The efficacy of deep learning models for tomato disease identification highly depends on the quality, diversity and annotation accuracy of the training datasets. Most models are trained on image-based datasets containing leaf samples captured in controlled environment. However, issues such as dataset bias, class imbalance and limited representation of real-field conditions remain as major challenges. The commonly used datasets for tomato disease diagnoses is listed in Table 2.

CNNs formed the basis for plant disease identification from images, because of their ability to learn hierarchical spatial features automatically. Architectures like AlexNet, GoogLeNet, VGGNet, ResNet and Inception have consistently showed high performance in detecting plant and tomato-specific diseases. Figure 3 illustrates the typical architecture and flow of CNNs in disease prediction. In the earliest studies, Mohanty et al. (2016) trained AlexNet and GoogLeNet on the PlantVillage dataset, achieving over 99% classification accuracy across 38 disease classes, including multiple tomato diseases. This pioneering work established the feasibility of using deep CNNs for automated plant pathology.

Building on this, Ferentinos (2018) assessed five CNN models (AlexNet, VGG, GoogLeNet, ResNet, and LeNet) using a large dataset containing 87,848 images of healthy and diseased plant leaf images, including tomato. The study found that classification accuracies, especially with AlexNet and ResNet, exceeded 99.53%. These results demonstrate the deep CNN’s robust learning ability when trained on large, well-curated datasets.

But, these models comprise tens of millions of parameters and rely on powerful GPUs or cloud-based infrastructure. For example, ResNet-50 has about 25.6 million parameters, which makes real-time inference on embedded or mobile phones, difficult. The deployment of AI models in remote or resource-constrained agricultural settings, where such infrastructure is unavailable, is limited by this high computational demand limits (Kamilaris and Prenafeta-Boldú, 2018).

In order to address this gap, researchers have switched to lightweight CNN architectures that retain accuracy while reducing complexity, an approach we explore next.

Even though traditional CNN models perform well, their large model size and high computational demands make them unsuitable for low-power device deployment, particularly in remote agricultural regions. In order to overcome these challenges, lightweight CNNs that balance efficiency and performance have been developed, allowing real-time inference on edge devices namely Raspberry Pi, NVIDIA Jetson Nano and smartphones. Sandler et al. (2018), developed MobileNetV2, which significantly reduces model parameters and computation by using depthwise separable convolutions and inverted residuals. The work by Ahmed et al., 2022, using a lightweight MobileNetV2-based architecture, achieved 99.3% accuracy, with a reduced computational footprint, on the Plant Village dataset (tomato). It is among the most popular models for mobile-based disease diagnosis due to its high efficiency and small memory footprint. Tan and Le (2019), introduced Efficient Net-BO, which uses a compound scaling technique to scale network depth, width and resolution. This helps to make better use of the model’s capacity without significantly adding complexity. Efficient Net-BO with an integrated attention module (Efficient NetB0-Attn) demonstrated high performance and efficiency appropriate for edge deployment, achieving 99.39% accuracy on plant disease classification (González-Briones et al., 2025).

The efficacy of a siamese network-based lightweight framework for edge-AI deployment was demonstrated by its 96.97% accuracy on the tomato subset of the Plant Village dataset with roughly 2.96 million parameters (Thuseethan et al., 2024). These lightweight models represent a key advance in expanding access to AI-based disease diagnostics, especially for smallholder farmers in regions with limited data availability. Figure 4 illustrates the overall pipeline for DL based tomato disease detection.

While CNNs are extremely effective in extracting local spatial features, they usually fail to extract long-range dependencies and global context, particularly in complex field images where background noise and overlapping symptoms is common. To address these limitations, transformer-based architectures initially created for natural language processing were utilized for computer vision tasks, including plant disease classification.

Dosovitskiy et al. (2021), introduced the Vision Transformer (ViT), which uses a self-attention mechanism to process image patches as sequences, thereby capturing both local and global interactions without the need for convolutions. On a balanced diverse tomato dataset, a multispectral approach utilizing ViT-B16 produced strong overall performance with average scores of 83.3% accuracy, 90.1% precision, 90.75% recall and 89.5% Fl-score, demontrating the efficacy of Vision Transformers in plant disease detection (De Silva and Brown, 2023). However, it is challenging to operate on low-power edge devices due to the model’s large size (about 86 million parameters), which results in significant processing demands.

To address this, researchers have developed MobileViT, a lightweight hybrid architecture that combines the inductive learning of CNNs with global modeling capability of transformers. MobileViT introduces local-global fusion blocks, that allows the model to achieve accuracy while maintaining computational efficiency. Based on this concept, a compact Vision Transformer (PMVT), modeled after Mobile ViT, demonstrated its suitability for mobile and resource-constrained deployment by achieving 94.9% accuracy with just 5.06 million parameters on plant disease detection tasks (Li G. et al., 2023). In order to achieve a strong balance of generalization, inference speed and overall performance, hybrid architectures such as CoAtNet (Dai et al., 2021) and Conformer (Peng et al., 2021) combines convolutional modules in early stages with self-attention mechanisms in deeper layers. These models offer great potential for future agricultural AI applications, even if they haven’t been thoroughly tested on tomato-specific datasets. The primary deep learning architectures for tomato disease detection are compiled in Table 3 which also highlights the dataset used, parameter size, accuracy, and edge deployment suitability. The Transformer-based and hybrid architectures constitute a paradigm shift in AI for agriculture, offering enhanced generalization, interpretability and stability in challenging real-world situations.

Most lightweight deep learning models for tomato disease diagnosis is based on RGB imagery, recent research is going on in multimodal and multi-source sensing to improve robustness under real-field conditions. There are more challenges like illumination, partial occlusion, overlapping symptoms, and background clutter in image only systems. In order to address these issues, the developing edge AI techniques combine visual information with spectral, physiological, and environmental signals, which enables earlier and trusted disease detection.

Micro-near-infrared (micro NIR) sensors can now be combined with low-power controllers such as ESP32, as a result of the advancements in portable spectroscopy. This makes it possible for real-time reflectance measurements at wavelengths linked to water stress, chlorophyll degradation, and pathogen-induced biochemical changes. Research has demonstrated that NIR-RGB fusion can increase the accuracy of early disease detection by 5-12%, especially for fungal diseases, where pre-symptomatic physiological changes takes place before the development of visible lesions (Zhang et al., 2022; Park et al., 2023). These low-cost NIR modules (<10 mW power draw) can be embedded directly into edge-AI devices without significantly increasing latency.

In the same way, lightweight CNNs are being used in combination with chlorophyll fluorescence probes (Fv/Fin) to distinguish abiotic stress from bacterial or viral infection. For example, Ruiz-Gómez et al. (2021) demonstrated that TYLCV classification has been enhanced by combining chlorophyll fluorescence with RGB features under field conditions by increasing sensitivity to early chlorosis, which is sometimes challenging to detect visually.

Additionally, environmental sensors are also essential. It is found that the accuracy of disease prediction, for climate-dependent pathogen severity such as early blight and septoria, can be increased by combining soil moisture, canopy humidity, and temperature data with late-fusion neural architectures (Haque et al., 2022). These multimodal systems enhance model generalization across growth contexts and help in differentiating disease symptoms from temporary physiological stress.

Recent research explores lightweight multimodel fusion architectures that can operate on edge devices like Jetson Nano or ARM-based mobile processors, which goes beyond simple sensor integration. Techniques such as late fusion of CNN features, MobileNet-based hybrid networks, and attention-based fusion layers optimized for embedded systems (Koirala et al., 2023). When compared to image-only models, these models can show greater resistance in situations like uneven lighting and occlusion, while maintaining inference time suitable for real-time field deployment.

All of these advancements suggests that multimodel sensing can act as a new and crucial area for tomato disease detection, that is both useful and deployable. Moreover, edge-AI systems can overcome a number of drawbacks of RGB-based methods by utilizing spectral, physiological, and environmental data. This allows for earlier and more reliable diagnoses for precision agriculture applications. Multimodal and multi-source edge AI approaches for tomato disease detection is summarized in Table 4, which includes the data modalities used, edge platforms, and their advantages over RGB-only models.

Deep learning has recently shifted from research labs to real-time agricultural conditions for tomato disease diagnosis. This change is largely influenced by edge-AI, which allows trained models to be installed directly on local hardware devices without the need for cloud-based inference. Raspberry Pi 4B, NVIDIA Jetson Nano and Android smartphones are the most common platforms used. Raspberry Pi 4B, a credit-card-sized microcomputer has a quad-core CPU and up to 8 GB of RAM, which balances both cost and performance. With the use of TensorFlow Lite, it easily supports lightweight models such as MobileNet, with inference speeds that are roughly 90% faster than the entire TensorFlow framework (Zuhair et al., 2023). Although it works well for simple tasks, its lack of GPU power may make it less effective for deeper models.

In contrast, real-time inference of moderate-sized CNNs such as ResNet and MobileNet is made possible by the NVIDIA Jetson Nano, which has a 128-core GPU and 4 GB of RAM. Because of this it is particularly useful in robotics and drone applications that need quick decisions and low latency (Tobiasz et al., 2023). Although it performs better, its increased cost and power consumption may be a drawback in farming areas with limited resources. Modern Android devices can run compressed deep learning models like MobileNetV2, with the help of frameworks like TensorFlow Lite or ONNX Runtime, especially those with adequate RAM and hardware acceleration (like mobile GPUs or NPUs). Depending on the model architecture and hardware optimization, these devices can achieve inference latencies of less than 500 milliseconds (Li Z. et al., 2023). However, in practical situations consistency may be impacted by hardware variability, thermal throttling and OS-level constraints. The features, capabilities and limitations are compiled in Table 5 to help decide whether they are fit for different deployment situations.

Model optimization is necessary for running deep learning models efficiently on edge devices, which usually have constrained memory and processing power. The most widely used optimization strategies are pruning, quantization and knowledge distillation.

Deeper architectures can be run on edge devices like Jetson Nano due to pruning techniques, as it drastically decreases the model size without affecting performance. For example, on devices like Jetson Nano, the FuPruner technique has shown efficient model compression with minimum loss of accuracy (Li et al., 2020).

Quantization decreases the size of the model and speeds up inference by reducing numerical precision typically from 32-bit floating-point weights into 8-bit integer format. Quantized models offer significant benefits in speed and energy economy, even though they lose very little accuracy (Jacob et al., 2018). They are therefore perfect for the deployment of micro-controllers and smartphones.

Knowledge distillation improves deployment efficiency by training a lightweight “student” model to simulate the output behavior of a more accurate, heavier “teacher” model. According to Hinton et al. (2015), this method can significantly reduce model complexity while maintaining a large portion of teacher’s accuracy.

The main goal, benefits and examples of each optimization method along with its limitation is listed in Table 6 to provide a comparative summary.

Edge AI has a lot of potential for smart agriculture, but there are numerous challenges to overcome before it can be used in real life settings. One major problem is dataset shift, which occurs when models were trained on ideal lab conditions (e.g., PlantVillage) are unable to generalize to field conditions because of variations in lighting, background clutter or disease stage (Kamilaris and Prenafeta-Boldú, 2018).

Inference instability caused by low-quality images such as incomplete, occluded or blurred images can reduce accuracy, which is an another major problem. Moreover, continuous usage of devices like Raspberry Pi in outdoors during hot weather might affect the performance of it due to battery limitations.

In order to address these issues, researchers are developing automated retraining processes, adaptive confidence thresholds and real-time image pre-processing that enables the model to be updated based on fresh field data. These methods are expected to be helpful in improving AI-based field diagnosis systems.

Figure 5 shows a simplified flow that illustrates the normal inference process on an edge AI device.

Plant pathogens are suppressed by microbial biocontrol agents (MBAs) by using variety of mechanisms such as:

Together these processes can produce a protective microbiome inside and around the tomato plant (Harman et al., 2004). The microbial biocontrol agents that are commonly used in tomato disease management with different modes of action are Trichoderma, Bacillus, Pseudomonas, Beauveria, and Metarhizium. For example, Bacillus subtilis competes for ecological niches by producing antibiotics whereas Trichoderma harzianum is well-known for its myco-parasitic activity and ability to trigger plant immune responses. The target pathogens and their representative biocontrol agents are shown in Table 7.

Microbial biocontrol agents (MBAs) have been shown in multiple field studies to be extremely successful in managing a variety of plant diseases. For instance, Trichoderma species are mostly recognized for their role in the biological control of plant fungal and nematode diseases through mechanisms like mycoparasitism, competition, and induced systemic resistance (Yao et al., 2023). In the same way, foliar applications of Bacillus subtilis have demonstarted efficacy against bacterial spot (Xanthomonas spp.) and early blight (Alternaria solani), with treated plants having increased photosynthetic activity, improved leaf retention and enhanced fruit development (Chowdhury et al., 2015).

Apart from single-strain applications, microbial consortia with the combinations of two or more suitable biocontrol organisms are becoming more and more popular due to their synergy. By focusing on several infections or combining anti-fungal and insecticidal qualities, these formulations frequently offer broad-spectrum protection. For example, studies have demonstrated the synergistic antogonistic activity of Bacillus-Pseudomonas consortium, leading to improved control of fungal pathogen Alternaria solani (Jia et al., 2023). Depending on the pathogen’s infection pathway, these consortia are formulated as foliar sprays, soil drenches, root dips and seed coatings.

Additionally, field-based microbial trials are becoming more focused on evaluating the formulation’s stability, climate resilience and their effectiveness in different temperature and humidity levels. In order to improve soil health and plant resistance against diseases, recent research has investigated the use of organic amendments such as fermented, composted or fresh organic treatments in arable soils (van der Sloot et al., 2024).

Despite encouraging outcomes, soil microbiota, pH and crop type, all of these have a significant impact on field success. So, for scalability and adoption, region-specific strain selection and formulation is important. The suggested microbial agents, their doses, application methods and environmental factors are listed in Table 8.

Microbial Biocontrol Agents (MBAs) are becoming more popular as environmentally friendly substitutes for chemical pesticides. But their efficacy in the field is still variable and sometimes inconsistent among research studies. These inconsistencies highlight several important challenges that must be addressed before microbial control can be consistently incorporated into precision disease management systems.

The foremost problem in the usage of microbial biocontrol is that its efficacy is sensitive to microclimatic factors. For example, Bacillus subtilis is the most commonly used foliar biocontrol agent but during high humidity, the antogonistic activity of it gets reduced. This is due to the inhibition of sporulation by excessive moisture and also it reduces persistance in leaves. In the same way, when Trichoderma species are applied in alkaline or low-organic matter soils, their efficiency is resduced. This shows that suitable environment is also essential for microbial biocontrol (Harman et al., 2004; Yao et al., 2023).

The next underlying problem is the strain adoption and colonization mechanism of the microorganisms. Because, most of the strains used for developing microbial consortia are grown in controlled environments like laboratory setups and tested in greenhouses. So, they fail to address the competition and stress conditions in agricultural fields. Field studies have demonstrated that microbial colonization is strongly dependent on soil microbiome composition, pH, temperature fluctuations and nutrient availability that leads to uneven disease control outcomes across regions (Jiao et al., 2021).

Another important conflict is that the chemical incompatibility between MBAs and certain classes of fungicides, insecticides and bactericides. For example, the viability and metabolic activity of Bacillus and Trichoderma is reduced by the combined usage of copper-based bactericides and broad-spectrum fungicides (Benítez et al., 2004). These interactions may unintentionally interfere with beneficial microbial colonization or decrease the biocontrol efficiency.

Further, the majority of MBAs have a short duration, which means they work best when they are used before or at the very early stages of pathogen infection. If application of MBAs is too late, then its efficacy can be greatly reduced, mostly affecting smallholder farming systems, because disease recognition often occurs late. It is common in quickly spreading diseases such as TYLCV or early blight. This limitation suggests the need for timely and precise application, which may be made possible by AI-based early detection systems.

At last, one of the most often constraint is the uneven field performance of the microbial consortia.

This unevenness is because of the complex interactions among environmental factors, crop genotype, native soil biota and agricultural management practices. These complications shows the importance of choosing region-specific strains, improve formulation stability and integrate digital decision support systems to increase acceptance and reliability.

All of these difficulties shows that although microbial biocontrol has a lot of potential, its practical application needs to be handled with considering all the external factors like biological interactions, environmental conditions etc. Understanding all these factors is crucial for developing strong and integrated control strategies that combine the advantages of AI-based diagnostics with microbial biocontrol strategies.

Microbial biocontrol works best when it is applied right. Mostly biocontrol agents has to be used at the earlier stages of infection or pathogen colonization as they are defensive rather than beneficial. This makes them suitable for integration with AI-powered disease detection systems, particularly those capable of real-time, in-field diagnosis.

As discussed in Section 3, lightweight deep learning models such as MobileNet or ViT-lite can be deployed on edge devices (e.g., smartphones, Raspberry Pi, Jetson Nano) to detect early visual symptoms of foliar diseases like TYLCV or early blight. After detection, the system will suggest the suitable microbial agent, the application mode and its dosage. For example, if AI detects TYLCV symptoms in tomato then it will suggest the use of Beauveria bassiana to reduce whitefly populations before they reach outbreak levels.

According to the recent study by Rejeb et al. (2022), drones are being increasingly used in agriculture from monitoring crops to delivery of biocontrols due to their ability to provide focused treatments and more efficient resource management. This reduces the environmental impact and also minimizes the input cost for farmers. Moreover, geotagging of disease instances can help schedule repeated microbial applications or automatically create maps for agricultural extension services.

In order to provide a hybrid strategy for predicting disease outbreaks and optimizing the timing of microbial interventions, some advanced platforms are linking AI detection models with IoT-based soil and weather sensors. This level of linking is needed for closed-loop precision disease management, where AI not only diagnoses but also mitigates and monitors environment friendly treatment strategies.