1 Introduction

Indian agricultural farming has an important role in the sector of crop production or crop yields. Since agriculture is the cornerstone of economic development in any country, a large portion of the global population depends on it for their nourishment. A massive amount is contributed by agriculture makes a significant contribution in India towards the GDP of the country. However, the share of agriculture in the GDP has shown a big decrease from 40% in the 1960s to just 18% now. One reason is that the domain has not modified itself as per the advancements in technology as done by other sectors. Nearly all the techniques that are in use are decades, and sometimes, centuries ancient. This has happened at a much lesser rate compared to other industries despite some modernization introduced. This paper, suggests a two-part method for the development of agricultural technology. The first section deals with an IoT implementation, whose goal is to use a variety of sensor networks together with cloud functionality for actual environmental monitoring of factors like lighting, soil moisture, temperature, etc., and the transfer of pertinent data via a modest wireless network.

This targeted data gathering might be useful in the creation of a user-friendly irrigated control application that manages irrigation tasks including crop irrigation and plant health checking. Additionally, persistent illness symptoms that harm crops and cause problems for the farmers are challenging to combat. This sector of land farming creates different employment and job opportunities to ensure benefits for a large number of people. The healthy or disease-affected condition of the plants has an essential role to play in improved earnings for the people engaged in agriculture [1]. This can be attained by constant monitoring of the plant health required at different phases of plant development for the prevention of diseases affecting the plants during their growth [2]. This presence of diseases and pests may influence the activity of evaluation of rice cultivation and is capable of reducing the yield of rice crops primarily. In recent times, the system relies on monitoring using the naked eye, which is regarded as a time-consuming process. Therefore, automated detection of plant diseases is necessary and useful for the apprehension of crop diseases to the highest extent even at the start.

Farmers have put forward many disease-tracking mechanisms at a constant duration of time with the aim of preventing plant diseases. This automated disease identification and eradication system has been presented through a combination of the benefits of the IoT agricultural land monitoring system and some Machine Learning algorithms. Precision agriculture (PA) is a recent technology, which provides complex methodologies for the optimization of farm yield. The exploitation of these complex technologies helps the achievement of economic progress in agriculture. Application for PA includes weed monitoring, crop-yielding growth, plant disease diagnosis, and the detection of plant pests. A farmer makes use of pesticides to eliminate pests, eradicate diseases, and improve crop yield. Crop diseases are becoming a concern for farmers owing to reduced output and financial losses. Therefore, illness identification and importance require emphasis for being appropriate [3]. Crop diseases can be detected in leaves, stems and flowers and other places. Leaf highlights certain aspects of natural goods and blooming throughout all of the periods across the globe [4].

For plant contaminant acknowledgment and assurance, several studies have been conducted using machine learning (ML) approaches, such as random forest, fuzzy logic, K-means system, convolutional neural networks (CNN), etc. ML techniques are consistently utilized as tools for the location of the rice plant diseases [5]. The objective behind the set tests is to locate the presence of infections in rice plants. ML counts were used for the collection of fine development, an infectious organism, which has effect on the yield in different harvests. Hence, performance increase is still essential, especially for decision-support frameworks, which assist transformation of the massive quantity of data into useful suggestion. Additionally, in [6, 7] the methods of machine learning techniques exhibit with an accurate method is utilized to find the diseases in leaf. These methods use pre-trained models to provide output of higher quality. Even with all the study done on crop disease classification and detection, improvements are still possible.

The main objective is vast variations seen in leaf size, color, shape, and positioning, have made any exact and effective automatic detection and categorization of plant illness a difficult role. The detection method is complicated by the intensity variations that occur when images of leaves are taken. This work has suggested deep learning models through the formation of an agricultural IoT system for finding a solution to these drawbacks and for improvement in yield and rice crop disease detection with IoT technologies. The IoT system’s deep learning module uses feature compensation for building the DC-GAN-MDFC-ResNet. Compared to the methods meant for the identification of agricultural rice crop diseases already in use, this framework is capable of finding the seriousness of crop diseases. In terms of actual agricultural production activities, it is highly instructive. The primary goal is to design an automated localization and classification of crop disease using the deep learning method. The significant contributions made by the projected model are shown below.

  • This work introduces DC-GAN-MDFC–ResNet model for computing the features to ensure improvements in detection and classification accuracy crop while reducing the training and testing time consumed.

  • The proposed technique accurately localizes the infected section of the rice crop owing to the reliable nature of the DC-GAN-MDFC–ResNet model that attains increased classification accuracy with respect to crop diseases owing to the capability of the ResCV Net model in dealing with the over-fitted training data of the proposed method.

The proposed technique is computationally effective for the detection of rice crop leaves disease and elaborate experiments have been performed compared with others on a typical Plant Village database, with the differences in terms of disturbances such as blurriness, luminance, high density noise, intensity variations, rotational and scaling variances to illustrate the validity of the suggested method, modern crop illness identification technologies have been used. The below Fig. 1 shows the first step we given the input of the rice crop disease dataset, then the input data is preprocessed in the second step and the final step is to involves rice crop disease detection using proposed DC-GAN-MDFC-ResNet which helps to accurate disease detection.

Fig. 1
figure 1

Flowchart of the proposed work

Full size image

The rest of the work is divided into related work in Sect. 2. Suggested deep learning methodology in Sect. 3. Efficiency in detection attained by the suggested framework is examined in “Experiment and Results. “Finally, in Sect. 5”Conclusion and future work” concludes the survey along with further future enhancements.

2 Related work

The majority of crop disease identification techniques concentrate on enhancing the accuracy of recognition on datasets that are publicly accessible while neglecting the techniques anti-interference capability, which leads to subpar recognition accuracy when applied d to actual scenes. A high-order residual convolutional neural network (HOResNet) has been recommended in [8] for accurate and successful agricultural disease identification. To increase the anti-interferenceability, the HOResNet can simultaneously exploit high level characteristics with abstract models and low-level characteristics with object details. In [9] the authors have examined the effectiveness of deep learning systems in the diagnosis of problems seen in agriculture. The study mainly uses two aspects for optimization of the convolutional neural network as the model architecture, and other as the training label. ResNet-50, InceptionV3, MobileNetV2, and other technologies are used in the basic convolutional neural network. Transfer learning is utilizes in model training to assess experimental data based on each framework’s performance.

The alternative framework's Top1 reliability is enhanced by almost 0.7%. In [10] the author presents an IoT system for the identification of crop fine-grained illness using deep learning and IoT technologies. This technology will greatly assist automatic detection of crop illnesses of any kind with the farmers getting the findings of the process of undiagnosed. In this study, a framework for fine-grained illness to find the system is proposed using a multidimensional residual neural network (MD-ResNet). In order to integrate the outcomes of multidimensional recognition, MD-ResNet builds a compensation layer that utilizes a compensation technique. This layer does detection from three aspects, including species, fine-grained illness and coarse-grained disease.

Using Rice Net, the second stage recognized the rice disease patch dataset obtained in the first two steps [11]. The four main rice diseases are rice panicle neck blast, rice leaf blast, rice false smut and rice stem blast can be identified using the symptoms of YoloX was used in the first instance to determine which rice segments were sick. To build a new rice disease patch dataset, the orginal photographs of the sickness were cropped based on the detection results. In the second step, Siamese Network was used to identify the rice disease patch datset that was obtained in the first step. Leaf spot, gray spot and rust are the three main diseases that afflict rice crop. The authors of [12] have presented a method for accurately diagnosing these illnesses that relies on K-means clustering and an improved deep learning algorithm.The sample images are first clustered using the K-means technique for the diagnosis of three illnesses, and then they are input into the enhanced deep learning technique.

According to [13] using an ensemble of several TL models improves the effectiveness of the system of diagnosing rice illness and deficiency disorder. However, in every one of these deep learning based examinations, the diseased are extraction is left to the discretion of the DL model, which doesn’t have to partition the region of interest before hand. Three rice diseases are examined in the present investigations leaf smut, brown spot and leaf blight. Included were the inadequacies in potassium (K), phosphorous (P) and nitrogen (N) in rice. In [14] employed a convolution neural network for automatic detect crop illnesses. The data set was collected from the publicly accessible AI Challenger Competition data set from 2018, with the inclusion of 27 images of diseases affecting 10 different crops. The Inception-ResNet-v2 model has been used in this survey. The residual network unit's cross layere direct edge and multi-layer convolution. On completion of the integrated convolution operation, its activation was done using the link into the ReLu function. An experiment finding shows total recognition accuracy in the model as 86.1 percent, demonstrating its effectiveness. This study has developed and used a We chat applet for the identification of agricultural illnesses and bug infestations.

A deep convolutional neural network (CNN) architecture framework (modified LeNet) for classifying rice diseases of the leaves has be presented by the authors in [15]. the tests make use of images of rice leaves that were taken from the Plant Village collection. The training of the proposed CNNs was performed for the identification of this work dissimilar class (three diseases and one healthy class). The model yielded an accuracy of 97.89%. Hassan S and Maji A K [16] the researchers have presented a fresh deep learning model that relies on the residue link and inception layer. Interpolation that may be separated based on depth was found useful for reducing the number of variables. Training and testing of three different plant disease databases has helped the suggested model has achievement of performance accuracy of 99.39% on three different plant for the rice disease data, and 76.59% for the cassava data. In this research study [17], various types of rice like Jehlum Sr-1, Mushkibudji, Sr-2 and Sr-4 were gathered from local grain sources and utilized in this study research analysis. A Deep Convolutional Neural Network based computer vision modal called “RiceNet” has been developed to improve the precision of distinguishing five distinct categories of rice grain kinds. InceptionV3 and InceptionResNetV2 models, which are pretrained architectures based on deep leaning were also used to categorize five distinct groups of rice species.

The authors of G.K.V.L Udayananda et al. [18] used convolutional Neural Network, which can be utilized. This method offers a grasp of the illness that currently impacts rice plants and the deep learning strategies that are employed to detect them. In [19] the authors introduced a CNN model, which helped identification of the disease yielding enhanced clarity and reliability with reduced training epochs. The suggested approach provided a useful method for identification of leaf transmitted illness linked to a particular agricultural zone. Increase in various region-based diseases development was found necessary owing to climatic and seasonal variations. Researchers have described a hybrid approach in [20] for rice crop leaf disease identification that combines auto encoders and convolutional neural networks. This technical paper proposes a novel method for crop disease detection on a rice crop leaf photos by employing convolutional encoder networks. A 900-image dataset, of which 600 acts as the training set and 300 as the test set, was used to get the work’s results. In this paper, three crops and five different forms of crop infestations have examined. With the use of leaf images, training in the suggested network was done in a way that allowed distinction between crop diseases. The suggested study made use of a variety of convolution filters, including 2 × 2 and 3 × 3.

In [21] introduced a novel method for implement of the categorization of the LR pictures of rice illnesses, a novel Super-Resolution (SR) technique called Wider activation for Attention techniques based on a Generative Adversarial Network (WAGAN) was used. In this technique a major block in the generator network, the discriminator network, and the adversarial loss make up the WAGAN. This work involved reconstruction of SR plant illnesses pictures from LR images utilizing a WAGAN for assessment of the effectiveness of the proposed approach in identifying plant diseases. In [22] a group of automated deep learning techniques BR-CNNs, considered as image-based networks for automatically identifying crop diseases and estimating their severity, have been proposed. They were found careful in the simultaneous detection of crop species, classification of rice crop diseases, and severity calculation of rice crop diseases. Deep convolutional neural network (CNN) techniques and the binary relevance (BR) multi-label learning algorithm enabled the effective recognition of seven crop species, ten crop disease kinds, including healthy, and three crop illness severity categories (normal, general and severe). Performance with BR-CNNs was seen as significantly better overall than that of MLP- CNNSs and LP-CNNs. The BR-CNN technique on ResNet50 delivered the highest examination quality of 86.70%, demonstrating the viability and efficiency of the networks.

In [23] the authors have suggested ResNet-18 that offered improved recognition efficiency, forever highly informative in the real agricultural production. However, the image segmentation had he problem of increased overhead, leading to the incapability of the recognition model in getting installed in an end-to-end manner. Three sub models were chosen in Ruoling Deng et al. [24] and included into the Ensemble method. The Ensemble Method reduced misdiagnosis of the crop illness by minimizing misunderstanding between the many disease kinds. Given the similarities in appearance between various forms of rice illness, an total accuracy of 91% was obtained when six types of rice diseases were diagnosed using Ensemble model. This is regarded as pretty acceptable. Zhang, Y et al. [25] this author utilizes Inception-ResNet-v2 Inception module considering its lower computing complexity than the original Inception module. They did, however, lose some information from the source image as a result of neglecting the complementary information of image features at different scales. This effort has been inspired by the urgent need for greater agricultural sustainability and production through precise accurate plant health monitoring. Our main goal is to provide a novel method for accurately detecting and classifying figure leaf illness by fusing cutting edge image processing method with support vector machines.

The dynamic clustering method used in [26, 27] is closely paired with Back-Propagation Neural Network and supported by Particle Swarm Optimization (BPNN-PSO). We utilize a Bayesian Neural Network, to implement the entire architecture, with the underlying neuron layer acquiring weights from the BPNN-PSO outputs in [28]. This project is interesting because it uses convolutional neural networks to detect Brown spot.

disease in rice paddies early for the first time. This image recognition and preprocessing has has been used for training and testing data that was manually gathered from rice fields. It is based on real time data. But because they need large, cumbersome sensors and accurate instruments, they are inefficient and expensive. In [29] collected real time data in agricultural fields using a variety of sensors, including temperature, humidity, PH level monitoring and soil moisture sensors. A single Raspberry PI 3 module (RPI3) was utilized to control all of the sensors that were deployed in different parts of the farms. You can see the camera interacting with the RPI on leaf disease. Convolutional neural network design is employed in the identification and categorization of leaf disease in [30]. Using the CROPCARE mobile application, the suggested intelligent system’s main purpose is to identify crop illnesses. To create a decision model trained over a variety of disorders, it makes use of the pretrained model Mobile Net-V2 and the super resolution convolution network (SRCNN).

The summary can be observed from the above benchmark techniques that the focus of the research carried out in diseased crop leaves Computer vision and machine learning are largely used for identification, especially the current progress made in deep learning applied in agriculture. But the techniques find scarce application in crop leaf disease detection, in which accuracy and efficiency are equally not balanced. Recent trends in DL have led to solutions that give very reliable data, and the hardware that is available allows for fast processing. But the way decisions are made could be better. At the moment, the models that are available will not perform adequately when evaluated in actual situations. Because of this, and predicated on the authors' earlier study, they came up with a new way to find plant diseases and get around the major problems with using it in real life. The majority of previous image synthesis works have either minimized the usage of compensation layers or used the traditional deep convolution architecture in the generator and discriminator of GANs. This is where the suggested work differs, t since it looked into the effectiveness of a design for GANs that combines a convolution layer with compensation.

3 Method

Figure 2 shows the total experimental steps with two stages. Data improvement, normalization, and SVD procedures are performed in the data set images through out the data processing step. The goal is to minimize the negative impact that the data set has on the training of the model. During the model training step, the data set is split into a test set, a validation set, and a training set. Model training is carried out on the training set. The validation set determines the fulfillment of the conditions. The model is kept in place of fulfilled if not, the model training parameters are tune-up to ensure fulfillment of the requirements. The test set is useful in confirming the correctness of the model. For the deep learning model of the system, this paper suggests the DC-GAN-MDFC-ResNet model, which may help diagnosis of common and serious crop illnesses and provides additional information on actual agricultural production operations and predictions of the result.

Fig. 2
figure 2

General Framework of proposed detection model

Full size image

3.1 Dataset

Plant Village dataset has been considered with analysis of 54,306 images of rice plant leaves made, spread over 38 class labels designated to them (https:// www. kaggle. com/ datasets/ mohitsingh1804/plantvillage). Every class label creates a pair of crop diseases, and the rice leaf picture alone is used in the estimation of the crop illness pair. The different image sources seen within the database itself, resulted in the taking setting, the technologies used, and the changes in the crop species, the information not instantly useful for image classification. Three distinct variations of the dataset have been employed in investigation in this work. The dataset has been in its original, color form at first; then, experiments have been conducted on a gray scale version of the dataset and, at last, all the demonstration were conducted on a version of the dataset in which the leaves have been segmented, with elimination of all additional background details that may be competent of introducing some intrinsic bias in the data. Automation of the segmentation was performed using a script adopted to show good performance on this work specific dataset. In this study, a technique based on a collection of masks produced by analyzing the brightness, brightness, and saturation mechanism of several picture segments in multiple color spaces (Lab and HSB) has been used. One among the steps involved in this processing too helped easy fixing of color casts, which has good potential in a few subsets of the dataset, thereby, eliminating one more probable bias. This series of tests have been planned for determination of the neural network in fact picking up on the "idea" of crop illnesses, or the intrinsic biases in the dataset are only learnt by it. Figure 3 illustrates the various forms of the rice crop for an arbitrarily chosen.

Fig. 3
figure 3

Sample Leaf Image from the dataset

Full size image

3.2 Data preprocessing

Considerable variations in the number of image kinds, irregular image quality, and uneven images sizes can lead to issues in image recognition. In order to find a solution to these issues, the dataset was processed prior to model training. Three crucial phases made up the process: data improvement, data normalization, and Singular value decomposition (SVD). The issue of major variation in the quantity of images across a variety of groups was sent by the first stage of data augmentation. The issue of significant difference in the number of images among the groups was resolved in the first stage of data enhancement. The biggest group contained 2473 images, while the least group included just 22 images. Model training was forward with influence from the number of images in the groups, leading to the test accuracy being reduced. Data augmentation technology extended the groups with a least number of models in the actual data set. Random cropping may lose the appropriate representation through removal of any unhealthy area of the image. As a result, rotation and horizontal flip have been used in this study to improve the dataset. In this work, groups with below 1,000 images have been extended to nearly 1,000 and with a decrease in longer groups to nearly 1,000. The influence of the number of images on classification accuracy has been eliminated through maintenance of a balance in the number of images in the different categories.

Figure 4 illustrates data augmentation. The actual image is seen going through a horizontal flip and then rotated by 90◦, 180◦, and 270◦. All to enable training of the model the images have been converted to a fixed size in the second stage of data normalization. During the operation, the picture in the dataset is normalized to 224 pixels. This size of the image has been used by several deep learning techniques. The third stage of singular value decomposition founds a solution to the issue of image quality. Noise was removed, helping in the extraction of important information from the real image. Most of the content in the certain images were seen concentrated in a smaller portion of the data, and the remainder was found superfluous. The real data was seen having inconsistent picture quality. An illustration is depicted in Fig. 5, and the singular value 0.9 was selected for an analysis of the pictures included in the data set. It adjusts the singular value correspondingly and the influence on the image is noted.

Fig. 4
figure 4

Image flip and horizonatal rotation

Full size image
Fig. 5
figure 5

Original and SVD-Processed Images

Full size image

The Singular Value Decomposition (SVD) is considered a promising transformation to the decomposition of the input matrix into three matrices. SVD of array A is the factorization of A into the give of three matrices \({A}_{mn}={U}_{mm }{D}_{mn }{V}_{nn}^{T} ,\) where the columns of U and V are orthonormal and the array D is diagonal with positive real. In many applications the data of array A is near to an array of low rank. It is shown that the singular value decomposition of A can get the matrix B of rank k which best approximates A. Singular value decomposition is defined for all matrices (rectangular or square) not similarity the more commonly used spectral decomposition in linear Algebra. The reader familiar with eigenvectors and eigenvalues required conditions on the array to ensure orthogonality of eigenvectors. The columns of V in the singular value decomposition are called the right singular vectors of A, always form an orthogonal set with no assumptions on A. The columns of U are called the left singular vectors. A simple consequence of the orthogonality is that for a square and invertible array A, the inverse of A is\(V {D}^{-1} {U}^{T}\).

3.3 Crop disease detection using DC-GAN-MDFC-ResNet

The proposed model consists of three important elements: In this study, valuable training data were produced using a deep convolution generative adversarial network (DC-GAN) for preprocessing. The data from an MDFC-ResNet dataset was utilized for providing segmentation, G Feature for extraction and proposed a crop disease detection framework which is based on DC-GAN and MDFC-ResNetused for classification is shown in Fig. 6.

Fig. 6
figure 6

Proposed frame work of Dc-GAN and MDFC

Full size image

3.3.1 DC-GAN

DC-GAN Generative adversarial nets (GANs) constitute a novel deep generative method had introduced [24]. Contrary to the conventional model, a GAN helps the implementation of two diverse networks and a technique for defensive training. GAN employs a back-propagation technique in complicated with Markov chains found not necessary and with exact and more practical samples generated [25]. The model of GAN consists of a pair of models, which include a discriminator \(Dr\) and generator \(Gr\). As the generator tries to perplex the discriminator, the discriminator's target is to differentiate among the real training data and the generated pictures to improve discriminant accuracy. To put it briefly, D and G play the subsequent game on \(V\left(Dr,Gr\right)\) is depicted in Eq. (1).

$$\underset{\mathit{Gr}}{\text{min}}\underset{\mathit{Dr}}{\text{min}}V(Dr,Gr)={\text{\rm E}}_{x\sim {p}_{data}}\left[\text{log}Dr\left(x\right)\right]+{\text{\rm E}}_{x\sim {p}_{z}\left(z\right)}\left[\text{log}Dr\left(Gr\left(z\right)\right)\right]$$
(1)

Here, \({p}_{data}\) specifies the distribution of the original data. \(Gr\) and \(Dr\) provide a discriminator model and a generator model, respectively. \(z\) Indicates the input noise, when \(x\) refers to a leaf image as actual data. There is a distinct solution, seen provided there are random functions \(Dr\) and \(Gr\), with \(Gr\) performing the recovery of the training data distribution, and \(Dr\) equivalent to \(1/2\) in every place. It has been observed to have a better performance practically for the generator for increase in \(\text{log}Dr\left(Gr\left(z\right)\right)\) rather than the minimization of \(\text{log}1-Dr\left(Gr\left(z\right)\right)\).However, GAN will have problems such as instability during training process. Compared with the original GAN, DCGAN uses convolution layer with stride instead of upsampling layer and convolution layer instead of a fully connected layer, which can play a better performance in extracting smoke image features. Almost every layer in the generator and discriminator uses the batch norm layer to normalize the output layers of the features, which speeds up the training and improves the stability of the training process. Moreover, the leaky ReLU activation function is used in the discriminator to prevent gradient sparseness. The convolutional neural network is a deep feedforward neural network that extracts features by learning the input picture layer by layer. CNN uses a convolution kernel to extract features, and it contains a three-layer structure, convolutional layers, pooling layers, and fully connected layers. Different layers have different functions. These functional layers are composed of many neurons, and each neuron connects only a part of the neurons in the adjacent layer. It reduces the complexity of the network and improves the calculation efficiency.

DC-GAN is a novel network framework [29] that depends on GAN. CNN replaces the \(Dr\) and \(Gr\) in the actual GAN, and the fully connected layer is replaced with the convolutional layer. The generator and discriminator are found symmetric. A addition of fractional stride convolutional layer has been made to the whole network in place of an up sampling layer and pooling layer to facilitate increase in the accuracy of training. The DC-GAN makes a significant improvement to the consistency of GAN training and the calibre of the results generated DC-GAN has been used for the purpose of creating the enlarged training samples in this study. In particular, the generation network, which creates a goal blood cell picture sample from a random input, uses a five-layer CNN in sequential steps, where \(Gr\) network utilizes 5-layer transferred solution, and the \(Dr\) network makes use of 5-layer solution.

3.3.2 MDFC–ResNet

The deep residual network, whose optimization was carried out as reported in the previous section, is a prerequisite for the MDFC-ResNet. The three dimensions of the MDFC-ResNet model are rice crop, illness, and disease level. Link between the three levels is provided by a compensatory layer. By returning the identification results of the rice crop disease dimensions in the form of a compensation and error correction strategy, the accuracy of rice crop level identification is improved. The initial species' size is shown by the uppermost layer. The enhanced ResNet-34 network is useful for classifying the rice crop which the image belongs. The illness dimension is the second dimension in the middle layer of MDFC-ResNet. It employs an optimized version of the ResNet-50 to identify illnesses. The bottom layer’s third dimension has produced the dimension of illness level. The dimension of disease and disease severity parameter sharing is used by ResNet-50 models to provide better-recognized performance and speeding up classification times. The probability distribution matrix of the rice crop, disease, and disease criticality is the result of a single crop disease image.

Post-three-dimensional, a compensating layer is used. The compensation layer uses three-dimensional probability matrices and the species and illness probability matrices in the form of feedback data as compensation for the probability of the acquired stage of disease. The dimensions of the results obtained from the crop and disease recognition, and disease level detection are quite diverse. The species dimension and illness dimension outcomes are “extended” depending on disease-level findings before being mixed using Eq. 2.

$$\left(\begin{array}{c}{Pr}_{z{0}{\prime}}\\ {Pr}_{z{1}{\prime}}\\ \begin{array}{c}\begin{array}{c}{Pr}_{z{2}{\prime}}\\ \vdots \end{array}\\ {Pr}_{z{59}{\prime}}\end{array}\end{array}\right)=\alpha \left(\begin{array}{c}{Pr}_{x0}\\ {Pr}_{x0}\\ \begin{array}{c}\begin{array}{c}{Pr}_{x1}\\ \vdots \end{array}\\ {Pr}_{x9}\end{array}\end{array}\right)+\beta \left(\begin{array}{c}{Pr}_{y0}\\ {Pr}_{y0}\\ \begin{array}{c}\begin{array}{c}{Pr}_{y1}\\ \vdots \end{array}\\ {Pr}_{y35}\end{array}\end{array}\right)+\left(\begin{array}{c}{Pr}_{z0}\\ {Pr}_{z1}\\ \begin{array}{c}\begin{array}{c}{Pr}_{z2}\\ \vdots \end{array}\\ {Pr}_{z59}\end{array}\end{array}\right)\left({y}_{j}\in {x}_{i}\&z\in {y}_{j}\right)$$
(2)

In the model,\({Pr}_{xi}\) refers to the probability of the ith species, \({Pr}_{yi}\) indicates the probability of the ith disease, \({Pr}_{zi}\) indicates the probability of the ith disease level, and \({Pr}_{z{i}{\prime}}\) signifies the last of the elaborate feature recognition outcome. During the process of choosing \(\alpha =1\) and \(\beta =1\) values, the best value of α is decided. Post many trials, the maximum correctness of the test set is achieved when \(\alpha = 10\) and \(\beta = 1.5\).

3.3.3 Compensation layer

A 21 matrix reflecting the likelihood of the species rice served as the identification result for the species dimension. A 4 × 1 matrix provided the identification result for the illness dimension. The matrix indicates an 8 × 1 matrix provided the result of the disease level, which typically includes different types of rice disease. The matrices now displayed a number of dimensions. Using the size of the disease level dimension as a guide, the findings from the application of the species dimension and the illness dimension were extended to an 8 × 1 matrix during the extension method, Because the species dimension increased from 2 to 8 by 1, each element was worked through numerous times before going on to the next. Similar to this, every element of the illness dimension was repeated double previous to moving on to the next element since it runs from 4 to 8 times. The three matrices were seen as equivalent in size following this expansion operation. The flow of algorithm used for crop disease in compensation layer is given below:

Algorithm 1
figure a

DC-GAN-MDFC–ResNet

Full size image

DC-GAN produces new samples, which are utilized as additional input to MDFC–ResNet and the output are, applied the final compensation layer shown in Fig. 7.

Fig. 7
figure 7

Outline of the proposed DC-GAN- MDFC–ResNet method

Full size image

4 Results and discussion

While training the method, the initialize and the optimizer played a significant role with considerable influence on the ultimate test outcomes. The residual network was generated by a deep network, and in order to reduce the risk of gradient explosion and vanishing gradient, deep networks required appropriate weight initialization.The cost varies around the lowest value when the cost gradients are extremely large. Prior to the neural network starting to train, variables were initialized through a procedure known as weight initialization. All the neural nodes without initialization were seen identical, and the weight values were found as zero. During the process of back propagation, the value of each weight gradient was determined by multiplying the input value x of the node by the gradient of the layer that came before it. In the event of the weights being same, every neural node in the neural network was updated in the same manner, and the nodes were found identical. During the training process, the neural network did not have the ability to learn meaningful information. A suitable initialize decided the initial weights, permitting efficient training for the network in. The objective of deep learning was to constantly make necessary changes to network parameters to permit them carry out several nonlinear changes of the input to efficiently match the output. It was mostly related to the procedure of determining the Cross-entropyloss function's ideal solution. Deep learning research focuses on updating parameter update techniques.

These algorithms are referred to as optimization techniques in this learning. In the discipline of deep learning, selection of an optimizer is one of a most crucial decision for the model. The use of different optimization techniques can alter training outcomes even when the information and model architecture are equivalent. Hence, the fusion of network and dataset was validated with the application of with the application of various optimizes for the choice of choosing the mainly efficient technique. Considering the progress made in deep learning in the last few years, there have been a few newly developed optimizes and initializes introduced. Therefore, the model was modified before the experiment for the choice of the suitable initialize and optimizer. The Keras framework, which relies on Tensor Flow and used in this research for a change in learning rate of this model, batch parameters and epoch, was used in a GPU context. Here pre-processing of the data set utilized in the tuning process of the technique was done through normalization of the size of the images. In this work, SVD technology has been used for the noise suppression in the image and the initialize and optimizer of the model was decided. This part, deals with focus on the performance of the work achieved with the present technique. In the DC-GAN-MDFC–ResNet was compared with popular techniques utilize in rice crop disease detection, referred to as MDFC-ResNet [10], AlexNet and ResNet-50 [9]. The investigation findings support the suggestion made in this paper, namely residual network with multidimensional feature comparison outperforms conventional frameworks. The hyperparameter values for proposed method are Batch size is 18, Learning rate 0.0026763419023990384 and Momentum is 0.34556072704304774. The total results depicted in the table below show DC-GAN-MDFC-ResNet with better performance in validation accuracy, training accuracy and test accuracy. Particularly the Validation Accuracy of the training set improved compared with previous method like AlexNet 85.25%, ResNet-50 88.65%, MDFC-ResNet 93.96% and Proposed DC-GAN-MDFC-ResNet 95.99%. Training Accuracy is improved compared with existing method like AlexNet 85.70%, ResNet-50 86.79%, MDFC-ResNet 89.82% and Propsed DC-GAN-MDFC-ResNet 91.86% and also Test Accuracy is also increased compare with existing methods like AlexNet 83.10%, ResNet-50 82.04%, MDFC-ResNet 85.22% and Propsed DC-GAN-MDFC-ResNet 88.66%. The experiment’s outcomes confirm the modal’s efficacy once more, since the DC-GAN-MDFC-ResNet proposed in this study outperforms the traditional model. The total results are shown in Table 1 and show that the DC-GAN-MDFC-ResNet model performs best on all data sets, including the training verification and test sets. The reason for this is that crop disease recognition substantially benefits from the rice crop characteristics that DC-GAN-MDFC-ResNet is able to extract from the images of crop leaf disease.

Table 1 Shows the comparison of Training accuracy
Full size table

The performance of training Accuracy is shown in the below Fig. 8.

Fig. 8
figure 8

Comparison of Training Accuracy

Full size image

The qualitative results seen with the use of the proposed model were compared with those of previous method Precision, Recall and Fl Measures in the case of Average, Minimum and Maximum value.

Below Fig. 9 shows a Comparison of Precision, Recall and Fl with Existing Methods.

Fig. 9
figure 9

Comparison of Precision, Recall and Fl with Existing Methods

Full size image

The DC-GAN-MDFC-ResNet task was compare to the best task of the previous three models, in terms of recall, precision and F1. For every network and static, the maximum, average and minimum scores were verified. The DC-GAN-MDFC-ResNet average was used to determine which of the three other networks had the highest average. The same procedure was performed for the maximum and minimum and maximum scores. All networks with perfect 100% scores thus, were considered good. Table 2 depicts the outcome. DC-GAN-MDFC-ResNet carried out improved as more the previous methods on every other measure with the utmost average accurateness, the utmost choice (minimum to maximum), and exact value for precision, recall, and F1 values. It can be seen that this is because the DC-GAN-MDFC-ResNet based crop disease identification method mainly relies on automatic extraction of crop features, and there is no unified conclusion about the quality of the features. Also, this work has effective preprocessing methods such as augmentation and SVD this could avoid overfitting problem. This method uses the powerful feature extraction ability to extract the image features of crop disease identification, and the extracted features are more comprehensive. Therefore, this method performs better in small sample image data set, which further shows that the proposed method can be better applied to image data set with small sample size.

Table 2 Depicts the Avg,Min, Max value of Current and Existing methods
Full size table

4.1 Friedman test analysis

The Friedman test obtained a significance value of Balanced Accuracy, Sensitivity and Specificity measures, respectively. Therefore, the null hypothesis for both Balanced Accuracy and Sensitivity measures are rejected, showing that there are significant differences in the performance of the compared algorithms. Table 3 shows the ranking obtained by the Friedman test and Ranking obtained by the Friedman test. As can be seen in Table, the first algorithm in the ranking for precision, recall and F1 score was DC-GAN-MDFC–ResNet.

Table 3 Shows the value obtained on the significance of Friedman test
Full size table

4.2 Validating the proposed model in other dataset

The imagery data came from a variety of sources. The bulk of the data was collected in-field by National Paddy Crop Improvement Centre. The remainder of the data were sourced from public images found on Google Images. For this challenge, external data, other than the data provided, was prohibited. Below are a few examples of data by category viz., healthy Rice Plant, leaf rust and stem rust.

There were 16, 225 images in the data that were provided to train the AI model (2405 healthy, Bacterial Leaf Blight 648, Bacterial Leaf Streat 505, Bacterial Panicle Blight 450, Black Stem Borer 506, Blast 2351 etc.) shown in Fig. 10. The test data, on which the final performance would be measured, had 610 images and their labels were not revealed to the participants.

Fig. 10
figure 10

Sample images of disease name on different categories

Full size image

From the Fig. 9, the results shows that the proposed DC-GAN-MDFC-ResNet carried out improved as more the previous methods on every other measure with the utmost average accurateness, the utmost choice and exact value for precision, recall, and F1 values.From the explanation, the model managed to latch on to the rust portions of leaf and stem to accurately classify the category.When scaled, this approach can help in digitally monitoring crop health and could lead to significant improvement in the agriculture productivity and yield.

5 Conclusions and Future work

In this study, agricultural rice crop disease identification systems can be applied directly too.

many types of crops by fusing deep learning with IoT technologies. To differentiates various level of rice crop disease, other than identify the disease. Using sensors, weather, soil, and air quality data can be aggregated with the aim of improving the accuracy achieved with crop disease identification. The limitations of DC-GAN-MDFC-ResNet are desirable for several applications in agriculture. Most importantly, accurateness is high, even evaluate with human recognition. In future we will focus on image quality, which might be an issue in some situations where we can define standards and specifications for images in the dataset collection. Pre-trained models can assist address this problem. Consequently, a significant number of images are still needed for the GAN training in the suggested method. In case of the dataset size being tiny, it would not be capable of extracting sufficient information for the generation of new labeled images. The CNN models can be tried and the association between the real picture database size and the effectiveness in preventing and managing crop diseases.