C-MAN: A Multi-attention Approach for Precise Plant Species Classification and Disease Detection Using Multi-scale, Channel-Wise, and Cross-Modal Attentions

The study of plant leaf detection and disease diagnosis [1] occupies a pivotal position in modern-day research, offering multifaceted implications that extend across various domains, including agriculture, environmental science, and botanical studies [2]. In the ever-evolving landscape of global challenges, understanding the health and conditions of plant leaves assumes profound significance, driven by the intricate relationships between plant life and the environment.

1.1 Background

One of the fundamental aspects of this field is the identification of plant species. In a world teeming with various plants, the ability to discern one species from another is not merely an academic pursuit; it forms the basis of agriculture, forestry, and ecological conservation. Apart from the species identification, another critical task lies in this field is disease detection. The ability to swiftly and accurately diagnosing plant diseases is helpful in implementing timely interventions, minimizing crop losses, and safeguarding the environment against the spread of pathogens [3].

Inspired from the above [2, 3], in this study we determined to focus on plant leaf detection and disease diagnosis to promptly identify and classify these diseases to reduce the crop losses and minimizing ecological impacts.

Historically, researchers have primarily relied on single sensing modalities to analyze plant leaves. These methodologies require the examination of leaves through a single lens, focusing on one aspect of leaf characteristics such as color analysis, shape recognition or texture examination. For example, Sunil et al. [4] addressed the critical issue of disease detection in cardamom plants using a neural network architecture variety, specifically EfficientNetV2. Similarly, Silviya et al. [5] explored the application of deep learning for plant leaf disease detection and classification using general deep convolution techniques. In their study, Harshavardhan et al. [6] utilized the deep ResNet-34 model to identify and diagnose various plant leaf diseases at an early stage, aiming to prevent crop damage.

1.2 Problem statement

Although these methods [4-7] have provided valuable insights, they encountered the inherent challenges of relying on a single sensing modality are spatial resolution, sensitivity and specificity.

Spatial resolution [8] identified by its subsections Color analysis, Shape recognition and Texture analysis is a challenge where the traditional deep learning methods often struggle to discern fine details on a leaf's surface, making it complex to detect subtle deformations, anomalies, or early-stage disease symptoms. Focusing exclusively on color analysis at identifying prominent discolorations, will create struggles with subtle variations that characterize early-stage diseases, especially on leaves with natural color variations. For example, it is hard to differentiate subtle fungal spots from markings on variegated leaves based solely on color intensity. Similarly the Shape recognition method thrives on well-defined structures and clear boundaries, making it ineffective in detecting diseases like leaf curl or wilting that cause deformations. While accurately identifying leaf rust, which manifests as slight bumps on the leaf surface, would be challenging for shape recognition alone. In same way the Texture analysis approach is susceptible to environmental factors like lighting and leaf orientation, leading to misinterpretations of disease-related changes in texture.

Similarly the sensitivity to relevant features and specificity in identifying disease-related patterns are essential for accurate diagnosis. Single-modal methods may excel in one aspect while falling short in another, leading to potential inaccuracies in disease detection [8, 9]. These methods can miss crucial disease symptoms hidden by visual elements like shadows, overlapping leaves, or background clutter. These limitations can impede model capacity to achieve a comprehensive understanding and diagnosis of plant health, particularly when dealing with complex real-time data and disease dynamics.

1.3 Proposed model (C-MAN)

To address the inherent limitations of single sensing methods in plant leaf detection and disease diagnosis, our research proposed a Custom Multi Attention Network (C-MAN) innovatively designed to revolutionize plant leaf detection and disease diagnosis limitations of conventional single sensing methods. This approach allows us to harness the full potential of multi-modal data and provide a comprehensive solution to plant disease analysis. We customized the traditional multi attention networks model [10] into a sophisticated deep learning C-MAN architecture by employing the Multi-Scale attention, Channel-wise attention and Cross-modal attention features.

To address spatial resolution issues, our C-MAN model is utilizing the “Multi-Scale attention” mechanism at different scales of network to capture both overall leaf structure and fine-grained details, enabling detection of subtle color variations, deformations, and textural changes associated with early-stage diseases. To overcome the sensitivity issues, our C-MAN contained “Channel-wise attention” mechanisms focus on specific channels within the feature maps, allowing the model to selectively attend to disease-relevant information even when masked by other visual elements. Finally our C-MAN having “Cross-modal attention”, which is implemented to manage the specificity issues by learning the relationships between different modalities and allowing the model to leverage the contextual information from various sources to differentiate between disease symptoms and other stresses.

Our novel C-MAN model with features Multi-Scale attention, Channel-wise attention and Cross-modal attention will empowers our system to extract vital information from plant leaf images, encompassing aspects such as color, shape, texture, and disease-related patterns. Our key contributions in this research paper are:

• Developing a novel Multi-Attention Network architecture with specific features tailored for plant disease diagnosis.
• Demonstrating the effectiveness of our model in addressing the limitations of single-modality methods through comprehensive experiments.
• Providing insights into the importance of multi-modal data and attention mechanisms for improved plant disease detection.

We plan to conduct experiments using the plant leaf dataset, containing 45 thousand leaf images of various plant species with different diseases. We will evaluate our C-MAN model's performance using standard metrics like accuracy, precision, recall, and F1-score, and compare it with existing methods to showcase its advantages. By addressing the limitations of single-modality methods and leveraging the power of multi-modal data and attention mechanisms, our research aims to significantly contribute to the development of more accurate and reliable plant disease diagnosis systems.

In the subsequent sections, we will detail our methodology, dataset, experiments, and results, demonstrating how our multi-attention network-based approach significantly improves plant detection accuracy and disease diagnosis compared to previous single-modal methods, reaffirming its potential for transformative impact in various domains, including agriculture, environmental monitoring, and botany.

This research leverages a novel Custom Multi-Attention Network (C-MAN) architecture specifically designed for plant leaf analysis and disease diagnosis. C-MAN addresses the limitations of traditional deep learning methods by incorporating three key attention mechanisms: Multi-scale attention, Channel-wise attention, and Cross-modal attention. The methodology of this research is grounded in the innovative application of Custom Multi-Attention Networks (C-MAN) to plant leaf analysis and disease diagnosis. We begin by providing a detailed introduction to C-MANs, explaining their historical development and relevance in image analysis. A comprehensive exploration of the theoretical foundations of MANs follows, delving into the mathematical principles underpinning their attention mechanisms.

The architecture of our proposed C-MAN model for plant leaf analysis and disease detection is presented with Figure 1 in thorough detail, outlining network layers and component functionalities. Emphasis is placed on the integration of multi-attention mechanisms and their specific roles within the model. The deep learning techniques employed for feature extraction, including CNNs and the integration of attention mechanisms, are thoroughly explained.

image003.png

Figure 1. C-MAN model based plant leaf analysis and disease diagnosis architecture

3.1 MANs in plant leaf analysis

MANs represent an in-depth neural network architecture designed to improve the accuracy and interpretability of image analysis tasks, making them particularly valuable in the context of plant leaf analysis. These networks have evolved over time, becoming a foundational component of modern image analysis techniques [17, 18]. Historically, the development of MANs can be traced back to the growing need for models that can effectively capture intricate patterns and nuanced details within images [19]. Traditional MAN’s often struggle to discern subtle variations in plant leaves, particularly when it comes to detecting early-stage disease symptoms or distinguishing between species. Our C-MANs address these challenges by introducing a custom mechanism with Multi-scale attention, Channel-wise attention and Cross-modal attention, which allows the C-MAN model to selectively focus on specific regions of interest within an image while simultaneously considering the broader context.

In the domain of plant leaf analysis C-MANs offer a transformative solution by harnessing the power of attention mechanisms, C-MAN will extract the critical features related to leaf color, shape, texture, and disease-related patterns with remarkable precision [20]. A distinguishing feature of MANs is their ability to understand the broader context of plant images. While concentrating on specific regions of interest, C-MANs also consider the relationships between these regions and the overall image composition. This contextual understanding aids in generating comprehensive and accurate insights into plant health. C-MANs introduce the concept of selective attention using Channel-wise Attention, allowing the model to dynamically prioritize regions of interest. This selective focus enhances the extraction of relevant features while effectively filtering out irrelevant or noisy information.

3.2 Proposed C-MANs architecture

This section delves into the intricate details of our proposed C-MAN architecture specifically designed for plant leaf disease diagnosis. As depicted in Figure 1, the network leverages the combined power of Multi-Scale Attention, Channel-wise Attention, and Cross-modal Attention to achieve superior accuracy and interpretability.

Process Flow: The C-MAN begins with an Input Layer that receives plant leaf RGB images with labeling information as raw data is fed into the network for feature extraction. A sequence of CNN layers forms the backbone of feature extraction. Each layer progressively extracts and abstracts spatial features such as edges, shapes, and textures. This hierarchical approach, leveraging the inherent structure of CNNs, enables the capture of increasingly complex spatial information from the input images.

Followed by CNN, the Custom Multi-Attention Network (C-MAN) is defined with Multi-Scale Attention Module, Channel-wise Attention Module and Cross-modal Attention Module. Multi-Scale Attention module tackles the challenge of limited spatial resolution by simultaneously learning features at different scales. Parallel branches with varying kernel sizes extract low-level details (texture) and high-level features (overall structure). The extracted features are then fused to create a comprehensive representation encompassing both fine-grained and global information. Sensitivity to disease-relevant information is enhanced by Channel-wise Attention module, which focuses on specific "channels" within the feature maps [21]. Through mechanisms like squeeze-and-excitation, it learns to selectively attend to disease-specific patterns, even when masked by other elements (e.g., shadows, background clutter). Cross-modal Attention module facilitates communication and knowledge exchange between various attention channels of attention modules. This fusion of diverse information empowers the model to learn complex relationships and improve its ability to differentiate disease symptoms from other stresses based on the combined context.

After the C-MAN model, the final layers consist of fully connected neurons and a Softmax activation function. These layers leverage the extracted features to output two potential results are: identifying the specific plant species the leaf belongs to the learned features and classifying the leaf as either healthy or infected with a specific disease. The C-MAN system provides a combined output encompassing both plant species classification and disease detection results. This comprehensive information offers valuable insights for various applications in agriculture, environmental monitoring, and botanical research, assisting in timely interventions and informed decision-making.

The key advantages of our C-MAN architecture are:

• Multi-scale feature extraction: Captures both fine-grained details and overall leaf structure for improved disease detection.
• Channel-wise attention: Enhances sensitivity by focusing on disease-specific features within the data.
• Cross-modal attention: Leverages the power of multiple attentions for improved disease differentiation and specificity.
• Interpretability: The attention mechanisms offer valuable insights into which image regions are crucial for diagnosis.

3.3 CNNs for spatial feature extraction

The dataset connected Input Layer, where plant leaf images are represented as a set of pixel values in a two-dimensional matrix. Each image is denoted as $I$ and it can be symbolized as, where i and j represent the row and column indices of $I=\left( {{I}_{ij}} \right)$ the matrix, respectively. The input data consists of a batch of images for more efficient computation, which can be represented as a tensor $X$, where $X=\left\{ {{I}_{1}},{{I}_{2}},\ldots ,{{I}_{n}} \right\}$, and $n$ is the number of images in the batch. The preprocessed input tensors fed to the CNN [22] layers block for spatial feature extraction. Here, each CNN layer operates as a series of convolutional and pooling operations, which can be represented mathematically as:

${{C}_{i}}=f\left( {{W}_{i}}*I+{{b}_{i}} \right)$

Here, ${{C}_{i}}$ represents the feature map produced by the i^th convolutional layer. ${{W}_{i}}$ represents the learnable convolutional kernel (filter) for the i^th layer. $f$ denotes the activation function, which introduces non-linearity (commonly ReLU). ‘$I$’ represents the input feature map from the previous layer, and ${{b}_{i}}$ is the bias term. The '*' operator denotes the convolution operation. Alongside convolution, pooling operations are often employed in CNN layers to reduce the spatial dimensions of the feature maps can be mathematically represented as:

${{P}_{j}}=\max (I,size=s)$

where, the ${{P}_{j}}$ represents the output of the j^th pooling operation (to reduce the spatial dimensions of the feature map), ‘$I$’ is the input feature map and $\text{size }\!\!~\!\!\text{ =s}$ denotes the size of the pooling window used for pooling, which determines the region of the input feature map that is considered during the pooling operation. In a CNN, feature abstraction occurs as the data passes through multiple layers. Each layer performs a series of mathematical operations, including convolution and pooling, to abstract features from the input data.

We represent the input feature map [12] at a layer as ${{I}^{\left( l \right)}}$ (with '$l$' as the layer index) and denote the resulting feature map after abstraction as ${{F}^{\left( l \right)}}$. This process is mathematically represented as:

${{F}^{(l)}}=\otimes \left( {{I}^{(l-1)}},{{W}^{(l)}} \right)+{{b}^{(l)}}$

${{F}^{(l)}}=\sigma \left( {{F}^{(l)}} \right)$

here, ${{F}^{\left( l \right)}}$ is the feature map produced by layer '$l$' after feature abstraction, $\otimes \left( {{I}^{\left( l-1 \right)}},{{W}^{\left( l \right)}} \right)$ represents convolution operation ‘$\otimes $’ between the input feature map ${{I}^{\left( l-1 \right)}}$ from the previous layer and the learnable convolutional kernel ${{W}^{\left( l \right)}}$ for the current layer with bias ${{b}^{\left( l \right)}}$ for layer '$l$'. The activation function ‘$\sigma $’ (ReLU) [23] applied element-wise to the result of the convolution operation as $\sigma \left( x \right)=\max (0,x)$ to introduce non-linearity and enabling the network to learn complex features.

As data passes through each layer in the CNN, it undergoes convolution operations, feature abstraction, and non-linear transformations through the activation function. This progressive abstraction allows the network to capture increasingly complex spatial information, starting from simple features like edges and gradually recognizing more intricate patterns and textures in the plant leaf images as it moves deeper into the network.

3.4 Custom multi attention networks (C-MAN) model processing

After the spatial feature extraction through CNN layers, the final CNN layer output is seamlessly integrated into C-MAN model. This C-MAN model, equipped with multi scale attention [24], channel wise attention and cross model attention to play a pivotal role in refining the extracted features by focusing on salient regions within the plant leaf images. These attention mechanisms allow the network to prioritize certain areas of the image that contain valuable information related to plant species and disease presence. This integration is crucial for enhancing the precision and interpretability of the model's predictions.

3.4.1 Multi scale attention

Within the C-MANs, multi scale attention mechanisms are organized into individual attention heads [20]. Each attention head is responsible for attending to specific aspects of the image, such as color, shape, texture, or disease-related patterns [16]. The structure and training of these attention heads are optimized to ensure that they collectively capture diverse and relevant information from the input images. The attention heads are trained in parallel, and their outputs are later fused to provide a comprehensive feature representation that combines the strengths of each attention head. The integration of attention mechanisms within the C-MANs can be mathematically represented as follows:

Let ${{F}^{\left( CNN \right)}}$ be the feature map produced by the CNN layers, and ${{F}^{\left( MAN \right)}}$ be the feature map output by the Multi-Attention Networks. The attention mechanisms within the C-MANs can be symbolized as ${{A}^{\left( 1 \right)}},{{A}^{\left( 2 \right)}},\ldots ,{{A}^{\left( n \right)}}$ representing individual attention heads. Each attention head ${{A}^{\left( i \right)}}$ operates on the input feature map ${{F}^{\left( CNN \right)}}$ and produces an attention-weighted feature map ${{F}^{\left( at{{t}_{i}} \right)}}$ [25]. The combination of the outputs from all attention heads is represented as:

${{F}^{(MAN)}}={{A}^{(1)}}*{{F}^{\left( at{{t}_{1}} \right)}}+{{A}^{(2)}}*{{F}^{\left( at{{t}_{2}} \right)}}+\ldots +{{A}^{(n)}}*{{F}^{\left( at{{t}_{n}} \right)}}$

This equation illustrates how the attention mechanisms contribute to the final feature map ${{F}^{\left( MAN \right)}}$, where each attention head focuses on distinct aspects of the input to enhance feature extraction [10]. Integrating attention mechanisms in C-MANs is pivotal for locating essential regions in plant leaf images, boosting accuracy in species classification and disease detection.

3.4.2 Channel-wise attention

Building upon the multi-scale attention module, our C-MAN architecture incorporates a Channel-wise Attention Module to further refine feature representation and enhance disease specificity. This module operates within each feature map extracted by the multi-scale attention module, effectively analyzing the "channels" that hold disease-relevant information.

Spatial Squeeze and Channel Excitation: The Channel-wise Attention Module operates independently on each feature map ${{F}^{\left( MAN \right)}}$ obtained from the Multi-Scale Attention Module. This attention model incorporates the Spatial Squeeze mechanism which aggregates spatial information across each channel, collapsing the feature map into a single channel representing overall channel importance. This is typically achieved through two common methods are Global Average Pooling (GAP) and Global Max Pooling (GMP):

$GAP=S(F)=\frac{1}{(H.W)}*\sum\limits_{i}{\sum\limits_{j}{{{F}^{MAN}}}}(i,j,c)$

$GMP=S(F)={{\max }_{i}},{{\max }_{j}}\forall \left( {{F}^{MAN}}(i,j,c) \right)$

where, ${{F}^{MAN}}\left( i,j,c \right)$ represents the feature map value at position (i, j) in channel c, H and W are height and width of the feature map and $S\left( F \right)$ is the resulting single-channel vector. For channel excitation, the resulting single-channel descriptor $S\left( F \right)$ then undergoes a transformation involving two fully connected layers and a non-linear activation function (e.g., ReLU) to learn a weighting factor for each channel based on its global importance for disease detection. The excitation process ($U\left( s \right)$) for a single channel descriptor $S\left( F \right)$ can be represented as:

$U(s)=\sigma \left( {{w}_{2}}*\operatorname{Re}LU\left( {{w}_{1}}*S(F) \right) \right.$

where, ${{w}_{1}}$ and ${{w}_{2}}$ are the weight matrices of the two fully connected layers, and $\sigma $ is the activation function.

Weighted Feature Map: Finally, the learned weight vector is multiplied element-wise with the original feature map, effectively amplifying disease-relevant channels and suppressing less informative ones. This refined feature map is then fed into subsequent layers for classification as follows:

${{F}^{C-MAN}}(i,j,c)={{F}^{MAN}}(i,j,c)*U(s)$

here, ${{F}^{C-MAN}}\left( i,j,c \right)$ represents the weighted feature map for the specific channel $c$ extracted from the excitation vector, helps to perform the channel wise attention. These refined weighted feature maps are then concatenated or summed to create a more robust and disease-specific representation before being fed into the classification modules.

Focusing on relevant channels, this module reduces the influence of noisy or irrelevant background information, leading to more accurate disease classification. By analyzing the generated weights, we can gain insights into which channels are crucial for different diseases, aiding in understanding the model's decision-making process. Compared to processing the entire feature map, focusing on individual channels reduces computational overhead, making the model more efficient.

3.4.3 Cross model attention

Following the Channel-wise Attention Module, our C-MAN architecture can incorporate a Cross-modal Attention Module to generalize the multiple feature maps utilized for disease diagnosis. This module facilitates communication and knowledge exchange between multiple attentions, leading to improved disease differentiation and specificity. In this module, each channel attention is processed separately through individual CNNs feature maps resulting in distinct feature maps. Each feature map undergoes its own Spatial Squeeze and Channel Excitation steps as described in the previous section. This ensures individual channel importance is assessed within each attention. An attention matrix is computed, capturing the relationships between channels across different feature maps. This matrix represents, for each channel in one feature map, how much attention it should pay to each channel in the other feature maps of same group is scaled using a dot-product attention mechanism.

$A=\operatorname{Soft}\max \left( {{F}^{C-MAN}}*\frac{{{K}_{{{2}^{T}}}}}{\sqrt{{{d}_{k}}}} \right)$

where, A is the attention matrix with dimensions m×n, ${{F}^{C-MAN}}$ is the weighted feature map from the excitation vector of attention-1, ${{K}_{{{2}^{T}}}}$ weighted feature map from the excitation vector of attention-2, $\sqrt{{{d}_{k}}}$ is the dimensionality of the key vector. The attention matrix is used to weight the feature maps effectively amplifying relevant inter-modal relationships and suppressing irrelevant ones. These results in a fused feature representation that captures complementary information from different modalities will be discussed in feature fusion section.

By leveraging the information from different attention feature maps, the model can better distinguish disease symptoms from other stresses or background noise, leading to more accurate diagnosis. Cross-modal relationships help the C-MAN model focus on specific disease-related patterns across modalities, improving the overall specificity of disease detection.

3.5 Feature fusion

Feature fusion [26] is a critical step in the multi-attention network architecture for plant leaf detection and disease diagnosis. It involves the combination of outputs from the multi-attention mechanisms with features extracted from the CNN layers [13]. This fusion process enriches the overall feature representation and plays a pivotal role in enhancing the model's capacity to capture both spatial and salient features from plant leaf images. Mathematically, feature fusion [26] can be expressed as follows:

Let ${{F}_{m{{a}^{\left( l \right)}}}}$ represent the feature map produced by the multi-attention mechanisms at layer '$l$' and ${{F}_{cm{{n}^{\left( l \right)}}}}$ denote the feature map obtained from the CNN layers at the same layer '$l$'. These feature maps are represented as matrices with dimensions dependent on the layer, then the feature fusion at layer '$l$' is described as:

${{F}_{\text{fusion}{{\text{ }\!\!~\!\!\text{ }}^{\left( l \right)}}}}=\oplus \left( {{F}_{m{{a}^{\left( l \right)}}}},{{F}_{cm{{n}^{\left( l \right)}}}} \right)$

In the equation above, Concatenate signifies the operation of concatenating feature maps along a specified axis, often the channel axis. This operation merges the spatial features extracted by the CNN layers with the salient features accentuated by the multi-attention mechanisms at layer '$l$'. The resulting feature map, ${{F}_{\text{fusion}{{\text{ }\!\!~\!\!\text{ }}^{\left( l \right)}}}}$, integrates both the fine-grained spatial details captured by the CNN layers and the contextually important regions highlighted by the attention mechanisms at layer '$l$'. This enriched feature representation becomes the foundation for subsequent plant species classification and disease detection tasks, contributing significantly to the model's accuracy and interpretability.

3.6 Plant species classification and disease detection

The goal of Plant Species Classification is to assign each analyzed leaf to its corresponding plant species category based on the learned features from the previous stages of the model. Mathematically, this can be represented as:

Let ${{F}_{\text{fusion}{{\text{ }\!\!~\!\!\text{ }}^{\left( l \right)}}}}$ represent the fused feature map [10] obtained after feature fusion, where '$l$' indicates the layer at which feature fusion was performed. This feature map encapsulates both spatial and salient information for plant spices classification as follows:

\[P=Softmax\left( {{W}_{P}}*{{F}_{\text{fusio}{{\text{n}}^{(l)}}}}+{{b}_{p}} \right)\]

where, $P$ represents the predicted probability distribution over plant species classes, ${{W}_{P}}$ is the learnable weight matrix specific to plant species classification, ${{b}_{p}}$ is the bias term and $\text{Soft }\!\!~\!\!\text{ max}$ [23] is the activation function that normalizes the output into a probability distribution. The output probability distribution $P$ provides the likelihood of the analyzed leaf belonging to each plant species class. The model assigns the leaf to the class with the highest probability, making it a powerful tool for plant species identification.

Concurrently, the processed features of plant spices classification are employed for disease detection, a task involving the discrimination between healthy leaves and those exhibiting signs of diseases. The disease detection can be represented mathematically as:

$D=\text{Sigmoid}\left( {{W}_{d}}*{{F}_{\text{fusion}{{\text{ }\!\!~\!\!\text{ }}^{\left( l \right)}}}}+{{b}_{d}} \right)$

where, $D$ represents the predicted probability [27] of disease presence, ${{W}_{d}}$ is the learnable weight matrix specific to disease detection, ${{b}_{d}}$ is the bias term and Sigmoid is the activation function [23] that maps the output to a probability between 0 and 1. The output $D$ indicates the likelihood of the analyzed leaf being diseased. If $D$ is close to 1, it suggests a high probability of disease, whereas a value close to 0 indicates a healthy leaf. This dual classification system is crucial for assessing plant health comprehensively.

Our C-MAN model has undergone rigorous testing on a dedicated test dataset, ensuring an unbiased assessment of its capabilities. The results clearly demonstrate the model's ability in accurately classifying plant species and detecting diseases within plant leaves. The evaluation process included testing on a dedicated test dataset that was not used during training, ensuring unbiased assessment. By quantifying its performance, we can ascertain the model's overall accuracy, precision, recall, F1-score, and loss functions [31, 33, 34].

5.1 Plant species classification results

Robust performance in plant species classification is crucial. We employed 5-fold cross-validation [36] for unbiased evaluation and hyperparameter tuning to optimize C-MAN's accuracy, precision, and recall. These experiments will compare the C-MAN against the existing methods like CNN [21], ResNet [6], Multi Attention [37] and Single Attention [38]. This procedure provided an unbiased estimate of the model's performance and helped detect overfitting. Several key hyperparameters [33] were fine-tuned, including the learning rate, batch size, and the selection of optimal attention mechanisms. Through a systematic search, the best combination of hyperparameters was determined to maximize model accuracy, precision, and recall. Techniques such as dropout [34] and batch normalization [33] were explored to mitigate overfitting and enhance model effectiveness. These strategies contributed to the model's ability to classify plant species accurately.

In the context of plant species classification, the performance of various models is summarized in Table 3. The CNN model [21] achieved an average precision of 0.850 (±0.030), recall of 0.882 (±0.035), accuracy of 0.860 (±0.040), and F1-score of 0.865 (±0.025), with an average processing time per evaluation of 180.103 (±20.000) seconds. The ResNet model [6] demonstrated an average precision of 0.891 (±0.035), recall of 0.920 (±0.040), accuracy of 0.901 (±0.045), and F1-score of 0.905 (±0.030), with an average processing time of 245.678 (±25.000) seconds.

Table 3. Performance metrics for plant species classification

Model Name	Precision (±)	Recall (±)	Accuracy (±)	F1-Score (±)	Process Time (±)
CNN	0.850 (±0.030)	0.882 (±0.035)	0.860 (±0.040)	0.865 (±0.025)	180.103 (±20.000)
ResNet	0.891 (±0.035)	0.920 (±0.040)	0.901 (±0.045)	0.905 (±0.030)	245.678 (±25.000)
Sin_Attn	0.912 (±0.040)	0.943 (±0.045)	0.920 (±0.050)	0.927 (±0.035)	315.432 (±30.000)
Mul_Attn	0.930 (±0.045)	0.963 (±0.050)	0.940 (±0.055)	0.946 (±0.040)	238.917 (±28.000)
C-MAN	0.953 (±0.050)	0.984 (±0.055)	0.967 (±0.060)	0.968 (±0.045)	274.167 (±22.000)

image007.png

Figure 5. Plant spices classification accuracy comparison

The Single Attention Method [38] showed an average precision of 0.912 (±0.040), recall of 0.943 (±0.045), accuracy of 0.920 (±0.050), and F1-score of 0.927 (±0.035), with an average processing time of 315.432 (±30.000) seconds. Multi Attention model [37] achieved an average precision of 0.930 (±0.045), recall of 0.963 (±0.050), accuracy of 0.940 (±0.055), and F1-score of 0.946 (±0.040), with an average processing time of 238.917 (±28.000) seconds. Notably, our C-MAN model outperformed others with an average precision of 0.953 (±0.050), recall of 0.984 (±0.055), accuracy of 0.967 (±0.060), and F1-score of 0.968 (±0.045), while maintaining an average processing time of 274.167 (±22.000) seconds. These results highlight the effectiveness of the C-MAN model in plant species classification, offering both high accuracy and reasonable processing time is presented in Figure 5. Similarly, Figure 6 presenting the comparison of the classification time across different plant disease detection models.

image008.png

Figure 6. Comparison of classification time across different plant disease detection models

Table 4. Each plant level classification performance metrics by C-MAN

Plant Type	Precision	Recall	F1-Score	Accuracy
Mango	0.953	0.984	0.968	0.964
Arjun	0.952	0.967	0.961	0.955
Alstonia Scholaris	0.967	0.983	0.975	0.973
Guava	0.925	0.981	0.952	0.949
Jamun	0.961	0.892	0.925	0.928
Jatropha	1.000	1.000	1.000	1.000
Basil	0.954	0.959	0.957	0.958
Pomegranate	1.000	0.961	0.981	0.979
Lemon	0.954	0.958	0.956	0.957
Pongamia Pinnata	1.000	1.000	1.000	1.000
Chinar	0.943	1.000	0.970	0.969
Average	0.964	0.971	0.968	0.967

In order to provide the robust information about the experiments, Table 4 is presenting the performance metrics at each plant level as plant species classification results. From this, we observed the slight variations in our C-MAN model's performance across different plant types. For Mango, Arjun, and Alstonia Scholaris, our model demonstrated high precision, recall, F1-Score, and accuracy values, with Alstonia Scholaris achieving exceptional precision (0.967) and recall (0.983). Guava showed slightly lower precision (0.925), but compensated with high recall (0.981). On the other hand, Jamun exhibited a trade-off between precision (0.961) and recall (0.892), resulting in a balanced F1-Score (0.925).

Jatropha, Basil, Pomegranate, and Pongamia Pinnata performed exceptionally well with perfect scores in precision, recall, F1-Score, and accuracy. In contrast, Chinar displayed a specific pattern with high recall (1.000) but slightly lower precision (0.943), resulting in an F1-Score of 0.970. Our C-MAN obtained plant type classification metrics related graph with precision, recall, F1 score and accuracy are presented in Figure 7.

image009.png

Figure 7. Plant type classification metrics comparison graph

The average metrics indicate strong classification performance across all plant types, with an average precision of 0.964, recall of 0.971, F1-Score of 0.968, and accuracy of 0.967. These results showcase our C-MAN model's effectiveness in distinguishing between different plant species, with certain plant types achieving near-perfect classification.

5.2 Plant disease detection results

The Figure 8 shows a collection of detected diseased leaf images obtained through our disease detection model using C-MAN model. These images represent a diverse range of plant species and leaf types affected by various diseases. Each image provides a visual insight into the symptoms and manifestations of plant diseases, including discoloration, lesions, spots, and other identifiable characteristics. These detections demonstrate the model's ability to accurately identify and localize disease-afflicted areas on plant leaves, aiding in the early diagnosis and management of plant health.

In the context of plant disease detection, we present the comparison of our proposed model C-MAN performance metrics with its counterparts are CNN [21], ImageNet [39], GoogLeNet [40] and Inception-V4 [41] as summarized in the Table 5.

image010.png

Figure 8. C-MAN detected disease plant leaves from dataset

The results from Table 5 reveal several noteworthy findings: The baseline CNN model [21] achieved reasonable precision and recall values of approximately 0.847 and 0.819, respectively, indicating decent accuracy in identifying diseased plants, although there is room for improvement. Figure 9 is presenting the comparison of the disease detection time across different plant disease detection models.

In contrast, the ImageNet [39] and GoogLeNet [40] models exhibited enhanced performance metrics, including precision, recall, accuracy, and F1-scores, surpassing the CNN baseline. This improvement can be attributed to their utilization of more intricate architectures and pre-trained features, resulting in superior disease detection capabilities. InceptionV4 [41], while excelling in precision (0.943), demonstrated a relatively lower recall (0.909), indicating its proficiency in correctly identifying diseased plants but with potential limitations in capturing all instances.

image011.png

Figure 9. Comparison of disease detection time across different plant disease detection models

image012.png

Figure 10. Plant leaf disease detection accuracy comparisons

Remarkably, the C-MAN model emerged as the top-performing model, boasting a precision of 0.940 and an impressive recall of 0.968, highlighting its effectiveness in disease identification while minimizing false negatives. The observed trends in accuracy and F1-score largely paralleled precision and recall, underscoring the importance of achieving a balance between these metrics. Additionally, the processing time is a crucial consideration, where CNN offers faster processing while ImageNet [39] and GoogLeNet [40] demand greater computational resources. Figure 10 is presenting the leaf disease detection accuracy improvements across different epochs and emphasizing that the our C-MAN model exhibits superior disease detection capabilities compared to its counterparts, holding promise for precise plant disease diagnosis with substantial implications for agricultural practices.