1. Introduction
Aphanomyces root rot (ARR), caused by the oomycete
Aphanomyces euteiches Drechs. in pea (
Pisum sativum L.), results in severe root damage, thus reducing pulse quality and yield [
1]. Plants are susceptible to this disease during any stage of their growth and development. Seed treatments and fungicides are not completely effective, and the pathogen can survive in the soil for many years without a host. Once the pathogen builds up in the soil due to favorable conditions, it can cause damage to successive susceptible crops as well [
2]. Initially, the lateral roots are prone to the infection, and eventually spread to the epicotyl. The pathogen can spread up to a distance of 18 cm from the infected plant and affect nearby healthy plants [
3]. The disease may cause loss of crop up to 86% [
4]. Thus, the development of resistant cultivars is crucial to limit yield losses.
Breeding and phenotyping have assisted in developing cultivars with better resistance to diseases [
5,
6,
7,
8]. Often, the assessment of disease resistance traits (phenotypes) for a broad set of genotypes is performed by observing their visual features [
9,
10]. However, since large numbers of plant materials are evaluated during cultivar development, standard phenotyping methods can be tedious and sometimes subjective. As an alternative approach, these visual characteristics can be processed for quantitative selection of disease resistance through deep learning-based image processing techniques such as convolutional neural networks (CNNs) [
11,
12]. Phenotypic features such as the disease status, morphology, and growth dynamics can be extracted automatically by assimilating prior knowledge and expertise [
13].
Deep learning has demonstrated its potential in numerous applications of machine vision—classification, object detection, semantic segmentation, and regression tasks [
14,
15]. Numerous CNN-based deep learning models have been developed for classification purposes. A typical CNN is designed usually using the following: a convolution layer, which extracts features from the input or previous layers; a pooling layer, which generalizes the features and minimizes the size for computational performance; and a fully connected layer, which classifies an image. The convolutional layers [
16] are defined by the convolution filters, which help in transforming and highlighting the patterns in the input image. The pooling layers reduce the dimensions of the data by linking a cluster of neurons from the previous layer to a single neuron. The image classification then takes place in the fully connected layers, where the activations are processed in the form of flattened matrices.
Deep learning models have gained popularity in dealing with agricultural problems such as crop and weed species identification [
17], plant disease detection [
18], fruit counting and grading [
19], food and grain quality monitoring [
20], yield prediction [
21], and crop stress phenotyping [
22,
23]. Phenomics techniques integrated with deep learning approaches can increase the throughput of plant phenotyping. Transforming the acquired images into authentic, reliable, and wide range of phenotypic features is a key factor for the successful application of image-based tools. Numerous approaches based on CNNs have been proposed by researchers for performing image-based plant phenotyping. An open-source tool called the Deep Plant Phenomics was introduced to implement CNNs for performing several common phenotyping tasks [
24]. An accuracy of 96.88% was obtained for classification of five different mutants of Arabidopsis, and a mean absolute difference of 20.8 h was observed for age regression task (prediction of crop age, measured in hours after germination to relate it to plant maturity). A deep learning technique was used to identify the plant stress level due to nitrogen deficiency, in which the CNN outperformed machine learning algorithms and had an accuracy of approximately 75% [
25]. A digital plant phenotyping platform for early-stage drought detection and quantification in Arabidopsis was designed using deep learning and chemometrics [
26]. The researchers processed close range spectral images with deep learning techniques and validated its feasibility based on an experiment for drought stress quantification in semi-controlled environments.
In this study, a CNN based classification model, DeepARRNet, was implemented to facilitate the evaluation of resistance to ARR in pea cultivars. Visible symptoms of ARR include honey-brown discoloration of pea roots, poor lateral root growth with minimal root hairs, and wilting of lower leaves [
1]. The reliability of identification of diseases in crops and severity prediction have improved with the application of deep learning algorithms. However, acquisition of massive amounts of data is a laborious and skill-demanding task [
27]. In addition, in many situations, image data for phenotyping are often not balanced between classes, where fewer images may be available in some classes. This situation is sometimes referred to as imbalanced or unbalanced data in data analytics. In existing plant phenotyping studies that are based on deep learning approach, the model does not reflect the features of the minority class owing to an under-sampling problem. Therefore, a proper data balancing technique should be utilized to develop a robust model that can replicate the original form of the unbalanced image data.
The random resampling method has been extensively applied in other fields such as toxicology [
28], biotechnology [
29], and drug discovery [
30] to deal with unbalanced data. In a study on tomato disease detection [
31], a deep learning model was used in conjunction with generative adversarial networks (GANs) [
32] for generating synthetic images of tomato plants to increase the amount of image data. The model was able to achieve a 10-class classification accuracy of 97.1% and concluded that augmentation through GANs increases the generalizability of the model and prevents it from over-fitting problem. On a similar note, Giuffrida et al. [
33] and Espejo-Garcia et al. [
34] proposed GAN models to synthesize artificial images of Arabidopsis plants and tomato plants for augmentation purposes. In the former study, the GAN was conditioned by leaf count, generating a plant image with the specified number of leaves. The feasibility and benefits of GAN-based image augmentation for multiple-disease identification were also assessed [
35]. The deep learning model achieved an accuracy of 93.7% when trained with both real and GAN-synthesized images. Madsen et al. [
36] also applied GAN to generate images of multiple plant species seedlings using a single network for improving the performance of plant species classification models and found better results with an average recognition accuracy of 58.9% for the generated images. Nevertheless, the benefit of the GAN approach over other resampling approaches needs to be further evaluated prior to its application. Therefore, in this study, three class-balancing techniques were enforced to identify the effective technique for improving the DeepARRNet model performance to evaluate ARR disease severity in peas. The three techniques used to address class asymmetry were: (i) random oversampling with image geometry and intensity-based transformations, (ii) synthesizing artificial images for class with low sample size using GAN, and (iii) loss function with class weighted ratio.
The main contributions of the presented work are listed as follows: (i) agriculture data is often limited by small and unbalanced sample size, and the validation of different approaches and its effect on the results is critical information that may be useful to those in the agricultural domain; (ii) the applications of machine learning and/or deep leaning approaches in root sample analysis are highly sparse, though several can be found for crop and leaf samples; and (iii) disease resistance is an important trait that plant breeders need to measure, given that root phenotyping for disease resistance is still based on visual estimation, image-based approaches such as one developed in this project (RGB imaging with CNN-based approach) can be useful.
4. Discussion
Deep learning algorithms can facilitate quantifying disease resistance in crops, as in this study, where DeepARRNet was used to evaluate the ARR resistance in pea cultivars. The model was developed to provide an end-to-end assistance to classify pea roots among three ARR severity classes: resistant, intermediate, and susceptible. An overall F1-score of 0.83 was observed, although the susceptible class accuracies were low. This can be anticipated due to the unbalanced distribution of image data, especially in the underlying class, though the overall performance was acceptable.
Unbalanced classes are a common issue for the application of deep learning algorithms, especially in the agricultural domain [
41,
42,
43]. One of the major objectives in this research was to evaluate multiple class balancing approaches to mitigate the problem with unbalanced class sizes, especially since there may be some overlap between the intermediate and susceptible classes on visual characteristics. All the three approaches utilized in this study (random oversampling-based image augmentation, GAN based image augmentation, and inclusion of weighing functions during classification) improved the overall performance of DeepARRNet. Amongst these results, the GAN-based image synthesis of a susceptible class showcased a highest overall F1-score of 0.92. The GAN-based approach may be computationally intensive, depending on data size, image resolution, and GAN network. The benefits of GAN-based image synthesis in improving model performance should surpass its limitations for successful implementation. Thus, it should be noted that the significance of selecting the effective class balancing technique would depend on the characteristics of the dataset, deep learning model, and the optimization techniques adopted. Previously, Marzougui et al. [
39] adopted a CNN-based model and machine learning algorithms of selected image features to evaluate the severity of ARR infection in lentils. The generalized linear regression model resulted in an accuracy of up to 91% for classification of three disease severity classes. Many studies in the literature have dealt with similar problems using hyperspectral imagery. For instance, Nagasubramanian et al. [
44] deployed a novel CNN model that had a classification accuracy of 95.7% to identify the soil borne fungal disease charcoal rot in soybean crops using hyperspectral images.
Rebalancing the dataset can change the decision boundaries of the classification model, thus improving the classification accuracies. This increases the chances of resulting in a better performance by converting the false negatives into appropriate predictions [
29]. This will improve the recall rate of the underlying class, as observed in this study (comparing the results in
Table 3 with the performances when class-balancing was implemented, i.e., in
Table 4,
Table 5 and
Table 6). Zhou et al. [
45] reported that combining GAN with classification network improved the average recall rate by 19% for identifying five stored-grain insect species. Similarly, there was a significant improvement in accuracy (+5.2%) when GAN-generated images were used to support the training of tomato disease identification model [
46]. However, there is a risk of decrease in precision value due to misclassification of negative samples as false positives. This theoretical intuition was in par with the results of this experiment, as the precision value of susceptible class decreased when the dataset balancing was attempted. Thus, class balancing improves the decision boundary, associating positive and negative samples into positive note. This slightly reduces the precision but can boost the recall rate, hence improving the F1 score.