A Transfer Learning-Based Hybrid Deep Learning Framework for Multi-Crop Plant Leaf Disease Classification

The world's population is expected to reach close to 10 billion by 2050, necessitating a 70% increase in food production over current levels, posing significant challenges for global food supply [1]. Cassava (Manihot esculenta) [2], wheat (Triticum spp.) [3], and tomato (Solanum lycopersicum) [4] are globally important food commodities, significantly impacting food security and agriculture. Cassava, a staple food for many people in sub-Saharan Africa, supports approximately 700 million people and provides essential starch for various industries, including food processing and biofuels [5]. Wheat, a major cereal crop, is crucial to diets worldwide, especially in temperate regions. Its versatility in culinary applications has encouraged widespread cultivation, making it an important contributor to the global food production system [6]. Meanwhile, tomato, a major vegetable crop, valued not only for its nutritional content but also for its economic importance in the global market, faces threats from pests and diseases that can affect yields and economic returns [3]. Therefore, continued investment in agricultural research and management strategies is crucial for the resilience of this staple food crop to biotic and abiotic stressors [7, 8].

Cassava, wheat, and tomatoes are susceptible to various leaf diseases that significantly reduce productivity. Examples include Cassava Mosaic Disease and Cassava Brown Streak Disease in cassava [9, 10], rust and blight in wheat [11], and Early Blight, Late Blight, and Yellow Leaf Curl Virus in tomatoes [12, 13]. Plant diseases significantly threaten agricultural productivity by causing substantial annual food losses and reducing product quality [14].

The development of deep learning, particularly transfer learning, has opened up significant opportunities in image-based plant disease detection [15]. Models such as ResNet, VGG, and MobileNet have been shown to achieve high accuracy in leaf disease classification [16, 17]. However, most research still focuses on a single crop species or relatively large and clean datasets. Real challenges such as data limitations [18], symptom variation between species, and the need for implementation on resource-limited devices remain largely unexplored [19].

Based on the identified research gaps, this study proposes an efficient and lightweight transfer learning based hybrid deep learning framework with approximately 10 million parameters. The proposed model integrates DenseNet121 and a scaled MobileNetV2 (α = 0.35), enhanced by attention mechanisms, batch normalization, Swish activation, and Grad-CAM++ for improved interpretability. Experiments are conducted on three publicly available datasets, cassava, wheat, and tomato, to evaluate cross-crop generalization capability. The main contributions of this study are summarized as follows:

1. a parameter-efficient hybrid architecture (~10 million parameters) that balances representational capacity and computational efficiency by combining DenseNet121 and MobileNetV2;
2. a structured feature fusion strategy employing 1×1 convolutional compression to reduce feature redundancy while preserving discriminative information;
3. the integration of attention mechanisms, namely Attentive Transition and Spatial Attention, to enhance disease symptom localization;
4. a multi-crop evaluation protocol across cassava, wheat, and tomato datasets, addressing the limitation of single-crop focus in prior studies; and
5. edge-device deployment validation using TensorFlow Lite (TFLite), including latency and FPS analysis, which is rarely considered in existing works.

Thus, this study emphasizes not only accuracy but also efficiency and scalability, making it more relevant for practical applications in artificial intelligence-based plant disease detection systems.

This study aims to develop an efficient transfer learning-based deep learning model for multiclass classification of leaf diseases in three main commodities: cassava, wheat, and tomato. The architecture of the proposed hybrid DenseNet121 and MobileNetV2 framework, shown in Figure 1, is as follows: input images are processed in parallel by two backbones, enhanced with attention mechanisms, compressed using 1 × 1 convolutions, aggregated via Global Average Pooling (GAP), and fused for classification.

Figure 1. Design of the proposed hybrid deep learning framework combining DenseNet121 and MobileNetV2 with attention-based feature fusion

3.1 Dataset

The datasets used consisted of three public sources: the Cassava Leaf Disease Dataset [39], which includes Cassava Blight (CB), Cassava Mosaic (CM), and Cassava Healthy (CH). The Wheat Leaf Disease Dataset [40], which includes Wheat Septoria (WS), Wheat Stripe Rust (WSR), and Wheat Healthy (WH).  The dataset for Tomato leaf disease [41], which includes Tomato Bacterial Spot (TBS), Tomato Early Blight (TEB), Tomato Late Blight (TLB), Tomato Leaf Mold (TLM), Tomato Mosaic Virus (TMV), Tomato Septoria Leaf Spot (TSLS), Tomato Spider Mites Two-spotted Spider Mite (TSMTSM), Tomato Target Spot (TTS), Tomato Yellow Leaf Curl Virus (TYLCV), and Tomato Healthy (TH).  Thirteen disease classes and one healthy class comprised the total of 10,635 images. To ensure the preservation of the original class distribution, the dataset was divided into training (80%) and testing (20%) sets using a stratified sampling approach. This strategy was applied consistently across all datasets and is particularly important for handling class imbalance, such as the cassava dataset with relatively few samples per class and the tomato dataset with substantially larger class sizes. Example images are shown in Figure 2. The detailed distribution of image counts is presented in Table 2.

Figure 2. a) CB, b) CM, c) CH, d) WS, e) WSR, f) WH, g) TSLS, h) TMV, i) TEB, j) TSMTSM, k) TTS, l) TBS, m) TLM, n) TH, o) TLM, p) TYLCV

Table 2. Multi-crop dataset collection: number of training and testing images per leaf disease class [39-41]

3.2 Data preparation and augmentation pipeline

All input images were standardized to a 224 × 224 resolution [42] and scaled to the [0, 1] range via division by 255. An augmentation strategy was subsequently employed to suppress overfitting [43], applying transformations such as rotation (±30°), spatial shifts (0.2), shear and zoom (0.2), horizontal flipping, and brightness variation (0.7–1.3). After augmentation, the amount of training data increased significantly, increasing diversity.

3.3 Hybrid model architecture

The proposed hybrid model integrates two complementary CNN backbones: DenseNet121 and MobileNetV2 (α = 0.35), aiming to balance representational richness and computational efficiency. DenseNet121 is employed to capture detailed and hierarchical feature representations through dense connectivity, while MobileNetV2 provides lightweight and efficient feature extraction using depthwise separable convolutions.

Feature extraction is performed at the final convolutional stage of each backbone. For DenseNet121, features are extracted after the last dense block (conv5_block16_concat), followed by a newly defined Attentive Transition module that applies channel-wise attention to emphasize disease-relevant features. For MobileNetV2 (α = 0.35), features are taken from the output of the final convolutional layer (Conv_1) and enhanced using a Spatial Attention module implemented via a 1 × 1 convolution followed by a sigmoid activation.

In both branches, the resulting feature maps are compressed to 384 channels using 1 × 1 convolutions and then individually aggregated using GAP. The pooled feature vectors from DenseNet121 and MobileNetV2 are subsequently concatenated to form a unified representation. The classification head consists of fully connected layers with dimensions 256 → 128 → 64, where the Swish activation function is consistently applied in all layers.

1. Swish Activation Function

$f(x)=x \cdot \sigma(x)=x \cdot \frac{1}{1+e^{-x}}$   (1)

where, x: input neuron, σ(x): sigmoid function.

Swish is a non-monotonic activation function that is smoother than ReLU, thus improving training stability.

Batch Normalization and Dropout (rate = 0.5) are employed after each dense layer to improve training stability and reduce overfitting. Finally, a Softmax layer is used to produce class probabilities according to the number of disease categories.

2. Softmax

$\operatorname{Softmax}\left(z_i\right)=\frac{e^{z i}}{\Sigma_i e^{z j}}$   (2)

where, z: logit (output value before normalization) for class c, K: number of classes, P(y = c∣x): probability that input x belongs to class c. Softmax ensures that the output is a probability distribution with a sum of 1.

Attentive Transition aims to improve feature retention in the transition layers of DenseNet architectures. In traditional DenseNets, these layers reduce spatial resolutions and channel dimensions, which can inadvertently lead to the loss of critical diagnostic information. By incorporating channel and spatial attention modules within these transition layers, the model dynamically assesses and retains important features. This is achieved by calculating channel importance scores through GAP and leveraging gating mechanisms, alongside generating spatial saliency maps via convolutional aggregations of feature responses. This dual mechanism creates a sophisticated attention mask that modulates feature maps prior to pooling, ultimately enhancing the model's capability to discern and retain significant pathognomonic features, thereby improving generalization across dense blocks [44-46].

Furthermore, Spatial Attention complements the Attentive Transition by specifically targeting symptom-localizing areas within input images. This mechanism develops 2D attention maps that amplify features related to diagnostic symptoms such as lesions or necrosis while reducing the influence of irrelevant background noise. The creation of these attention maps through lightweight convolutional blocks leads to element-wise multiplication with the feature tensor, allowing the model to focus more sharply on clinically significant areas [46]. Research indicates that employing spatial attention enhances model accuracy and facilitates interpretability in the predictions made by neural networks [47, 48].

Moreover, the advantages of incorporating attention mechanisms extend beyond mere accuracies; they contribute to a better understanding of model behaviors. As shown in various studies, models utilizing attention can derive more interpretable insights, highlighting which areas of the input data were crucial for classification. This is particularly valuable in plant pathology, as determining symptoms associated with diseases can inform effective mitigation strategies and agricultural practices [49, 50].

Synthesizing Attentive Transition and Spatial Attention into the architectures of DenseNet121 and MobileNetV2 provides a robust framework for diagnosing plant diseases. This hybrid model significantly enhances the ability to extract and retain relevant features while offering clearer insights into the disease classification process. Implementing these advanced attention mechanisms aligns with the current trends in deep learning research that emphasize accuracy and interpretability [51-54].

3.4 Training process

Model training was conducted using a TensorFlow/Keras-based computing environment with Graphics Processing Unit (GPU) acceleration via CUDA on an NVIDIA T4 device with 15 GB of memory, enabling efficient and rapid training. Adam was used as the optimizer with an initial learning rate of 1 × 10⁻⁵, chosen for its ability to adaptively adjust the learning rate during training, thus promoting stable convergence. The batch size was set at 8, aligning with GPU memory limitations and ensuring good gradient quality during weight updates. The training process ran for a maximum of 50 epochs, but included an early stopping mechanism that monitored the validation loss. This meant that training would be stopped early if there was no decrease in the validation loss over several consecutive epochs, preventing overfitting and saving computational time. To improve model generalization, a dropout regularization technique with a dropout rate of 0.5 was applied, which randomly deactivates half of the neurons in a given layer during training, forcing the network to learn a more robust representation. For performance comparison, two specialized models for leaf disease detection were used: LeafDoc-Net and its modified variant, LeafDoc-Net ReLU. This comparison aims to comprehensively evaluate the performance of the proposed models, including accuracy, efficiency, and generalization capabilities for leaf image data.

3.5 Performance metrics

Model evaluation was performed using standard multiclass classification metrics: accuracy, precision, recall, F1-score, and Area Under the Curve (AUC). Analysis was performed using a confusion matrix to examine the distribution of predictions between classes. Some of the mathematical formulations used are as follows:

3. Accuracy

Accuracy $=\frac{T P+T N}{T P+T N+F P+F N}$   (3)

with: TP: true positive, TN: true negative, and FP: false positive, FN: false negative.

4. Precision

Precision $=\frac{T P}{T P+F P}$   (4)

with: TP: true positive (correct prediction for the positive class), FP: false positive (incorrect prediction, thought to be positive when it is actually negative).

Precision measures the extent to which the model accurately predicts the positive class.

5. Recall

Recall $=\frac{T P}{T P+F N}$   (5)

with: TP: true positive, FN: false negative (incorrect prediction, mistaken for a negative class when it is actually positive). Recall measures the extent to which the model is able to find all positive examples.

6. F1-Score

$F 1=2 \times \frac{\text { Precision × Recall }}{\text { Precision }+ \text { Recall }}$   (6)

F1 is the harmonic mean of precision and recall, useful when the class distribution is imbalanced.

3.6 Efficiency and deployment testing

In addition to accuracy, this study also measures model efficiency through the number of parameters and inference latency. The model was converted to TFLite for testing on edge devices. The parameters evaluated included average latency (ms), p95 latency, p99 latency, and frames per second (FPS) to assess readiness for real-time implementation.