The dataset of rice disease is selected from IP102 and icgroupcas (http://www.icgroupcas.cn/). IP102 is a dataset of crop disease and pest data and widely used in crop disease research. It contains 75,000 images of 102 crop diseases and pests, which covers a wide range of crops including rice, wheat, corn, tomato, etc. In this paper, five diseases Bacterialblight, Blast, Brownspot, Leaf smut, Tungro are selected from IP102 dataset, because they are the most prevalent rice diseases and can cause significant yield losses. Figure 1 illustrates the examples of the rice disease images and the healthy one.

In our experiments, each category of the rice pest dataset was randomly divided into the training set and the test set at the rate of 7:3. The training set contained a total of 4322 pictures, the test set contained a total of 1853 pictures, and the whole dataset contained a total of 6175 pictures. Table 1 shows the detail of the rice disease image dataset used in this paper.

The dataset used for transfer learning in this paper was selected from the PlantVillage dataset to ensure a diverse representation of plant diseases. 10 plant diseases from the PlantVillage dataset are selected, including Apple black rot, Apple healthy, Apple rust, Apple scab, Blueberry healthy, Cherry healthy, Cherry powdery mildew, Corn common rust, Corn gray leaf spot, and Corn healthy. The details of dataset for transfer learning are shown in Table 2 . In our experiment, each category of the disease dataset was randomly divided into a training set and a test set at the rate of 7:3. The training set contained 8143 images, the test set contained 3867 images, and the whole dataset contained 12010 images in total.

By including a wide range of plant diseases, the effectiveness of the proposed transfer learning method in addressing the challenge of limited training data for specific plant diseases can be evaluated. Furthermore, this diverse dataset can be used for the developing a more robust and generalized model, which can be applied to a wider range of plant disease diagnosis and monitoring tasks. Figure 2 is an illustration of 10 plant disease images in the PlantVillage dataset for transfer learning in our experiments.

To ensure uniformity across the dataset and elevate the quality of input data, we adhered to the following pivotal procedures:

In this paper, the SENet neural network was used, which is an improved version of the ResNet50 architecture (Joshy and Rajan, 2023; Zhu et al., 2023). The ResNet50 neural network is characterized by consisting of an initial convolutional layer, four Layer layers, two pooling layers, and one softmax activation layer. Each Layer layer is composed of multiple residual blocks, enhancing the network’s ability to effectively propagate gradients and learn complex features.

SENet incorporates a Squeeze-and-Excitation (SE) module after each residual block of ResNet50, which enhances the network’s capacity for selective feature emphasis. This module is designed to selectively weight the feature maps produced by the previous layer based on their importance for classification. Therefore, SENet allows for more focused attention on pertinent features in the input image, significantly improving performance in fine-grained image recognition tasks.

The choice of SENet for rice disease phenotype recognition is motivated by its advanced feature recalibration capability, which significantly enhances the model’s focus on relevant features. SENet’s unique Squeeze-and-Excitation block dynamically adjusts channel-wise feature responses, allowing the model to emphasize informative features while suppressing less useful ones. This is particularly effective in rice disease recognition, where distinguishing between subtle variations in disease symptoms is crucial. SENet’s ability to adaptively recalibrate feature maps leads to more accurate and robust recognition of rice disease patterns, making it an ideal choice for this application.

Table 3 provides a detailed summary of the ResNet50 architecture used in this paper, including the number of filters in each convolutional layer, the kernel size, and the stride. The final output of the ResNet50 network is input into the SENet module, culminating in a softmax probability distribution of the 10 plant subclasses in the PlantVillage dataset.

The structure of SENet is shown in Figure 3 . In SENet, each layer comprises two distinct residual blocks. The first one is referred to as residual block one, which consists of three convolutional layers with a kernel size of 1*1, and an additional convolutional.

layer with a kernel size of 3*3. The input to residual block one is a matrix with a channel number of C, length of L, and width of W.

However, in our work, we performed preprocessing on the images to resize both the length and width to W. Consequently, the input matrix dimensions of residual block one become (C,W,W). The output of residual block one is a matrix with 4C channels, and the length and width are halved, resulting in a data matrix of size (4C,W/2,W/2). Therefore, the residual block increases the number of channels to extract features. The first type of residual block structure is depicted in Figure 4 .

The residual block 2 is used to accomplish identity mapping and illustrated in Figure 5 . It is composed of two convolutional layers with a kernel size of 1*1, along with an additional convolutional layer with a kernel size of 3*3. Similar to residual block 1, the input to residual block 2 is a matrix with a channel number of C, length of L, and width of W. However, in our study, we performed preprocessing on the images to ensure that their length and width are both adjusted to W. Consequently, the input matrix dimensions of residual block 2 become (C,W,W). The output of residual block 2 remains identical to the input.

The number of convolution kernels in the convolution layer of each residual block in the model is summarized in Table 4 .

After four sets of layer, the final characteristic matrix is obtained. The last layer is the global average pooling layer, and the softmax is used as the activation function.

The softmax function is a commonly used probability distribution function in deep learning, which maps a set of output values x = [x ₁, x ₂,…, x_n ] to a probability distribution p = [p ₁, p ₂,…, p_n ]. The softmax function is given by Equation 1 :

The softmax function exponentiates each element of the input vector x and then normalizes them to ensure that their sum equals to 1. The resulting values p can be interpreted as class probabilities in multi-class classification tasks. The softmax function simplifies the computation of loss functions and the optimization of models.

Transfer learning aims to leverage knowledge acquired from previously trained tasks to solve related tasks more efficiently (Abbas et al., 2021; Rajpoot et al., 2023). In the context of rice disease recognition, this approach is particularly valuable due to the limited availability of large, annotated datasets specific to this task. This approach mitigates issues associated with overfitting when training on small datasets and reduces the time needed for training on large datasets (Fan et al., 2022; Saberi Anari, 2022).

As illustrated in Figure 6 , we first train the SENet on the PlantVillage dataset to obtain a pre-trained model. This pre-training step enables the SENet to learn general features of plant diseases, which can then be fine-tuned for the specific task of rice disease recognition. Then, we incorporate the parameters of the pre-trained model as transfer knowledge into the network training of the small dataset for the target task. This fine-tuning process tailors the model to the specific characteristics of rice diseases, enhancing its accuracy and efficiency for this application. Finally, we obtain a new model tailored to solving the target task. The integration of transfer learning with SENet thus allows for effective utilization of existing knowledge, significantly improving the performance of rice disease recognition.

By combining transfer learning with the SENet architecture, we effectively leveraged prior knowledge to enhance rice disease recognition. To fine-tune the pre-trained SENet model for the new task, we added new layers to the network and retrained the model on a small dataset of rice disease images.

The heuristic principle is adopted to determine which layers were most suitable for fine-tuning. Generally, the early layers of a network capture more generic features, such as edges and textures, while the later layers capture task-specific features. Thus, we typically fine-tune the later layers while keeping the early layers fixed. This process, known as fine-tuning, involves adjusting the parameters of the pre-trained model to the new task, enhancing its ability to handle task-specific features and improve performance on the rice disease recognition task.

In this paper, we utilized the cross-validation skill to determine the optimal fine-tuning strategy to enhance the performance of our SENet model in the task of rice disease recognition. Initially, we designed a candidate set encompassing a variety of potential fine-tuning strategies. These strategies considered the unfreezing of different numbers of layers, such as the last layer, the last two layers, the last three layers, and so on.

The attention mechanism has emerged as a pivotal technique in fine-grained image recognition, enhancing the nuances and subtleties of images (Lee et al., 2022; Hu et al., 2023; Jiang et al., 2023). The attention mechanism rapidly scans the global context, pinpoints the pertinent target regions, and dampens unrelated information. Consequently, the significant areas without necessitating auxiliary labeling information can be emphasized (Liu et al., 2022; Peng et al., 2023). Mathematically, the attention mechanism can be represented as Equation 2 :

In Equation 2, Q is the query vector and K and V stand for matrices of key and value vectors, respectively. d_k denotes the dimensionality of the keys. The attention process can be broken down into three core steps:

Gauging the similarity between the query and each key. Normalizing these similarity scores via a softmax function to derive attention weights. Taking a weighted sum of the value vectors, modulated by the aforementioned attention weights. This attention mechanism’s dynamism and adaptability lead to a marked enhancement in performance compared to conventional models. Its strength lies in its capacity to zoom into specific input components based on their relevance to the current query (Tang et al., 2023).

There are two broad types of attention mechanisms: self-attention and soft attention (Yang et al., 2023). This paper makes use of the advantage the of the soft attention mechanism with the Squeeze-and-Excitation Networks (SENet). Within our research, the soft attention mechanism hones in on regional aspects, assigning them attention weights between 0 and 1. This determinate attention profile is crafted through rigorous training, empowering the network to fine-tune feature selection based on these trained weights.

In this paper, the 10-fold cross-validation is adopted for testing the results. The dataset was divided into 10 subsets. For each iteration, one subset was used for validation while the other nine for training, cycling through all subsets. This approach reduces data-specific biases, ensuring a thorough evaluation. Importantly, we used precision as our evaluation metric, prioritizing the accuracy of positive identifications in our rice disease phenotype detection task.

The experiments were conducted on CentOS 2.1903 64-bit, included 8 vCPU cores and an NVIDIA V100 GPU with 32 GB of memory. The algorithms were implemented with Tensorflow 1.14.0, Keras 2.25, Python 3.6. GPU and CUDA 11.8.

The size of image pixel segmentation S is 64*64, the number of input channels C is RGB-3 color, the batch size of batch B is 8, the number of iterations E is 50, the discard rate P is 0.5, momentum β is 0.9, the initial learning rate η is 1e ⁻², and the fixed learning rate attenuates to 1e ⁻². We use VGG16, ResNet50, Inception as the baseline for comparison.

We evaluate the results using accuracy, Macro-Precision, Macro-recall, Macro-F ₁, and Cross entropy loss. In the following definition, TP and TN refer to correct predictions of positive and negative examples, respectively. FP and FN represent incorrect predictions, where FP is a positive example incorrectly classified as unfavorable, and FN is a negative example incorrectly classified as positive.

1. Accuracy (A) primarily measures the overall effectiveness of the model and indicates the proportion of correctly predicted samples among all samples. The equation is shown as Equation 3 :

2. Macro-Precision (P _macro) is chiefly concerned with the prediction results, specifically the average precision across all classes. It represents the proportion of samples that are correctly identified as positive, and is illustrated as Equation 4 :

3. Macro-Recall (R _macro) focuses on the true positive rate among actual positive samples. It calculates the average recall across all classes, and is given by Equation 5 :

4. Macro- F₁ -Score ( F 1 m a c r o ) is the harmonic mean between Macro-Precision and Macro-Recall, offering a balanced measure that takes both into account. The equation is defined as Equation 6 :

5. Cross Entropy Loss (L _ce) It is commonly used for classification and identification tasks. The cross-entropy function for multi-classification problems is shown in the formula, where y is the accurate probability distribution, q is the predicted probability distribution, N is the number of samples, and K is the number of label values. The equation is shown as Equation 7 :

Table 9 shows the results of the Macro-precision, Macro-recall rate, and Macro-F ₁ value, and comparisons with transfer learning and non-transfer learning on the test set. From Table 9 , we can see that our method has the best performance; the Macro-precision, Macro-recall rate, and Macro-F ₁ are 0.9012 respectively. This is due to the introduction of the SE module with an increased attention mechanism to enhance the recognition ability of the model for fine-grained rice disease images and the use of transfer learning methods to reduce the over-fitting of the model due to the need for more sample data.

The TRiP system is implemented based on PHP, MySQL and Redis, which can be used and deployed for edge computing. PHP-based RESTful APIs are adopted to regulate the client-server communications. The lightweight SQL server MariaDB is used for storing datasets of different formats, including images and phenotypic traits.

As illustrated in Figure 11 , the architecture of TRiP includes: 1)Web interface: Users can upload rice phenotype images through the web interface and view recognition results on the page; 2) Middleware: The middleware receives the uploaded images and passes them to the core module for recognition, and receives the model parameters provided by the user for model deployment; 3)Core: The core module uses a pre-trained SNet model for feature extraction and classification of the uploaded images. It can also use some pre-processing techniques to enhance the quality of the images. Task scheduling and task retrieval are performed; 4)Database: The database is responsible for storing the uploaded images and their corresponding recognition results. Keeps the relevant parameters of the model and the model performance.

The basic idea of TRiP is Machine Learning as a Service (MLaaS), which utilizes machine learning algorithms and models as cloud services to tackle the challenging problem of rice disease classification. TRiP enables developers to build, deploy, manage machine learning models and performance monitoring based on the cloud infrastructure.

The system overview of TRiP is shown in Figure 12 . The processing steps include: Firstly, the rice disease image samples are collected and annotated, with accurate annotations based on the knowledge of domain experts. Then, the obtained images are processed using image processing techniques, including grayscale conversion, image filtering, image sharpening, and resizing.

The data augmentation methods are used to increase the training dataset, such as random rotation, flipping, and translation. Subsequently, the sample images are input into the TRiP system for training. The model is pre-trained on the PlantVillage dataset firstly, then the transfer learning technique is employed and the model is migrated to the IP102 dataset for further optimization. After training and parameters fine-tuning, the model is used to predict the disease categories of rice disease.

The rice disease recognition system is developed using microservices architecture, such as Amazon Web Services (AWS). Microservices architecture is a loosely coupled architecture that splits a large application into a set of small and independent services. Each service can be developed, tested, deployed, and extended separately. The Microservice model structure of TRiP is shown in Figure 13 .

The user interface of TRiP is illustrated in Figure 14 . The interface and interactive design of TRiP with navigation cues, succinct layouts, and pertinent prompts messages for easy uasge.

Furthermore, TRiP’s adaptability across multiple devices and can be used in the field or an office setting. This cross-device compatibility, refined for desktops, tablets, and mobile phones, guarantees a uniform experience for different users.