To overcome across-fitting and diminishing ascent, Zhongqi Lin et al., (2019) [7] suggested a convolutional kernel matrix with parallel convolutional layers organized in a matrix with the combination of DropConnects, exponent linear unit, localized response normalization, and other techniques. As a result, non-linearity mapping will be created, and the representative ability will be improved with an acceptable increase in parameter size. Because of the substantial similarities across sub-categories, imagery of midwinter wheat leaf diseases was used as investigational samples. The matrix-built convolutional neural network surpassed AlexNet and VGG-16 with accuracy in the validation of 96.5 percent and accuracy in the testing of 90.1 percent.
Deepak Kumar et al., (2022) [8] according to this literature, the wheat disease is divided into three kinds (fungal diseases, bacterial diseases, and insect diseases). Most of the study in the literature has been discovered on wheat stripe rust disease (60.81 percent), and the most utilized prediction approach is ANN (13.32 percent). According to the findings, correctness (67 percent) is the most important performance parameter, and the largest number of articles will be published on the wheat disease by 2020. In addition, only five research employed hybrid methods, which combine SVM and NN techniques.
Zhencun Jiang et al., (2021) [9] focuses on two types of wheat leaf disease and three types of rice leaf disease, collecting and enhancing 40 illustrations of every disease. And it seeks to enhance the Visual Geometric Group Networks-16 model by using the pre-processing model on ImageNET for transformative learning and alternate learning. For rice leaf diseases, the model's accuracy is 97.22 percent, and for wheat leaf diseases, it's 98.75 percent. For greater recognition accuracy, Yuan Tian et al., (2011) [10] suggested a novel classifier system based on image recognition of wheat leaf diseases. The proposed combination model was built on stack generalization and had a 2-level structure: The first level was a support vector machines-based decision classifiers trained by feature sets generated through a data fusion method, and the second level was a support vector machines-based decision classifiers trained by feature sets generated through a data fusion method. This strategy is more versatile and has a greater identification success rate than other single classifiers and other ways of classifier ensemble for wheat leaf diseases.
Wenxia Bao et al., (2021) [11] has used stripe rust and powdery mildew as wheat leaf disease data, and this paper presents a method based on Elliptic Maxima Margins Criteria metric training learning for identifying wheat leaf diseases and their seriousness. The ideal elliptical metric matrices are constructed by applying the greatest margin criteria and the gradient rise method, which transforms the features space and reduces features correlation. The severity of stripe rust and wheat powdery mildew is then determined. The suggested approach outperforms classic support vector machines and other learning algorithms, according to the experimental findings. The suggested approach achieves a maximum accuracy of 94.16 percent for identification. Jiang Lu et al., (2017) [12] introduced an automated wheat leaf diseases diagnostic model built on multi-instance learning, which integrates wheat disease detection and disease area localization using image-levels annotation to train wheat leaf images in natural environments. Over 5-fold cross-validation, the system obtains the accuracy of 97.95 percent and 95.12 percent for two distinct architectures, namely VGG-FCN-VD16 and VGG-FCN-S, respectively.
Wenjiang Huang et al., (2014) [13] used a weighted mixture of a single band and a normalized wavelength difference of two bands to create new optimal spectral indices. The RELIEF-F method was used to extract the most and least important wavelengths for various diseases from leaf spectral data. These novel indices had classification accuracies of 86.5 percent, 85.2 percent, and 93.5 percent for healthy leaves and leaves affected with powdery mildew, yellow rust, and aphids, respectively. Jonas Franke et al., (2007) [14] conducted a spatial-temporal study of disease dynamic contrast using three high-resolution remote sensing pictures. The data was classified into areas indicating varying levels of disease severity using a decision tree based on the Normalized Difference Vegetation Index and mixture tuned matched filtering findings. The findings revealed that high-resolution multi-spectral data is typically acceptable for detecting in-field heterogeneities in crop vigor, but only somewhat suited for early identification of crop diseases. Lakshay Goyal et al., (2021) [15] described a new wheat disease classification approach in which a deep learning convolutional model is used to train the properly categorized wheat leaf diseases into ten categories. With a testing accuracy of 97.88%, the suggested approach is quite accurate. Experiments show that the suggested strategy outperforms others in terms of f-score, accuracy, and recall.
Artzai Picon et al., (2019) [16] offered numerous adaptations for early disease identification using a Deep RNN-based method to detect many plant leaf diseases under real-world acquisition situations. This study examines the performance of three important European endemic wheat diseases in terms of early detection. Davoud Ashourloo et al., (2014) [17], the purpose of this paper is to provide spectral disease indices for leaf rust identification. The spectral mixture analysis employed fractions of the different disease symptoms retrieved by the digital camera and the measured reflectance spectra of the diseased leaves as input. Then, using the least-squares approach and the spectral mixture analysis, the spectral reflectance of various illness symptoms was calculated. Tagel Aboneh et al., (2021) [18] advocated employing sophisticated deep learning methods and picture data to automate plant diseases detection. RGB picture data was acquired from the Bishoftu region for the experiment. The VGG19 model identified wheat illness with 99.38 percent accuracy based on the trial findings. Malusi Sibiya et al., (2019) [19] trained a CNN network to recognize and classify images of maize leaf diseases taken using a smartphone camera. The Neuroph Studio framework's image recognition capability allowed for the creation of a CNN with accurate feature extraction. Out of healthy leaves, the proposed model was able to distinguish three different forms of maize leaf illnesses.
Peifeng Xu et al., (2017) [20] converted a clear disease image of wheat leaf desire into a G single channel grey image of the RGB model. To extract binary feature points from a group of disease spots, the Sobel operator approach is utilized. The flood filling algorithm filters the noisy spots. Finally, the area ratio of infected patches to leaves is evaluated to provide a precise disease level diagnosis. It has been confirmed that the recognition rate is 96.2 percent, and the accuracy rate is 92.3 percent, which is about equivalent to human eyesight. Uferah Shafi et al., (2021) [21] suggested a machine-learning-based system for categorizing wheat yellow rust infection types. To begin, mobile cameras were used to capture an image dataset of various yellow rust diseases. Grayscale images from the gathered dataset were used to extract six Gray Levels Co-occurrence Matrix features and four Localized Binary Patterns texture features. CatBoost excelled on Gray Level Co-occurrence Matrix with an accuracy of 92.30 percent, according to the data. Cross-validation categorization accuracy of 93 percent was achieved by Christoph Römer et al., (2011) [22]. On the second day following inoculation, before any observable symptoms developed, this outcome was achieved. The approach presented here is useful for detecting pre-symptomatic pathogens using fluorescence spectra.
Cédric Bravo et al., (2003) [23] suggested a classification model based on exponential classification, which was built on a subset of wavebands produced by stepwise selecting features. Based on four distinct wavebands, this approach reduced confusion rates from 12 to 4% erroneous categorization. These findings are promising in terms of developing a low-cost optical device for detecting diseases like yellow rust in the field in the early spring. Davoud Ashourloo et al., (2014) [24] used a spectroradiometer in the 450 to 1000 nm wavelengths to obtain reflecting spectrum information for diseased and healthy leaves. The proportion of each disease symptom and the ratio of disease-infected area to the total area of leaf area were calculated using RGB digital pictures. The dispersion of all Spectral Vegetation Indices values rises as the severity of the disease grows.
The spectral vegetation technique outperforms the frequently used automated classification algorithm was proved in Yue Shi et al., (2017) [25]. The indices-based kernel discriminant technique provided accurate classification while preserving the pathological foundation of input data. The findings indicate that this strategy is reliable in terms of transferability and robustness in terms of detection. Uferah Shafi et al., (2022) [26] highlighted cutting-edge wheat rust detection tools including IoT, remote sensing, machine learning, and deep learning. The constraints and limits of these strategies are also examined, highlighting the practical consequences of applying them.
Mikhail A. Genaev et al., (2021) [27] discussed an approach based on the image hashing algorithm was used to decrease the degeneracy of the training data during the creation of this set. The disease-recognition method is based on the EfficientNet convolutional neural network architecture. A network with a training technique based on augmentation and transfer of imagery styles demonstrated the highest accuracy (0.942). The recognition approach was developed as a Telegram bot that allows users to assess plants in the field based on lesions. Under laboratory circumstances, E.Bauriegel et al., (2011) [28] used hyper-spectral imaging equipment to examine the wheat plant. Principal component analysis was used to distinguish spectra of sick and healthy ears tissue in various wavelength ranges. The degree of illness was properly identified (87%) using the image analysis technique Spectral Angle Mapper, assuming a visual rating error of 10%. SimpleNet was built using convolution and inverted residual blocks by Wenxia Bao et al., (2021) [29]. In this work, the Convolutional Blocks Attentions Modules, which combine spatially and channel attention mechanisms, were included into inverted residual blocks to improve the model's representation capacity for illness characteristics, allowing the model to pay attention to key aspects while suppressing unneeded features and reducing the effect of complicated backdrops in the pictures. The suggested SimpleNet model has a 94.1 percent identification accuracy on the test data set, according to the results.
At the initial stage in the growth of the wheat yellow rust illness, Dimitrios Moshou et al., (2004) [30] determined the difference in spectral reflectivity between healthful and sick leaves. A technique of normalization based on reflectivity and beam intensity modifications was created. The use of neural networks, especially multilayer perceptrons, improved the performance of classification from 95 percent to 99 percent employing a total of 5,137 wheat leaf spectral for assessment. Mohsen Azadbakht et al., (2019) [31] explored identifying wheat leaf rust at different leaf area index values and at canopy size. Various machine learning methods were developed to predict disease severity levels at the canopy scale, utilizing reflectance data collected in the situ field with a Spectro-radiometer. Alexander Koc et al., (2022) [32] developed a high-throughput plant phenotyping technique for scoring yellow rust in wheat breeding experiments by merging spectral vegetation indices data from spectral sensors in Random Forest models. Davoud Ashourloo et al., (2016) [33] used a non-imaging spectra-radiometer in the electromagnetic area to measure the spectra of infected and non-infected leaves in various diseases. The goals of this research were to use machine learning approaches to detect WLR diseases, as well as to assess the training sample size and the impact of illness symptoms on the algorithms' predictions.
Kshyanaprava Panda Panigrahi et al., (2020) [34] With the use of photographs of the plant, this work focuses on various supervised machine learning approaches for maize plant disease diagnosis. The above-mentioned categorization strategies are studied and compared to find the optimal model for plant disease prediction with maximum accuracy. In comparison to the other classification approaches, the Random Forest algorithm has the best accuracy of 79.23 percent. In-plant leaf disease classification, Ümit Atila et al., (2021) [35] introduced an EfficientNet deep learning architecture, and the performance of this model was compared to that of existing models. To train the models, the PlantVillage dataset was utilized. In the test dataset, the B5 and B4 models of the EfficientNet architecture earned the greatest values in precision and accuracy compared to other deep learning models, with 98.42% and 99.39% for precision, and 99.91 percent and 99.97 percent for accuracy, respectively. Vimal K. Shrivastava et al., (2021) [36] described a color-based image-based rice plant disease classification method. The article looked at fourteen distinct color spaces and retrieved four attributes from each one, totaling 172. Furthermore, the performances of the different classifiers were evaluated, and it was discovered that the support vector machine (SVM) classifier had the greatest classification accuracy of 94.65 percent.
R.Sujatha et al., (2021) [37], The author of this paper examined the effectiveness of several Machine Learning and Deep Learning algorithms for detecting citrus plant disease. The accuracy of illness categorization obtained through testing is outstanding since Deep Learning approaches outperform Machine Learning models in the case of diseases prediction as follows: VGG-16–89.5 percent > Inception–v3–89 percent > VGG-19–87.4 percent > SVM–87 percent > SGD–86.5 percent > RF–76.8%. From the results, it is found that random forest has the lowest classification accuracy, whereas VGG-16 has the highest classification accuracy. Aravind Krishnaswamy Rangarajan et al., (2018) [38], In this article, Tomato crop disease classification was accomplished with images from the PlantVillage dataset using neural learning architectures such as VGG16 net, and AlexNet. With 13,262 samples, VGG16 net had a classification accuracy of 97.29 percent while AlexNet had a classification accuracy of 97.49 percent. The models' performance was assessed by altering the series of images, mini-batch sizes, and weight vector learning rates. G.Sambasivam et al., (2021) [39], For the prediction of Cassava Mosaic Disease, this article focuses on approaches to reach an accuracy of 93 percent using focal loss, Synthetic Minority Over-sampling Technique, and class weight using deep neural networks.
Ramar Ahila Priyadharshini et al., (2019) [40] proposed a maize leaf disease categorization system based on a modified LeNet. Maize leaf photos from the PlantVillage collection were used in the experiment. The suggested LeNet has been trained to recognize four distinct classes, including three disorders and one healthy class. The learned model has a 97.89 percent accuracy rate. Ismail El Massi et al., (2021) [41] presented a technique for identifying illness and damage in plant leaves that is automated. The proposed system is based on the concept of combining classifiers, with two variants: the hybrid combination of three classifiers and a serial combination of two classifiers. The suggested methodology shows that using a hybrid method to detect illnesses and damages on plant leaves yielded substantial results.