%, and 99% respectively [34].

Attallah et al. propose a pipeline for the identification of tomato leaves. Compact convolutional neural networks are used, and transfer learning is applied to extract deep features. Additionally, a hybrid feature selection approach is employed to reduce dimensions. The results demonstrate impressive accuracy: 99.92% using K-nearest neighbor and 99.90% using support vector machine classifiers [35].

Borugadda et al. propose a new approach to classify leaf diseases of tomato plants using transfer learning with the VGG16 model. They use the Plant Village dataset, containing 18,160 images across 10 classes including nine disease categories and one healthy class. The model achieves an impressive accuracy: 95.68% in MLP and 95.79% in VGG16. Their work contributes significantly to early disease detection in tomato crops, potentially preventing crop damage and increasing yield [36].

Kaur, Harnal, Gautam, Singh, and Singh propose a novel approach using the modified CNN model InceptionResNet-V2 (MIR-V2) to detect tomato leaves. They achieve an impressive accuracy rate of 98.92% and an F1 score of 97.94%. The model is trained on both a public dataset and a self-collected dataset including seven different tomato leaf diseases as well as healthy leaves. This research focuses on the potential of deep learning to improve precision agriculture and crop management [37].

Liu G et al. The scientists explored deep convolutional networks for plant disease recognition. They proposed the Selective Kernel MobileNet (SK-MobileNet) model, achieving an impressive 99.28% accuracy on a public dataset. This lightweight approach outperforms existing methods while maintaining computational efficiency. The study contributes to automated plant disease detection, crucial for agricultural productivity, using visible range images despite background complexity and precise parasite localization challenges [38].

Nayak A. et al. focused on the study of rice diseases and nutrient deficiencies in images taken with smartphones. They employed image processing techniques and model optimization, utilizing 2,259 smartphone images of rice plant parts across different classes. The study achieved over 90% accuracy in diagnosing nutrient deficiencies using deep convolutional neural networks (DCNNs), with DenseNet121 performing exceptionally well. The dataset contains 250 live validation images representing 12 different rice diseases and nutrient deficiencies. Their work contributes to enhancing agricultural practices through technology, enabling early plant health issue detection and informed decision-making [39].

Sapna Nigam et al. proposed an innovative approach to automatically detect major wheat rusts using deep learning techniques. They prepared the WheatRust21 dataset, including 6,556 images of healthy and diseased wheat leaves collected from natural field conditions. The authors experimented with classical CNN-based models, achieving accuracy ranging from 91.2% to 97.8%. However, their fine-tuned EfficientNet B4 model achieved an impressive test result of 99.35%, making it suitable for mobile applications in the field of wheat disease identification [40].

Han Jiang et al. aimed to identify plant disease species using a transfer learning algorithm applied to a ResNet model. They used an open-source dataset of black rot, bacterial spot, rust, and healthy leaf samples. The transfer learning approach significantly improved accuracy, achieving 83.75% identification accuracy, outperforming the ResNet-101 model. Their study highlighted the feasibility of transfer learning-based plant disease detection models, which is a promising approach to improving agricultural practices [41].

El Massi et al. suggested that a hybrid approach proved effective in their study. The hybrid combination(HC) gave the best results, with an overall detection rate of 91.11%, compared to 88.33% for the serial combination(SC). The first variant, known as SC, combines two SVM classifiers, S1 and S2, in series. The system they proposed employs two variants of combination: serial and hybrid [42].

Marino et al. proposed a methodology for the identification and classification of imperfections in potatoes. They developed a labelled dataset comprising six categories and multiple breeds, and employed a multi-camera setup to capture images of the potatoes.The combination of autoencoders and support vector machines was proposed for the identification of damaged and green areas in selected images, with the localization results being utilised as input for the SVM classification. The efficacy of this methodology is evidenced by its attainment of an average accuracy of 95% and an average recognition rate of 93% [43].

4. Review Table

This review provides a comprehensive analysis of the various technical research methods employed in relevant studies. Table 1 provides a detailed comparison, describing the specifics used in each study. Additionally, the Table summarizes the main findings of each method and provides citations for the studies. This comparative overview effectively highlights the diversity of techniques used by researchers to address research challenges.

These techniques encompass established methods like regression analysis and extend to cutting-edge machine learning tools such as neural networks and supervised post-hoc regression.

The primary objective of this comparative analysis is to equip researchers with an in-depth understanding of the strengths and limitations of each approach, thereby allowing them to make informed decisions when designing future studies of research.

Despite the satisfactory results yielded by existing models, they are deficient in terms of disease detection at an early stage.The proposed hybrid CNN-RNN architecture synthesises advancements in temporal-spatial feature fusion from extant literature [16,22,32,39], thus offering a roadmap for early disease detection. The integration of the proposed architecture with the Internet of Things (IoT)-enabled environmental monitoring aligns with emerging AI-IoT paradigms [44,46], which prioritise scalable, real-time agricultural diagnostics Figure 1.

Part 1: IoT Data Acquisition

This section focuses on data including temperature and humidity, collected by the system. The data is processed by a microcontroller before being wirelessly to the cloud for storage. This direct and continuous monitoring aligns with the advancements in the Internet of Things (IoT) for agriculture, as discussed by D. Thotho and P. Macheso. They highlight the critical role of IoT in precision agriculture for applications such as smart irrigation and environmental monitoring [44].

Part 2: CNN-RNN Processing in the Cloud

• Data Preprocessing: In this phase, both the sensor data and the plant images are subjected to preprocessing. It may be necessary to normalize or filter sensor data in order to ensure its quality. Similarly, images must be prepared for deep learning analysis.
• Feature Extraction: The convolutional neural network (CNN) is employed to analyze the preprocessed images, with the objective of extracting spatial features that indicate the presence of disease. The application of convolutional neural networks (CNNs) for pattern recognition in disease detection represents a crucial technology in precision agriculture, as observed by S. K. Javheri, who underscores the significance of artificial intelligence (AI) in crop health monitoring and disease detection [45].
• Sequence Modeling: Recurrent neural networks (RNNs) process the continuous data from sensors in order to capture temporal trends, which are crucial for understanding how plant health changes over time. This step is of particular importance for the early detection of diseases, which represents a central objective of the proposed architecture and is supported by the existing literature on the application of AI in agriculture.
• Feature Fusion and Classification: The features derived from the CNN and RNN are integrated to create a comprehensive representation, which is then classified using models such as SVM and Logistic Regression to determine the health status of the plant. This fusion exploits the spatial-temporal data to achieve accurate early detection and classification, thereby enhancing the precision of decision-making in agricultural contexts.

Core Projects and Literature Integration

The incorporation of the Internet of Things (IoT) for the continuous acquisition of environmental data and the combination of convolutional neural networks (CNNs) and recurrent neural networks (RNNs) for its subsequent processing represent a crucial step in the advancement of disease detection capabilities in agriculture. In their review of IoT-enabled smart agriculture, Mohy-eddine et al. also discuss the potential security concerns and highlight solutions such as blockchain for secure and reliable agricultural data management [46]. This highlights the necessity for robust systems such as the proposed hybrid CNN-RNN model, which not only detect diseases at an early stage but also guarantee data integrity and security. By adopting this hybrid model, the agriculture sector can capitalise on the advantages of both AI and IoT, achieving a higher level of precision in disease detection and enhancing overall crop management strategies, which are crucial for the sustainability and efficiency of modern agricultural practices.

5. Critical Analysis

The reviewed approaches demonstrate notable strengths, as evidenced by their high accuracy rates, which frequently exceed 95%. This underscores the efficacy of deep learning models, particularly convolutional neural networks (CNNs) such as ResNet and VGG, in extracting spatial features from plant images. Moreover, the application of transfer learning has been shown to be highly advantageous, substantially reducing the requirement for extensive, labelled datasets while maintaining robust performance across a range of crops and disease types. However, it is important to acknowledge several limitations. A notable weakness is the over-reliance on curated datasets, such as PlantVillage, which are often collected under controlled laboratory conditions and lack the variability and complexity encountered in real-world agricultural environments, limiting the generalizability of the models. Furthermore, the computational demands of advanced architectures like VGG16 and ResNet hinder scalability, particularly for deployment in resource-constrained settings. Another critical limitation is the insufficient exploration of early-stage disease detection and the simultaneous identification of multiple diseases, which are common challenges in practical agricultural scenarios.

To address these limitations, several research gaps have been identified. First, there is a pressing need for the development of lightweight models, such as MobileNet, which can operate efficiently on edge devices with limited computational resources. Secondly, the incorporation of multimodal data, encompassing IoT sensor inputs (e.g., temperature, humidity) and spectral imaging, has the potential to enhance the robustness and accuracy of disease detection systems. Finally, there is a need for research to develop models that are resilient to environmental variations, such as changes in lighting conditions, occlusions, and background noise, to ensure reliable performance in diverse agricultural settings. Addressing these gaps will be critical to advancing the practical applicability of deep learning in plant disease detection and management.

6. Results and Findings

The period from 2019 to 2024 saw transformative advancements in deep transfer learning for plant disease detection, with 59 reviewed studies demonstrating its efficacy across crops such as tomato, rice, and date palm. Hybrid architectures (e.g., CNN-RNN fusion) and lightweight models (e.g., SK-MobileNet) emerged as leading contenders, achieving accuracies of 78.80–99.92% (Table 1). It is noteworthy that models such as ResNet-101 and InceptionV3 consistently exceeded 95% accuracy for tomato diseases, while ensemble networks (e.g., Vallabhajosyula et al. [24]) outperformed single-model approaches through feature concatenation.

▪ Key challenges persist, including data scarcity, limited labelled datasets for rare diseases, high cost of labelling, and variability in plant disease manifestations across different environmental conditions. Computational demands, such as the training times for Faster R-CNN, and generalization gaps, like performance drops under variable lighting and backgrounds, also pose significant hurdles. Transfer learning has been demonstrated to reduce data dependency, with pre-trained models such as VGG16 and MobileNetV2 achieving over 90% accuracy even on small datasets like 1,200 date palm images. The integration of the Internet of Things (IoT), including sensor-RNN temporal analysis, has further enhanced early detection by capturing environmental correlations.

▪ Notable innovations include attention mechanisms (e.g., CBAM in Nawaz et al.) for localized disease features and data augmentation (e.g., Jiang et al.’s [16] rainbow concatenation) to reduce overfitting. However, real-world scalability remains constrained by hardware limitations and the need for farmer-friendly interfaces.

▪ Research underscores the potential of artificial intelligence, as efforts are focused on developing lightweight architectures (e.g., EfficientNet-B4 for mobile deployment) and multimodal fusion (e.g., spectral + visual data). These advancements position AI as a cornerstone for sustainable agriculture, offering rapid, precise diagnostics to safeguard global food systems.

7. Future Directions

Following a thorough review of Table 1, the following refined directions emerge to address current gaps and leverage technological advancements:

• Advanced Hybrid Architectures: Future models should prioritise hybrid frameworks (e.g., CNN-RNN, attention mechanisms) to integrate spatial-temporal data, enhancing early disease detection. This finding is consistent with studies such as Jiang et al. [16] and Zhang et al. [22], which underscore the efficacy of fused architectures for enhanced accuracy and real-time performance.
• Scalable and Diverse Datasets: Addressing dataset limitations, including class imbalance and a paucity of real-world variability, necessitates the curation of large-scale, multimodal datasets. The augmentation of synthetic data and the establishment of collaborative platforms for the dissemination of agricultural data in the public domain have the potential to mitigate biases and enhance the generalisability of models across a range of crops (e.g. millet, potatoes) and environments.
• Edge-AI and IoT Integration: The deployment of lightweight, energy-efficient models (e.g. MobileNetV3, EfficientNet) on edge devices will facilitate real-time, on-field diagnostics. This reduces reliance on cloud infrastructure, as seen in Nayak et al. [39], and supports IoT-enabled systems for continuous environmental monitoring and rapid response.
• Explainable AI (XAI): Bridging the trust gap between farmers and AI necessitates interpretable models. Techniques like Grad-CAM or attention visualization, as explored by Karthik et al. [20], can clarify decision-making processes, fostering adoption in practical settings.
• Cross-Domain Generalization: Advancing beyond the confines of a single crop (e.g., tomatoes), future research should prioritize the development of transferable models adaptable to diverse crops and geographies. This necessitates benchmarking against heterogeneous datasets and the incorporation of unsupervised and semi-supervised learning to mitigate the dependency on annotation.
• Sustainability-Driven Innovation: Developing resource-efficient algorithms (e.g. pruning, quantization) will democratize access for smallholder farmers. Moreover, the integration of artificial intelligence (AI) with climate-smart practices, such as the prediction of disease outbreaks under changing environmental conditions, has the potential to enhance global food security.

By addressing these priorities, AI-driven solutions have the capacity to exceed academic standards, delivering scalable, equitable, and actionable tools for sustainable agriculture.