Iterative segmentation and classification for enhanced crop disease diagnosis using optimized hybrid U-Nets model

Introduction

Precision farming has completely altered the monitoring and management of crop health. Late detection and inappropriate diagnosis are two major causes of crop loss; hence, precise and timely diagnosis of diseases becomes very imperative from the point of view of crop yield and sustainability. Crop diseases, on the other hand, remain one of the major challenges to improving productivity and hence economic viability in agriculture. In contrast, that has improved considerably due to advances in technology. This therefore underlines the critical need for such innovative solutions that not only identify but also classify crop diseases with a high level of precision and efficiency. Techniques that have been introduced as promising ways of addressing these challenges are deep learning techniques. Yet, the intricacy of symptoms and variability in the quality of captured image data calls for further improvements in this domain. The deficiencies of traditional models, in terms of precision, accuracy, and the reality of running the model in real-time, spite the need to find efficient models. Traditional approaches have been inefficient in handling noisy and high-dimensional data typical of agricultural settings found in real-world scenarios. This leads to the development of models that handle such data robustly while ensuring a high degree of accuracy in disease classification.

It is in the light of this that an all-inclusive framework is proposed for the management of crop disease diagnosis, incorporating novel image processing and deep learning techniques to achieve this task. One key building block of such a framework lies in the use of adaptive anisotropic diffusion for image denoising, greatly improving the quality of input data fed into the processing pipeline. Realizing the demands of proper segmentation in disease diagnosis, we use a Fuzzy U-Net++ model that applies fuzzy logic to enable accurate segmentation. The very original application of fuzzy decisions instead of binary ones by this model has resulted in a radical departure from the traditional approach and has given much more nuance and effectiveness to the segmentation of the diseased regions in crop images.

The Moving Gorilla Remora Algorithm (MGRA) scheme increases the capability of our framework through feature selection. Coupled with convolutional operations, this new paradigm would make it possible to identify and select the most discriminative features for disease classification, trying to improve model performance. For classification, a modified LeNet architecture was applied one of the most efficient models for pattern recognition problems, which has been beautifully fine-tuned for the task of crop disease classification.

The efficacy of the proposed model is duly tested with the aid of two preeminent datasets attendant to this sphere of activity: PlantVillage and PlantDoc. This evaluation goes on to confirm that results from our model turn out better compared with methods available today, having huge potentiality for enhancing precision, and accuracy, and raising the speed of crop disease diagnosis. We therefore place our contribution within the scope of precision agriculture by providing farmers, agronomists, and researchers with a robust tool against crop diseases. This research brings agricultural innovation to a totally different level by showing potential through cutting-edge AI techniques to improve crop management practices.

In-depth review of existing models

Looking through the spectrum of inventions, all targeted at improving accuracy, efficiency, and applicability within different agricultural contexts, a systematic review of the current methodologies for plant disease detection and classification gives way. This has been possible through a careful review of recent publications in the genre, which exposes a universal move toward advanced technologies that emerge clearly in deep learning, Internet of Things (IoT), and machine learning models, as well as contrastingly advanced data augmentation techniques and newer forms of convolutional neural networks.

The common factor contributing to these studies is the attempt to rectify and reduce the limitations of conventional plant disease diagnosis techniques. Notably, such includes generalizability across a wide variety of plant species and diseases, ease of handling, scalability, and processing in real-time proposed systems, and the practical concerns involved in the implementation of IoT in agricultural environments. Also, the use of high Volume and heterogeneous data in training models, together with the computational complexity of the methods, makes relevant challenges that should be carefully addressed.

Table 1 analysis evidences an important turn towards more sophisticated models elaborated to provide other than a more accurate refinement of the process related to disease detection and classification a way towards precision agriculture. Leveraging deep learning with IoT, further coupled with the strategic application of data augmentation and attention mechanisms, it is in these studies, contributing to the advancement of smart farming solutions. The road ahead demands balance in its stride through the intricate dance between technological innovation and the pragmatic challenges in its implementation process.

Looking through the detailed methodologies and findings presented by the reviewed articles goes a long way to show and provide passage to the transformative, evolutionary era of the domain of plant disease detection. Deep learning with IoT integration and the new use of data augmentation techniques like LeafGAN, along with attention mechanisms, signals a new era entering agricultural technology. This shows that their enhancement of the convolutional neural network (CNN) architecture and further the strategy of feature fusion have really improved the performance of disease detection systems in terms of precision and efficiency, supporting the potential of artificial intelligence to realize a revolution in the field of plant pathology.

Considerable limitations arise scalability of the systems to deal with real-time processing, model generalizability to a greater range of diseases and plant species, and logistical challenges to IoT deployment. Furthermore, large and very diverse datasets are required to train the models effectively in some cases, whereas many other methods are very computationally intensive, thereby making full-scale adoption very difficult.

In summary, the review emphasizes that further studies will be needed to unlock ways to confront the complexities and limitations accrued. The scope in the near future is immense for studies to investigate very new solutions by which plant disease detection systems could allow better generalization and scalability. This will build on the model development that is technologically advanced but practical and accessible for real-world agricultural applications to exploit the fullest potential of artificial intelligence (AI) in sustainable agriculture. The journey to smart farming is quite challenging, though replete with opportunities for ground-breaking research that can bridge the gap between theoretical models and practical, high-impact solutions in plant disease diagnosis and management.

Design of the proposed model for disease analysis

In view of the review of existing methods used for the analysis of plant diseases, it has been observed that most of these methods either have low efficiency or are complex to implement in real scenarios. In this section, an efficient deep learning process design will be elaborated by fusing efficient denoising, segmentation, and classification operations. According to Fig. 1, the application of adaptive anisotropic diffusion for denoising collected agricultural images is the key element of the framework designed for advanced crop disease diagnosis. Diffusing within the region of homogeneity without significantly affecting edge features is very important in effective disease identification operations. This methodology retools the original anisotropic diffusion method, as proposed by Perona, Shiota & Malik (1994), to best accommodate the variability that exists within agricultural imaging where minute lesion details are paramount to various disease classification operations. The mathematical formulation for adaptive anisotropic diffusion commences with the general heat conduction operation, represented via Eq. (1),

(1) $${\displaystyle \frac{\mathrm{\partial }I}{\mathrm{\partial }t}=\mathrm{\nabla }\cdot \left(c\left(x,y,t\right)\mathrm{\nabla }I\right)}$$where I represent the image intensity function, t represents the iteration time, $\mathrm{\nabla }$ signifies the gradient operator, and c (x, y, t) is the diffusion coefficient, dictating the rate and scope of diffusion.

This diffusion coefficient is crucial, as it is modulated to enable adaptive behavior in response to the image’s local features. This coefficient is adapted based on the edge-stopping function, which serves to preserve significant edges corresponding to disease symptoms or boundaries. The edge-stopping function, $f\left(||\mathrm{\nabla }I||\right)$, is articulated via Eq. (2),

(2) $$f\left(\parallel \mathrm{\nabla }I\parallel \right)={\displaystyle \frac{1}{1+{\left(k\parallel \mathrm{\nabla }I\parallel \right)}^2}}$$where k is a contrast parameter that dictates the sensitivity of edge preservation. The lower the value of k, the greater the sensitivity to edges. The diffusion process is iteratively updated using a discretization approach, which is translated via Eq. (3),

(3) $$In+1\left(x,y\right)=In\left(x,y\right)+\lambda \mathrm{\Delta }t\sum _{\mathrm{\forall }\left(i,j\right)\in \mathrm{\Omega }}ci,j\left(x,y,t\right){\mathrm{\nabla }}^{2}In\left(x,y\right)$$where λ is a scaling parameter that controls the rate of diffusion, Δt is the timestamp, and Ω represents the neighborhood around pixel (x, y) sets. The term ${\mathrm{\nabla }}^{2}In\left(x,y\right)$ represents the Laplacian, reflecting the second spatial derivative of image intensity, thereby indicating areas of intensity variation or noise levels. The adaptive nature of this diffusion process is further enhanced by modulating the diffusion coefficient c (x, y, t) based on the local variance within the image, intending to dynamically adjust the diffusion strength according to the local image structures. This is integrated into the framework via Eq. (4),

(4) $$c\left(x,y,t\right)=g\left(\sigma x,y\right)f\left(\parallel \mathrm{\nabla }I\parallel \right)$$where, g (σx, y) is a function of local variance, σ (x, y), tailored to adjust the diffusion in areas of varying noise levels, thereby enhancing the model’s adaptability to different disease manifestations and noise structures. The iterative updating mechanism ensures gradual denoising while maintaining the structural integrity of crucial image features. The process converges upon reaching a stable state where no further significant changes in the image occur, determined via Eq. (5),

(5) $$\mathit{l}\mathit{i}\mathit{m}\left(t\to \mathrm{\infty }\right)\parallel I\left(n+1\right)-I\left(n\right)\parallel =0$$

Upon convergence, the resultant output is a denoised image set, where noise has been substantially reduced while preserving essential diagnostic features such as edges, textures, and contrasts inherent in the agricultural plant images & samples. This output forms a pristine base for subsequent segmentation and classification processes, crucial for the accurate diagnosis of crop diseases.

The essence of this integration lies in enhancing the model’s performance through improved ambiguity handling and decision-making, which is paramount in the intricate task of segmenting diseased regions from denoised agricultural images & samples. U-Net++, at its core, is an advanced iteration of the original U-Net architecture, designed to improve segmentation performance through a series of nested, dense skip pathways and deep supervisions. The fundamental process representing U-Net++ architecture involves the iterative refinement of feature maps across different levels of the network. These feature maps, represented as Xi,j, where i and j represent the depth and layer within the network, respectively, are computed via Eq. (6),

(6) $$Xi,j=Hi,j\left(\left[X\left(i-1,j\right),X\left(i,j-1\right),...,X\left(i-1,j-1\right)\right]\right)$$where, Hi, j signifies a composite function comprising convolutional operations, activation functions, and concatenation represented by [], applied to the set of previous feature maps. The integration of fuzzy decisions into this architecture involves the implementation of fuzzy inference systems within the skip connections and at the final decision stage of the model. This necessitates the definition of membership functions for the segmentation classes, represented as μA(x) for a class A, where x represents pixel intensity or a feature vector derived from the input images & samples. The membership functions are designed to capture the uncertainty inherent in distinguishing between diseased and healthy tissue, which is mathematically formulated via Eq. (7),

(7) $$\mu A\left(x\right)=exp\left(-{\displaystyle \frac{{\left(x-cA\right)}^2}{2\sigma A^2}}\right)$$where, cA and σA are the center and spread of the Gaussian membership function within the range 0 to 255, respectively, tuned based on the characteristics of the disease signatures. The fuzzy logic rules are then applied to combine these membership values, deriving the fuzzy logic-based feature maps at each network level. For instance, if A and B are two classes representing different disease states, a rule is modeled via Eq. (8),

(8) $$IfX\left(i,j\right)isAandX\left(i,j-1\right)isB,thenY\left(i,j\right)isC$$where, Yi, j represents the output feature map influenced by fuzzy logic, and C is a class or state resulting from the combination of A and B pixels. The implications of these rules are computed using the Mamdani-type fuzzy implication functions, and are expressed via Eq. (9),

(9) $$Yi,j=min\left(\mu A\left(Xi,j\right),\mu B\left(Xi,j-1\right)\right).$$

The defuzzification process then converts the fuzzy results into an augmented set of non-fuzzy segmented image pixels. The centroid method calculates the center of the area under the curve (AUC) of the aggregated membership functions and is given via Eq. (10),

(10) $$COA={\displaystyle \frac{\sum x\times \mu C\left(x\right)}{\sum \mu C\left(x\right)}}$$where, x ranges over all possible intensity values, and μC(x) is the aggregated membership function for class C, indicating the degree to which each pixel belongs to the disease region following the application of fuzzy rules. The final segmentation map is generated by applying the defuzzification process across all pixels and layers in the network, formulated via Eq. (11),

(11) $$S=\left\{sx,y|sx,y=COA\left(Yi,j\left(x,y\right)\right),\mathrm{\forall }\left(x,y\right)\in I\right\}$$where, S represents the final segmented image, and I is the set of all pixels in the input image sets.

Next, as per Fig. 2, the MGRA, in conjunction with convolutional operations, marks a novel paradigm in the domain of feature selection, specifically engineered for enhancing disease identification in segmented images & samples. This innovative approach merges the intuitive search dynamics of the Gorilla Troops optimization algorithm with the adaptive attachment characteristic of Remora fishes, facilitating a robust yet flexible feature selection mechanism & operations. When integrated with the computational prowess of convolutional operations, this hybrid model sets new precedents in identifying optimal features crucial for accurate disease detection. Initially, the model estimates convolutional features from the segmented images via Eq. (12),

(12) $$F\left(I\right)=X=\sum _{r=1}^R\sum _{c=1}^{C}I\left(R-r,C-c\right)\ast w\left(r,c\right)+b\left(r,c\right)$$where, $I$ represents segmented image, $R,C$ represents image size, while $w\mathrm{\&}b$ are the corresponding weights & biases. Based on these features, the Gorilla Troops optimization algorithm is applied, which is fundamentally inspired by the social structure and foraging behavior of gorillas. It employs a metaphorical representation where each gorilla (agent) in the troop (population) represents a potential solution, i.e., a subset of features. The position of each gorilla in the feature space is updated based on the following equations, reflecting the gorilla’s movement towards better foraging grounds (optimal feature subsets) via Eq. (13),

(13) $$Xnew=Xbest+STOCH\times \left(Xbest-Xcurrent\right)$$where, Xnew represents the new position (feature subset), Xbest represents the position with the highest fitness (effectiveness in disease identification), Xcurrent is the current position, and STOCH is a stochastic number between 0 and 1 sets. This movement mimics the gorilla’s approach towards optimal feeding areas, analogous to the search for an optimal feature set. Next, the Remora, which is known for its ability to attach to a host for transportation, represents the adaptive aspect of the algorithm. In the context of MGRA, this behavior is modeled to enhance exploration capabilities, allowing feature selection processes to dynamically adapt based on the evolving landscape of the optimization tasks. This is mathematically represented via Eq. (14),

(14) $$Xadaptive=Xhost+f\left(d\right)\times \left(Xbest-Xhost\right)$$where, Xadaptive represents the new adaptive position influenced by the Remora, Xhost is the current host position (a selected gorilla or feature set), and f(d) is a function representing the adaptive attachment mechanism, with d indicating the distance to the best solutions. This process allows the algorithm to maintain diversity in the feature selection process, preventing premature convergence scenarios. The adaptive attachment mechanism is represented via Eq. (15),

(15) $$f\left(d\right)=\sum _{i=1}^N{\displaystyle \frac{X\left(i\right)-\mu }{N}.}$$

The convolutional operations are integrated into this framework to evaluate the fitness of each feature subset. The fitness of each feature subset, represented by a gorilla in the optimization landscape, is assessed based on its ability to contribute to disease identification operations. This involves defining a fitness function, which could be based on criteria including classification accuracy, inter-class separation levels. The fitness function, symbolically represented as J(F), is formulated via Eq. (16),

(16) $$J\left(F\right)=\alpha \times Accuracy\left(F\right)-\beta \times Redundancy\left(F\right)$$where, F is the set of convolutional features selected by this process, α and β are weights balancing the importance of accuracy and redundancy, respectively. The iterative process of MGRA involves alternating between the gorilla movement equations and the remora adaptive mechanism, coupled with the evaluation of convolutional feature subsets until convergence criteria are met. This is defined as

(17) $$\begin{array}{rl}& while\left(notconverged\right)\\ & \{UpdatepositionsusingMGRAdynamics;Evaluatefitness;UpdateXbest\}.\end{array}$$

Upon convergence, the algorithm yields a set of convolutional features optimized for disease identification, embodying the optimal balance between discriminatory power and generalizability operations.

Next, in the refinement phase of the model, the classification process is ingeniously orchestrated through an architecture inspired by LeNet, intertwined with the advanced mechanics of graph networks, thereby instituting a novel paradigm for the identification of various disease types. This fusion, termed the graph-based LeNet process (GLNP), leverages the convolutional features selected by the MGRA operations, transforming them through a graph-based framework aligned with LeNet’s foundational principles, hence catering to the complex nature of agricultural disease patterns.

The inception of GLNP begins with the structuring of convolutional features into a graph format, where each node represents a distinct feature, and the edges encode the relationships or dependencies between these features. This translation from spatial to graph-domain is formulated as $G=\left\{V,E\right\}$ where G represents the graph, V the set of vertices or nodes corresponding to MGRA-selected features, and E the set of edges representing the feature correlations between different sets. The initial feature-to-node assignment is mathematically depicted via Eq. (18),

(18) $$V=\left\{vi|vi=fi,\mathrm{\forall }i\in F\right\}$$where, fi corresponds to the i^th feature within the feature set F selected by MGRA process. Subsequently, the edge weights, symbolizing the feature correlations, are calculated using a cosine similarity and is expressed via Eq. (19),

(19) $$eij=similarity\left(vi,vj\right)={\displaystyle \frac{vi\cdot vj}{\parallel vi\parallel \parallel vj\parallel }.}$$

For each pair of nodes vi and vj in the graph. This step ensures the graph encapsulates the intricate inter-feature relationships pivotal for disease identification process. Incorporating the principles of the LeNet architecture, the GLNP advances by subjecting the graph-structured data to a series of graph convolutional layers. These layers adapt the conventional convolutional operations to the graph domain, enabling the hierarchical feature extraction from the graph structure. The operation of a graph convolutional layer is encapsulated via Eq. (20),

(20) $$H\left(l+1\right)=\sigma \left(D^{-{\displaystyle \frac{1}{2}}}{A}^{\prime }{D}^{-{\displaystyle \frac{1}{2}}}H\left(l\right)W\left(l\right)\right)$$where, H(l) and H(l+1) are the input and output node feature matrices for the lth layer, respectively, $A^{\prime }=A+IN$ (with A being the adjacency matrix of the graph, and IN the identity matrix), D is the degree matrix of A’, W(l) is the weight matrix for the lth layer, and σ represents the ReLU non-linear activation process. Post graph convolutional processing, the node features, now enriched with local and global contextual information, are pooled to form a unified graph-level representation, suitable for classification operations. This pooling operation is guided by the graph topology and is represented via Eq. (21),

(21) $$z=Max\left(H\left(L\right),G\right)$$where, z is the graph-level feature vector, H(L) is the final node feature matrix after L layers of graph convolution, and $Max$ represents the max graph pooling operation process. The culmination of GLNP is realized through the integration of this graph-level representation into a series of fully connected layers, akin to the traditional LeNet architecture, for the final disease classification process. This is mathematically depicted via Eq. (22),

(22) $$O=\sigma \left(W\left(fc\right)z+b\left(fc\right)\right)$$where O represents the output vector corresponding to different disease types, W(fc) and b(fc) represent the weights and biases of the fully connected layers, respectively, and σ signifies the softmax activation function facilitating the multiclass classification process. The graph-based LeNet process culminates with the categorization of the input features into distinct disease types, effectively harnessing the spatial-feature extraction capabilities of LeNet and the relational insights provided by graph networks.

This innovative amalgamation not only underscores the depth of feature interrelations but also enhances disease identification accuracy, leveraging the global structural properties inherent within the agricultural disease data samples. The model’s efficacy, underscored by its robustness and precision, sets a new benchmark in the domain of crop disease classification, promising significant strides toward informed and effective agricultural management. Next, the Fuzzy U-Net++ model, an innovative architecture for image segmentation, particularly augments the segmentation process by integrating fuzzy logic with the advanced capabilities of U-Net++ is shown in Fig. 3.

Discussions

Adaptive anisotropic diffusion for image denoising

Adaptive anisotropic diffusion is a fundamental preprocessing step in the proposed model aimed at improving the quality of agricultural images by reducing noise while preserving critical features such as edges. This technique modifies the traditional anisotropic diffusion method introduced by Perona, Shiota & Malik (1994) by incorporating an adaptive approach that accounts for local image characteristics. The key advantage of adaptive anisotropic diffusion is its ability to control the diffusion process based on image gradients, thereby allowing for smooth diffusion in homogeneous regions while restricting it at edges, where crucial disease-related features might be located. This is particularly important in agricultural disease diagnosis, where fine details, such as lesions or discolorations on leaves, need to be retained to ensure accurate classification. By iteratively updating the image based on local variance, the adaptive method balances noise reduction and edge preservation, resulting in cleaner, more diagnostic-friendly images that serve as input for subsequent segmentation and classification processes.

Fuzzy U-Net++ for image segmentation

The Fuzzy U-Net++ model represents a significant advancement in image segmentation, particularly suited to handling the uncertainty and complexity inherent in agricultural disease imagery. U-Net++, an extension of the original U-Net architecture, enhances the feature extraction process through nested, dense skip connections, allowing for more detailed and refined segmentation across multiple scales. The integration of fuzzy logic into this architecture introduces an innovative layer of decision-making that departs from the traditional binary classifications. Instead of assigning a pixel to a single class based on hard thresholds, fuzzy logic enables a more nuanced classification by calculating the degree of membership for each pixel across different classes. This flexibility is especially beneficial when dealing with blurred or ambiguous boundaries, such as those between healthy and diseased tissue in plants. The fuzzy decision-making process reduces the likelihood of misclassification in uncertain regions, leading to more accurate and reliable segmentation maps. These maps, which delineate the diseased areas from the healthy regions, are crucial inputs for the subsequent classification stages of the model.

Moving gorilla remora algorithm for feature selection

The MGRA is a novel approach to feature selection that enhances the model’s ability to identify the most relevant features for crop disease diagnosis. This algorithm combines the strengths of the Gorilla Troops optimization algorithm, known for its robust global search capabilities, with the adaptive behavior of Remora fish, which attach themselves to hosts and adapt their position dynamically is shown in Fig. 4. In the context of feature selection, each “gorilla” represents a potential subset of features extracted from the segmented images, while the “remora” dynamically adjusts the feature selection based on the evolving search landscape. By utilizing this hybrid optimization strategy, the MGRA effectively balances exploration and exploitation during the feature selection process, ensuring that the model focuses on the most discriminative features for disease classification. The convolutional operations integrated into MGRA further refine the feature set by extracting hierarchical patterns from the segmented regions. This combination of biological inspiration and convolutional feature extraction allows the model to achieve higher accuracy in disease identification by focusing on features that are not only relevant but also generalizable across different disease types and plant species. The adaptive nature of MGRA ensures that the model remains flexible and efficient, even as it encounters new or varied agricultural data samples. Each of these methods—adaptive anisotropic diffusion, Fuzzy U-Net++, and the Moving Gorilla Remora Algorithm—plays a pivotal role in enhancing the accuracy, precision, and robustness of the proposed model. Together, they form a cohesive and highly effective pipeline that addresses the challenges of noise, segmentation ambiguity, and optimal feature selection in crop disease diagnosis. Table 2 shows the Comparison and Justification for Hybrid Fuzzy U-Net++ Model over Standard U-Net.

Table 2:

Comparison and justification for Hybrid Fuzzy U-Net++ model over standard U-Net.

Criteria	Fuzzy U-Net++ model (Hybrid)	Standard U-Net	Discussion
Model architecture	Nested architecture with dense skip pathways, incorporating fuzzy logic	Simple encoder-decoder structure with plain skip connections	The nested architecture in Fuzzy U-Net++ enhances feature refinement through multiple layers of convolutions and skip connections, while the integration of fuzzy logic adds a level of interpretability and adaptability for uncertain boundaries.
Segmentation strategy	Uses fuzzy logic for handling uncertainty in segmentation decisions	Binary segmentation decision without uncertainty handling	Fuzzy U-Net++ applies fuzzy inference to capture the ambiguity in pixel classification, which is especially useful in agricultural disease images where boundaries are unclear or noise is present. In contrast, U-Net uses binary segmentation, which can fail in ambiguous regions.
Feature refinement	Iterative refinement through dense skip connections and deep supervision	Single level skip connection with limited feature refinement	The nested skip pathways of U-Net++ allow for more comprehensive feature propagation, refining features across multiple scales, whereas U-Net’s simpler skip connections limit the depth of feature refinement.
Handling of ambiguities	Uses fuzzy membership functions to represent the uncertainty of class labels	No explicit mechanism for dealing with uncertainty in pixel classification	The fuzzy logic layer in Fuzzy U-Net++ assigns degrees of membership to different classes, allowing for smoother and more accurate segmentation where class boundaries are not clear, unlike U-Net which directly classifies pixels into binary classes without accounting for uncertainty.
Adaptability to noisy data	Better adaptability through fuzzy decision-making and multi-scale feature maps	Poor adaptability to noise due to strict binary decisions	Fuzzy U-Net++ is inherently more adaptable to noisy data because fuzzy logic smooths the decision process across noisy regions, while U-Net tends to over-segment or miss regions when noise is present.
Denoising capability	Hybrid integration with Adaptive Anisotropic Diffusion enhances robustness against noise	Relies on pre-processing or additional models for denoising	The Fuzzy U-Net++ is part of a hybrid model that directly integrates denoising mechanisms (like Adaptive Anisotropic Diffusion), improving its capability to process noisy agricultural images without requiring separate pre-processing stages.
Decision logic	Fuzzy decision logic allows for gradation in pixel classification based on fuzzy membership functions	Hard thresholding results in binary decisions, often leading to errors in ambiguous regions	Fuzzy U-Net++ uses fuzzy sets to assign degrees of belonging to each class, making it more resilient to complex and overlapping regions in the image. In contrast, U-Net makes hard classifications that can misclassify pixels in uncertain regions.
Performance in agriculture use-cases	Improved accuracy in disease diagnosis with precise handling of disease boundaries and uncertain regions	Struggles with precise segmentation in complex agricultural images due to lack of uncertainty handling	Fuzzy U-Net++ has shown superior performance in handling the intricate and often noisy characteristics of agricultural images, making it more suited for crop disease diagnosis where boundaries are often blurred or ambiguous. U-Net lacks this precision due to its binary nature.

DOI: 10.7717/peerj-cs.2543/table-2

Benefits of proposed UNet++ process

The recommended model, Fuzzy U-Net++, is a hybrid due to its incorporation of fuzzy logic with the advanced U-Net++ architecture, which offers significant advantages in handling uncertainty, adaptability to noisy data, and refined feature extraction. The hybrid nature of the model leverages the strengths of fuzzy logic to manage ambiguity in disease boundaries, while the nested architecture of U-Net++ ensures better feature propagation and segmentation accuracy. In contrast, the standard U-Net lacks these advanced mechanisms, resulting in less precise segmentation, especially in noisy or ambiguous agricultural images & samples. This hybrid approach is strikingly different from U-Net in that it combines the strengths of fuzzy logic for decision-making with the dense, nested feature maps of U-Net++. These innovations lead to better adaptability in real-world agricultural settings, where diseases often manifest with unclear or overlapping boundaries and noisy image data. Consequently, the Fuzzy U-Net++ model provides more nuanced and accurate segmentation, significantly improving the overall performance in crop disease diagnosis.

Comparative result analysis

Our study is majorly set in the experimental setting for the advanced diagnosis of crop diseases using the application of the proposed model. It is minutely designed to validate the effectiveness and efficiencies of the model across different dimensions concerning accuracy, precision, and recall. We describe here the overall setup used in our experiments, considering the datasets applied, preprocesses conducted, parameters setting, and evaluation metrics used.

Datasets: Experimentation was performed on two most renowned agricultural datasets, namely PlantDoc and PlantVillage. The PlantDoc dataset contains 2,500 images of different plant species and diseases, with each image size ranging from 256 × 256 to 1,024 × 1,024 pixels. This dataset contains images of healthy and diseased leaves of plants under different light conditions against various backgrounds. The PlantVillage dataset has more images, with 54,306 images, classified into 38 classes based on the species of plants and types of diseases, and their pictures are standardized at a resolution of 256 × 256 pixels.

Data pre-processing: All images were resized to 256 × 256 pixels for uniformity before training. The authors applied rotations, flipping, and scaling data augmentation techniques to the dataset to reduce the overfitting of the model. The created model will be more robust to overfitting and can generalize better to unseen data. Further, normalization on each image was performed so that pixel values are scaled in the range [0,1], which also helps accelerate the model’s convergence during training.

Parameters setting: For the adaptive anisotropic diffusion step, ‘k’ was chosen equal to 30 with a time step ‘∆t’ of 0.15 and 20 iterations to achieve sufficient noise reduction while preserving essential details. In this article, the Fuzzy U-Net++ model is configured with a total of four deepest levels and a set of [32, 64, 128, 256] filters. The design of fuzzy logic rules was based on histogram analysis of the segmented regions, which gives an accurate distinction between areas that have been affected and those that are perfectly healthy.

In this regard, the Moving Gorilla Remora Algorithm feature selection was run with a population of 50 gorillas for 100 generations, while the probability function attached to a remora for attachment was set dynamically by the iteration count to ensure exploration in the early generations and exploitation in later ones. Convolutional operations within MGRA used 3 × 3 kernels with a stride of 1, seeking to get intricate features from the segmented regions. In the phase of LeNet with graph networks, the graph was constructed with nodes from features selected before and edges for the correlation of those variables, putting a threshold of 0.5 for creating an edge. The graph convolutional network had three layers, each of which was followed by a ReLU activation and a pooling operation to progressively abstract higher-level features.

Training and evaluation: For this problem, the dataset was split 80% for training and 20% for testing. The Adam optimizer was used with an initial learning rate of 0.001 and followed a learning rate schedule in which the learning rate is decayed by a factor of 0.1 every 25 epochs during the minimization process for the cross-entropy loss function. Hence, model performance was measured by using accuracy, precision, recall, the F1-score, and the AUC, which gave a comprehensive assessment of the disease classification capabilities of the model in different use-case scenarios.

Experimental results: The experiments were conducted on a computing setup with an NVIDIA RTX 3080 GPU, 32 GB RAM, and an Intel i9 processor. Each model component was trained and separately evaluated to ensure meticulous optimization before it was used in the full pipeline. Then, the integration and evaluation of the full model were run under identical settings for the sake of consistency and comparability.

What is expected to be the final proof of the model’s efficacy in classifying various crop diseases with a very high degree of accuracy, this research is arranged through a rigorous experimental setup. This article will further elucidate how adaptive anisotropic diffusion, Fuzzy UNet++, MGRA, LeNet with graph networks, and their combination have downstream implications for agricultural disease diagnosis. We evaluated our proposed model’s effectiveness rigorously on classifying various crop diseases using the PlantDoc dataset. In the light of this experimental setup, performance analysis of our model is relative to three existing methods referenced here as Moupojou et al. (2023), Hosny et al. (2023), and Masood et al. (2023). The evaluation metrics used in comparison are accuracy, precision, recall, F1 score and AUC.

Table 3 showcases the overall accuracy of the proposed model compared to other methods. The proposed model demonstrates a superior accuracy of 94.5%, which is a significant improvement over methods (Moupojou et al., 2023; Hosny et al., 2023; Masood et al., 2023). This enhancement in accuracy is attributed to the comprehensive feature extraction and robust classification capabilities introduced by the integration of adaptive anisotropic diffusion, Fuzzy UNet++, and the graph-based LeNet process. In Table 3, the precision metric is analyzed. The proposed model achieves a precision of 93.8%, reflecting its ability to minimize false positives and accurately identify disease instances & samples. This is particularly critical in agricultural applications where false positives can lead to unnecessary interventions & scenarios. The improvement over the existing methods underscores the efficacy of our feature selection and optimization strategies.

The proposed model outperforms others with a recall rate of 95.2%, indicating its effectiveness in identifying a higher proportion of actual disease instances is shown in Fig. 5. High recall is essential for early disease detection and reducing the risk of missed diagnoses, crucial for effective disease management in crops & its samples. Table 3 evaluates the F1-scores, which are a harmonious mean of precision and recall. The proposed model attains the highest F1-score of 94.5%, showcasing its balanced performance in both precision and recall. This balance is critical for practical applications where both false positives and false negatives have significant consequences in real-time use case scenarios. Table 3 details the AUC values, an aggregate measure of model performance at various threshold settings. The proposed model’s AUC of 96.8% illustrates its superior ability to differentiate between disease classes under varying conditions. This high AUC value indicates the model’s robustness and its effectiveness across different disease types and severities.

The Fig. 6, illustrate the substantial advancements achieved by the proposed model in the diagnosis of crop diseases using the PlantDoc dataset. The improvements in accuracy, precision, recall, F1-score, and AUC is attributed to the innovative combination of adaptive anisotropic diffusion for enhanced image quality, Fuzzy UNet++ for accurate segmentation, MGRA for optimal feature selection, and the incorporation of graph networks with the traditional LeNet architecture for robust classification operations. The enhancement in performance metrics not only signifies the model’s superior diagnostic capabilities but also emphasizes its potential in reducing false diagnoses and enhancing disease management strategies in agriculture. These results promise significant advancements in precision agriculture, enabling more effective and timely interventions to safeguard crop health and productivity.

Next, the evaluation of our proposed model was extensively also conducted using the PlantVillage dataset, aimed at demonstrating its efficacy in identifying and classifying a wide range of crop diseases. This section delineates a comparative analysis between our model and three established methods, represented as Noon et al. (2022), Chelloug et al. (2023), and Masood et al. (2023). The metrics used for this comparative study include accuracy, precision, recall, F1-score, and AUC, facilitating a comprehensive evaluation of the model’s performance.

Table 4 illustrates the accuracy comparisons across the different methods. The proposed model achieves a leading accuracy of 92.8%, surpassing the other methods (Noon et al., 2022; Chelloug et al., 2023; Masood et al., 2023). This superior accuracy indicates the model’s effectiveness in correctly classifying the diverse set of diseases present in the PlantVillage dataset. The improvement in accuracy is attributed to the synergistic integration of advanced image processing and deep learning techniques within our model. In Table 4, the precision metric is evaluated for different scenarios. The proposed model stands out with a precision of 91.5%, underscoring its ability to minimize false positives effectively. The precision improvement over methods (Noon et al., 2022; Chelloug et al., 2023; Masood et al., 2023) highlights the model’s refined feature selection and classification capabilities, which are crucial for ensuring the reliable diagnosis of plant diseases. Table 4 presents the recall values for the proposed model compared to the other methodologies. Our model exhibits the highest recall at 93.4%, indicative of its superior ability to identify true positive cases across the dataset samples. High recall rates are essential for comprehensive disease detection, ensuring that fewer disease instances are missed, thereby promoting more effective agricultural management practices. Table 4 compares the F1-scores, which provide a balanced measure of the model’s precision and recall.

The proposed model achieves an F1-score of 92.4%, indicating a well-balanced trade-off between precision and recall. This score, superior to those of methods (Noon et al., 2022; Chelloug et al., 2023; Masood et al., 2023), reflects the model’s overall robustness and reliability in plant disease diagnosis is shown in Fig. 7. Most important metric that measures model performance at class differentiation in several use case scenarios. Clearly, the AUC in this proposed model is 94.6%, reflecting excellent classification performance across varying thresholds. This high value of AUC underlines the fact that this model can maximize the differentiation between the different disease and health states in the samples from PlantVillage. In this background, the results produced in tables exhibit appreciable overhauling in performance by the proposed model against existing models used in analyzing and classifying various plant diseases in the PlantVillage dataset. The accuracy, precision, recall, F1-score, and AUC are improved due to the enhanced integration of methodologies that make up the model, where noise reduction is by adaptive anisotropic diffusion, accurate segmentation by fuzzy UNet++, optimal feature extraction by MGRA, and effective classification due to the amalgamation of LeNet’s principles with graph networks.

Conclusion and future scope

This article presents an integrated, advanced framework for crop disease diagnosis to mitigate existing challenges characterizing agricultural disease management today: inadequate precision, lower accuracy, and delayed response. Our methodology is supported by the novel integration of adaptive anisotropic diffusion for image denoising, Fuzzy U-Net++ for accurate segmentation of images, Moving Gorilla Remora Algorithm-linked convolutional operations in feature selection, and augmentation of the traditional LeNet architecture using graph networks for disease classification. Experiments on well-known datasets, PlantDoc and PlantVillage, show the dominance of our proposed model over existing methodologies. More importantly, the proposed model improved the precision, accuracy, recall, and specificity of classification and engendered remarkable improvements in reducing delay and AUC, making new benchmarks. These are improvements credited to the model’s prowess in reducing noise while retaining paramount image details, correct segmentation of disease areas with accuracy even under uncertainty, efficient selection of features of relevance, and frontier use of graph network data structures that enhance operations for disease classification. The successful deployment of such a framework will be instrumental in the agricultural regard due to the achievement of more accurate, timely, and efficient diagnoses of diseases and their management. It, therefore, at this level, reduces the prevalence of false positives and negatives to allow better use of resources to enhance crop yield and health.

Although the current research represents a new frontier in crop disease diagnosis, the field is replete with opportunities for further exploration and enhancement. One of these would be to further the applicability of the proposed model to more crops and diseases under different environmental conditions and through various stages of development of the disease. This would contribute to the generalization of the effectiveness of the model and ensure its applicability across global agricultural practices. Integrating this model into real-time imaging systems via drones or autonomous ground vehicles will realize real-time diagnosis and management of diseases in the field, dramatically reducing response times and greatly enhancing the model’s practical utility. We applied explainable AI techniques, which interpreted the model’s influence on the most important end-user features and patterns; this improved the model’s understanding and reliability. Hybrid models can also be developed by combining deep learning and classic agronomic knowledge, with transfer learning upon the model to other crops or diseases having less deck data, for both robustness and adaptability. There is a need to improve predictions at the output end with respect to climate data and other environmental variables, so as to allow anticipation of outbreaks based on environment and scenarios. A collaborative framework that will assist in sharing best practices, insights, and data amongst the farmers, agronomists, and researchers would enhance the impact of this model much more and hence make the adoption of practices very much easier with AI-driven agriculture operations. This model has been the most tremendous advancement in applying state-of-the-art machine learning techniques to agriculture disease diagnosis challenges and opens the way for future innovations to further revolutionize the field of precision agriculture operations.

Supplemental Information