Synergizing Convolutional Neural Networks and Big Transfer Learning for Enhanced Rice Disease Identification

Over the next three decades, the global population is projected to increase by approximately 2 billion individuals, going from the current 7.7 billion to 9.7 billion by 2050. The projected population peak around 2100 is estimated to approach nearly 11 billion individuals [1]. Rice, being a staple grain for much of the world's population, serves as a vital source of calories for over half of the people on Earth [2]. The significant increase in rice productivity, attributed to the adoption of using enhanced cultivars, fertilization methods, and irrigation techniques during the Green Revolution, played a crucial role in boosting production and causing a sustained decrease in rice prices. This development has been a key driver in alleviating the reduction of poverty in Asia over the preceding centuries. In spite of these prior accomplishments, the ongoing growth of rice productivity remains critical for various reasons. The pace of rice harvest improvement has considerably experienced a deceleration in the recent years, failing to retain pace in tandem with the expansion of the population. This shortfall has resulted in shortages and elevated prices, particularly impacting impoverished communities. The 2008 food crisis and the spike in rice prices underscored the vulnerability of food security. Despite substantial economic growth in many regions, the stability of food security remains somewhat precarious [3]. Plant diseases constitute a significant hindrance and play a crucial role in realizing the maximum potential yield. Rice, being the most vital global food crop, faces threats from various diseases caused by fungi, bacteria, and viruses. Changes in agronomy techniques, diminished cultivar variation leading to a slender inherited foundation, and observable weather-related shifts have altered the changing patterns of rice ailments throughout the duration. The primary diseases have increasingly developed potent and expanded to novel regions. Numerous ailments previously deemed minor have now gained economic significance in various areas [4].

Timely disease detection is crucial for optimizing agricultural output. Innovation is imperative in addressing this need, and the application of advanced techniques is key to more accurate solutions. Globally, Artificial Intelligence (AI) methods are being employed in agriculture to enhance the efficiency of crop health monitoring. These AI techniques surpass human capabilities, offering precision in crop management.

Agriculturists are increasingly integrating artificial intelligence and machine learning techniques to elevate the effectiveness of crop management. This involves the identification and treatment of crops affected by various diseases and pest infestations. Emerging technologies such as machine learning, computer vision, satellite imaging, artificial intelligence, and data analysis play pivotal roles in disease management across a wide array of crops. Their adoption signifies a paradigm shift in agricultural practices, ensuring enhanced accuracy and efficacy in disease detection and mitigation [5]. It is evident that models based on machine learning hold significant promise in the agricultural sector, particularly in the realm of detecting plant and crop diseases. Nonetheless, several challenges persist, including the dilemma encountered during model training with extensive datasets. While utilizing large datasets can undoubtedly enhance the precision of predicting disease models which comes at the expense of increased computational demands. In particular, computational models or edge devices, which are typically smaller and less computationally intensive, struggle to achieve optimal performance under these circumstances. This challenge is compounded by the issue of covariate shift, stemming from disparities between the distributions of the training data used to develop the model and the data on which the model is subsequently applied. Addressing these challenges is essential to ensure the effective integration of machine learning in agriculture and to strike a balance between accuracy and computational efficiency [6].

This research paper tackles the aforementioned challenge by integrating a CNN as a more scalable approach to image classification, alongside BiT. BiT embodies a collection of image models that have been pre-learned to exhibit exceptional adaptability, delivering outstanding performance on new datasets, even when limited examples are available per class. The integrated approach not only yields significantly improved results across diverse tasks but also demonstrates robust transferability, proving effective even when confronted with datasets containing only a sparse number of images per class.

BiT, developed by Google Research, is a comprehensive transfer learning framework designed to offer a robust and versatile feature extractor. It can be fine-tuned for a range of downstream tasks. However, the computational demands of advanced image processing algorithms, particularly for tasks like leaf detection on large datasets or real-time monitoring applications, can be substantial. This could pose challenges in resource-limited environments or for tasks requiring swift processing. Thanks to its thorough pre-training, BiT demonstrates strong generalization capabilities to new tasks even with limited data. Even when provided with a small number of images, BiT maintains good classification performance. The fundamental concept of BiT involves pre-training a model on a comprehensive and varied dataset, such as ImageNet, employing self-supervised learning methods. Self-supervised learning entails training the model to predict specific data properties without relying on explicit human annotations. Through pre-training in this fashion, the model acquires complex representations of the input data, facilitating their transfer to other tasks with minimal fine-tuning. Therefore, in this study, we opt for BiT for rice disease detection with limited datasets. BiT does have certain drawbacks. However, when paired with CNN, it has the potential to alleviate or counteract some of these disadvantages.

1. BiT focuses on learning high-level, abstract representations, while combining it with CNNs enhances the capture of fine-grained visual details. This fusion has the potential to produce more robust and comprehensive representations of the input data.
2. In contexts with limited labeled data, this integration could expedite fine-tuning processes, as CNNs excel at extracting task-specific features from restricted datasets.
3. CNNs are frequently employed in domain adaptation scenarios, where the target domain diverges substantially from the source domain. Integrating a CNN model with BiT can bolster the model's capacity to adapt to new domains or datasets by harnessing the domain-specific features extracted by the CNN.

One of the drawbacks associated with CNNs is their reliance on the Spatial Invariance Assumption.

4. The spatial invariance assumption within CNNs suggests that CNNs treat all regions of an image equally, regardless of their spatial position. In contrast, BiT captures extensive long-range dependencies and global context within the data, potentially mitigating the spatial invariance assumption inherent in CNNs.

So synergizing CNN with BiT innovatively for rice leaf disease detection with a limited dataset can lead to a more robust solution. By combining the strengths of both models, we can capitalize on the feature extraction capabilities of CNNs while harnessing the generalization power of BiT's extensive pre-training. This approach allows us to overcome the limitations of each model individually. Additionally, by leveraging the complementary aspects of CNNs and BiT, we can boost the model's capability to generalize to novel and data that has not been observed before, ultimately improving accuracy and reliability of rice leaf disease detection even with limited dataset sizes.

Incorporating a Big Transfer (BiT) model into a machine learning pipeline alongside a CNN also offers numerous advantages, particularly in the realm of image processing and classification tasks such as:

1. Utilizing Pre-Trained Features: BiT models undergo pre-training on extensive and diverse datasets, endowing them with a robust and versatile feature extraction capability. Through integration with a CNN, this combines the benefits of tailored feature extraction with the sophisticated, generalized feature comprehension inherent in the BiT model.
2. Improved Performance on Limited Datasets: In scenarios where training data is restricted, a model developed entirely from scratch may encounter challenges in acquiring effective features and achieving generalization. Nevertheless, the pre-trained elements of the BiT model offer a remedy by supplying a comprehensive, pre-learned set of features. Subsequently, the CNN layers can fine-tune these features tailored to the specific task, resulting in superior performance compared to a model trained solely on the constrained dataset.
3. Enhanced Generalization: The synergy between CNN and BiT enhances the model's skill to generalize effectively to novel, data that has not been observed before. The CNN layers adapt to the intricacies of the current task, and concurrently, the BiT model, with its comprehensive pre-training, contributes a broader understanding of image features. This collaborative approach results in heightened accuracy and robustness in real-world applications.
4. Streamlined Training Process: Employing a pre-trained BiT model reduces the computational costs and time typically associated with training a model entirely from the ground up. Subsequently, the CNN layers demand comparatively less data and training time for fine-tuning, rendering the overall process more efficient.
5. Mitigated Overfitting: In situations where the dataset is limited or lacks diversity, training a model from the ground up can frequently result in overfitting, wherein the model becomes excessively attuned to the training data but struggles with performance on new data. The incorporation of a pre-trained BiT model serves to alleviate this risk by infusing more generalized learning, enhancing the model's overall robustness.

The following sections of this document are structured as follows: Section 2 contains a review of relevant literature, exploring various research documents applying deep learning-based methods to improve the precision of identifying rice diseases. In Section 3, we outline the methodology, with subsection 3.1 detailing the dataset and 3.2 providing a comprehensive outline of the recommended approach for detecting crop ailments. Section 4 discusses experimental analyses, where thorough experiments are carried out and results are assessed through a comparative examination. This section includes an analysis of computational complexity and cost in 4.1, followed by a comparison of CNN_BiT with other deep learning models in 4.2. To conclude, Section 5 presents the paper's conclusion.

3.1 Dataset

This paper addresses five significant rice diseases that pose considerable challenges to rice farmers worldwide, as they have the potential to substantially diminish both the yields and quality of rice crops.

1. Rice Blast: Rice blast, a fungal disease caused by Magnaporthe oryzae, poses a significant threat to rice cultivation globally. It can result in substantial yield losses and jeopardize food security in affected regions. Symptoms include lesions on leaves, stems, and panicles, ultimately leading to plant death. It is characterized by the infection of rice plants through spores, leading to the development of abnormalities or marks on various parts of plants, including leaves, leaf collars, panicles, stems, and stem nodes. The infectious agent retains the ability to generate spores for more than 20 days, presenting a significant menace to rice crops susceptible to this disease [15].

2. Brown Spot: Bipolaris oryzae, the fungus responsible for brown spot, flourishes in warm and humid conditions, which heightens the vulnerability of regions with tropical and subtropical climates to outbreaks of this disease. During the winter, the fungus persists in plant debris and soil. Its spores disperse via wind, water, or human actions. It impacts rice plants, affecting different plant components including leaves, glume, Seedlings, leaf sheaths, stems, and mature plant grains. Particularly, dark coffee-colored spots emerge on the panicle, and in severe cases, spot formation may occur. on the grains, ultimately leading to diminished yield and compromised milling quality [16].

3. Bacterial Sheath Blight: Bacterial sheath blight, triggered by Burkholderia gladioli bacteria, impacts various crops like rice and maize. It induces distinct symptoms such as water-soaked lesions on leaf sheaths, with rapid spread often culminating in the formation of a mucilaginous bacterial sheath. It usually becomes evident during the heading stage of the rice plant. In mature plants, water-soaked and translucent lesions develop near the edges of the leaves. Over time, these lesions increase in size and display a wavy margin. Eventually, the affected areas turn a straw-yellow color, spreading to cover the entire leaf. If left uncontrolled, this disease can lead to decreased crop yields and compromised quality [17].

4. Rice Tungro: Rice tungro is characterized by stunted growth, yellowing of leaves, and reduced tillering in infected plants. The disease stems from the collaborative impact of two viruses transmitted by leafhoppers. The viruses are Rice tungro bacilliform virus (RTBV) and Rice tungro spherical virus (RTSV). RTBV induces symptoms of stunting and chlorosis, while RTSV enhances the transmission efficiency of RTBV by the insect vectors. The manifestation of this disease includes symptoms such as leaf discoloration, stunted growth, reduced tiller numbers, and grains that are only partially filled. Tungro disease also affects certain wild rice relatives and other grassy weeds commonly found in rice fields. The resulting yield losses can be substantial, particularly in regions where the disease is widespread [18].

5. False Smut: It is caused by the pathogen Ustilaginoidea virens infiltrates the rice spikelet over a minor aperture prior to heading. It's characterized by the formation of fungal spore masses within the rice panicle, resembling grains but actually composed of fungal structures. This disease is more prevalent in warm, humid climates, fostering ideal conditions for fungal proliferation and spore generation. Elevated nitrogen fertilizer usage and waterlogged environments can further elevate the likelihood of disease onset. In the vegetative phase of rice development, the fungus takes hold by infiltrating the tissue at the rising tips of the tillers. The effects of rice false smut predominantly influence the quality, altering the visual presentation of the crop. False smut can indeed lead to substantial reductions in yield and the quality of grains [19].

1.png

Figure 1. Comprehensive dataset: Six classes featuring healthy rice crops and five ailments affecting rice plants

In addition to these five categories of rice diseases, the dataset also incorporates images of healthy rice crops. In total, the dataset comprises six classes, with each class containing only 20 images for training purposes which is depicted in Figure 1. Current data comprising images of both diseased and healthy rice crops was gathered in the open fields of Melmaruvathur, Kavaraipettai, and Gummidipondi regions in South India during December 2021, the pictures were captured using Xiaomi and Redmi phones of 48 megapixels resolution. However, they were susceptible to noise and distortion due to the unregulated environmental conditions in the open field, along with fluctuations in lighting and backgrounds. Preserving reliability in image proportions poses a challenge, particularly when the data collection of imageries of fluctuating dimensions. To solve this issue, the training hired data expansion methods to change the size of all pictures to the specified input dimension of 224×224 for the CNN. Furthermore, image normalization was implemented to alleviate issues related to gradient propagation. Additionally, image enhancement methods such as Erosion, Dilation, Opening, and Closing were performed to augment areas of images with changing levels of intensity.

3.2 Synergizing BiT model within CNN framework for enhanced transfer learning in computer vision

The block diagram in Figure 2 illustrates the overarching process outlined in this research article. Elaboration on the procedural specifics will be provided in the subsequent subsections.

3.2.1 Convolutional Neural Network

Convolution represents a mathematical operation performed on two functions with real-valued arguments. This operation is commonly represented by an asterisk (*). It is essential that one of the functions, denoted as w, be a valid probability density function to ensure that the output represents a weighted average. Additionally, w must be zero for all negative arguments to prevent looking into the future, a capability beyond our practical reach.

In the context of CNNs, the first function x involved in the convolution is often called the input, while the second function (in this case, w) is referred to as the kernel. The result of the convolution is termed the feature map. It is represented as

$\mathrm{s}(\mathrm{t})=(\mathrm{x} * \mathrm{w})(\mathrm{t})$      (1)

Convolutional networks have the capacity to produce outputs that extend beyond traditional class labels in classification tasks or numerical values in regression tasks. Instead, they can generate high-dimensional, structured objects, typically in the form of tensors produced by standard convolutional layers. As an illustration, the model might generate a tensor S, where S_i,j,k represents the probability that the pixel at coordinates (j, k) in the input to the network belongs to class i. This capability enables the model to meticulously label each pixel in an image and create precise masks that accurately delineate the contours of individual objects [20].

2.png

Figure 2. Block diagram of proposed hybrid model framework of BiT with CNN

Fully connected layers, as their name suggests, establish extensive connections with preceding layers. Typically, these layers utilize activation functions like "sigmoid" or "softmax" in the final layer to generate predictions related to classes. Essentially, convolutional layers discern features extracted from input data, which are then condensed by successive aggregation layers. Utilizing the high-level features obtained, fully connected layers typically conduct the classification of input data into predefined categories during the final stages. Moreover, the classification layer not only categorizes the data but also extracts features crucial for both classification and detection tasks. Figure 3 illustrates the components of a typical CNN [21].

3.png

Figure 3. Components of a standard CNN layers

3.2.2 BiT-Big transfer model

Several recent advancements have been made in enhancing the training of deep neural networks. The objective is not to incorporate additional components or complexities, but rather to present a methodology that relies on the minimal number of techniques while achieving outstanding performance across various tasks. This approach is referred to as "Big Transfer" (BiT).

They conduct training on networks using three distinct sizes of data collections. The major, BiT-L, undergoes training concerning JFT-300M data collections, consisting of 300 million images with noisy labels. BiT is subsequently transferred to a variety of tasks, encompassing dimensions of data collections used for learning fluctuates from one instance per class to an over-all of one million examples. These assignments encompass a variety of datasets, including ImageNet's ILSVRC-2012, CIFAR-10/100, Oxford-IIIT Pet, Oxford Flowers-102 (including few-shot variants), and the VTAB-1k benchmark comprise 19 different datasets. Remarkably, BiT-L achieves effectiveness even when limited downstream data is available (see Table 3).

When BiT-M utilizes ILSVRC-2012 for pre-training (needed only once) and is additionally skilled on the public ImageNet-21k data collections, it delivers notable enhancements, providing cost-effective fine-tuning for subsequent tasks. BiT not only demands a concise refining practice for each freshchore nevertheless obviates the necessity for wide-ranging model fine-tuning on novel tasks. As an alternative, it employs an exploratory for configuring parameter tuning during transfer, proven effective across our diverse evaluation suite, is presented. It is crucial to highlight that the most significant aspect of BiT is its effectiveness, providing insights into the intricate relationship among dimensions, structure and learning parameters [22].

The essential elements identified for constructing a proficient network for transfer can be categorized into two groups: upstream components employed in pre-training at the upstream and transmission of downstream components for optimizing the process for an innovative task.

Pre-training at the upstream stage. The initial crucial factor is scale. In the realm of deep learning, it is widely acknowledged that larger networks exhibit superior performance in their designated tasks. Additionally, the relationship between larger datasets and the necessity for correspondingly expansive architectures is well-established. The efficacy of scale, particularly during the pre-training phase, is meticulously examined within the framework of transfer learning, extending to tasks characterized by a scarcity of data points. This analysis explores the intricate balance among the computational resources allocated (training time), the dimensions of the architecture, and the magnitude of the dataset. To explore this further, three BiT models undergo training on three significant data collections: ILSVRC-2012, which includes 1.3 million pictures (BiT-S); ImageNet-21k, comprising 14 million images (BiT-M); and JFT, boasting an impressive 300 million imageries (BiT-L).

Another crucial module involves Group Normalization (GN) and Weight Standardization (WS). Although Batch Normalization (BN) is a standard feature in the majority of advanced image models for stabilizing learning, it has been found that BN presents challenges for Big Transfer aimed at two specific motives. Firstly, during the learning of extensive models having batches of small size per device, BN demonstrates suboptimal performance or imposes synchronization costs across devices. Secondly, because of the need to update statistics in succession, BN is disadvantageous for transfer learning. In contrast, the pairing of GN and WS has shown enhancements in performance, particularly in situations where batches of small sizes are trained for ImageNet and COCO. This study showcases GN and WS proves beneficial for conducting training with sizable batch dimensions which also significantly influences the landscape of transfer learning.

Transmission to task that downstream. They suggest an economical refinement procedure that is applicable to a varied range of downstream jobs. Significantly, they have sidestepped the need for a costly exploration of tuning parameters for every original task and data volume. Instead, they experimented with a singular hyperparameter per task referred to as "BiT-HyperRule." This rule intelligently selects crucial tuning parameters based on a straightforward process with basic image resolution and the quantity of data instances.

Remarkably, during the downstream tuning phase, none of the following regularization forms were utilized: reducing weight decay to zero, reverting weight decay to initial parameters, or omitting dropout are all viable options. Remarkably, despite the significant size of the network, with BiT boasting 928 million parameters, its performance remains robust even without these techniques and their corresponding hyperparameters. This resilience persists even when transitioning to extremely small datasets. BiT adopts a streamlined training and fine-tuning framework, incorporating a select few components chosen with precision to strike a balance between complexity and performance. Leveraging extensive pre-training on a large scale is instrumental in achieving commendable performance [23].

3.2.3 Optimizinning in rice crop disease detection: integrating the BiT model into a CNN framework

Implementation utilizes TensorFlow, employing Keras as its high-level API for constructing and training deep learning models. After the datasets are stored, images undergo resizing to 384×384 pixels followed by cropping to 224×224.The dataset is partitioned within the realms of training and validation sets, with an 85% allocation for training and 15% for validation. As depicted in Figure 4, the information stream of the Hybrid CNN_BiT model is depicted. is presented where various preprocessing functions are applied to images, encompassing resizing, cropping, and normalization. Training and validation datasets are meticulously prepared, incorporating diverse data augmentation techniques such as random flipping, along with batching. This meticulous preparation optimizes the data for seamless integration into the neural network during the training process.

4.png

Figure 4. Information stream of the hybrid CNN_BiT model

The images undergo processing through a CNN block of layers. The BiT (Big Transfer) Model, denoted as BiT_model, is employed as a pre-trained model sourced from TensorFlow Hub. This model takes the output of a custom CNN as its input, with the CNN serving as a feature extractor. The resulting outputs are then channeled into the BiT model, offering the flexibility for further fine-tuning as necessary. This model features the integration of the BiT model along with a Dense layer tailored for the classification task into NUM_CLASSES categories. This deliberate integration strategy seeks to leverage the synergies between CNNs and transfer learning, capitalizing on the distinctive strengths inherent in both approaches.

The fusion of a BiT model with a CNN represents a potent alliance in tackling image classification tasks. This integration harnesses the benefits of both deep, pre-trained feature representations and tailored, task-specific feature extraction. Consequently, this methodology frequently yields models that exhibit heightened accuracy, resilience, and efficiency when confronted with diverse challenges in image classification.