Paper deep dive
Improving Fine-Grained Rice Leaf Disease Detection via Angular-Compactness Dual Loss Learning
Md. Rokon Mia, Rakib Hossain Sajib, Abdullah Al Noman, Abir Ahmed, B M Taslimul Haque
Intelligence
Status: succeeded | Model: google/gemini-3.1-flash-lite-preview | Prompt: intel-v1 | Confidence: 97%
Last extracted: 3/27/2026, 1:19:42 AM
Summary
This paper introduces a dual-loss framework for fine-grained rice leaf disease classification by combining ArcFace Loss and Center Loss. By integrating these metric learning losses with pretrained CNN backbones (InceptionNetV3, DenseNet201, and EfficientNetB0), the authors improve intra-class compactness and inter-class separability, achieving a state-of-the-art accuracy of 99.6% on the Rice Leaf Dataset.
Entities (6)
Relation Signals (3)
InceptionNetV3 → achievedaccuracy → 99.6%
confidence 100% · Our approach achieves significant performance gains, with accuracies of 99.6%, 99.2% and 99.2% respectively.
Dual-loss framework → improves → Rice Leaf Disease Detection
confidence 98% · Improving Fine-Grained Rice Leaf Disease Detection via Angular-Compactness Dual Loss Learning
InceptionNetV3 → utilizes → ArcFace Loss
confidence 95% · The proposed approach employs a transfer learning-based convolutional architecture... combining ArcFace and Center Loss
Cypher Suggestions (0)
No Cypher suggestions yet.
Abstract
Abstract:Early detection of rice leaf diseases is critical, as rice is a staple crop supporting a substantial share of the world's population. Timely identification of these diseases enables more effective intervention and significantly reduces the risk of large-scale crop losses. However, traditional deep learning models primarily rely on cross entropy loss, which often struggles with high intra-class variance and inter-class similarity, common challenges in plant pathology datasets. To tackle this, we propose a dual-loss framework that combines Center Loss and ArcFace Loss to enhance fine-grained classification of rice leaf diseases. The method is applied into three state-of-the-art backbone architectures: InceptionNetV3, DenseNet201, and EfficientNetB0 trained on the public Rice Leaf Dataset. Our approach achieves significant performance gains, with accuracies of 99.6%, 99.2% and 99.2% respectively. The results demonstrate that angular margin-based and center-based constraints substantially boost the discriminative strength of feature embeddings. In particular, the framework does not require major architectural modifications, making it efficient and practical for real-world deployment in farming environments.
Tags
Links
- Source: https://arxiv.org/abs/2603.25006v1
- Canonical: https://arxiv.org/abs/2603.25006v1
Trouble viewing inline? Open PDF directly →
Full Text
32,011 characters extracted from source content.
Expand or collapse full text
Improving Fine-Grained Rice Leaf Disease Detection via Angular-Compactness Dual Loss Learning Md. Rokon Mia1, Rakib Hossain Sajib2, Abdullah Al Noman3, Abir Ahmed4, B M Taslimul Haque5 Abstract Early detection of rice leaf diseases is critical, as rice is a staple crop supporting a substantial share of the world’s population. Timely identification of these diseases enables more effective intervention and significantly reduces the risk of large-scale crop losses. However, traditional deep learning models primarily rely on cross entropy loss, which often struggles with high intra-class variance and inter-class similarity, common challenges in plant pathology datasets. To tackle this, we propose a dual-loss framework that combines Center Loss and ArcFace Loss to enhance fine-grained classification of rice leaf diseases. The method is applied into three state-of-the-art backbone architectures: InceptionNetV3, DenseNet201, and EfficientNetB0 trained on the public Rice Leaf Dataset. Our approach achieves significant performance gains, with accuracies of 99.6%, 99.2% and 99.2% respectively. The results demonstrate that angular margin-based and center-based constraints substantially boost the discriminative strength of feature embeddings. In particular, the framework does not require major architectural modifications, making it efficient and practical for real-world deployment in farming environments. I Introduction Rice is a essential global food staple, however its yield is severely threatened by several leaf diseases (e.g., Bacterial Leaf Blight, Brown Spot, Leaf Blast, Leaf Scald, and Narrow Brown Spot), which may result in crop losses ranging from 10%10\% to over 60%60\% [13, 3]. Recent advances in deep learning have revolutionized agricultural automation, enabling highly effective image-based diagnosis of plant diseases. Convolutional Neural Networks (CNNs), including architectures such as ResNet, DenseNet, EfficientNet, VGG, and InceptionNet, have shown strong performance in these classification tasks[17, 16, 15, 7]. However, a significant challenge remains: fine-grained classification of rice leaf diseases is difficult due to subtle and visually overlapping symptoms—for instance, Bacterial Leaf Blight and Leaf Blast, or Brown Spot and Narrow Brown Spot, can exhibit highly similar visual patterns (see Fig.1). Most current deep learning methods rely primarily on the cross-entropy loss function. This function is limited because it fails to explicitly enhance intra-class compactness or inter-class separability, which are crucial properties for accurately discriminating between visually similar disease classes. Figure 1: Representative samples of rice leaf disease classes. To overcome this limitation, metric learning-based loss functions, such as Center Loss [27] (which minimizes intra-class variance) and ArcFace[6] (which maximizes inter-class separation with an angular margin penalty), have proven highly effective in domains like face recognition [14]. Triplet loss [10], particularly is frequently employed in person-reidentification and gait recognition, uses techniques such as increasing intra-class gap and decreasing inter-class gap. Despite their success, these methods remain underexplored in agricultural vision.This study introduces a dual-loss framework combining both ArcFace and Center Loss atop ImageNet-pretrained CNNs (e.g., InceptionNetV3, DenseNet201, EfficientNetB0). Our approach moves beyond conventional classification accuracy by encouraging more discriminative feature learning in angular space. This enables more reliable and robust separation of visually similar fine-grained rice diseases. The main contributions of this work are as follows: • We propose a dual-loss framework that combines ArcFace and Center Loss to improve fine-grained visual discrimination in the classification of rice disease. • We demonstrate that integrating metric learning improves the separability and robustness of CNN-based features across various pretrained backbones. • We evaluate our approach using benchmark rice disease dataset, attaining state-of-the-art outcomes with InceptionNetV3 under dual-loss supervision. • The method preserves computational efficiency, rendering it appropriate for real-time and resource-limited agricultural applications. I Literature Review Rice diseases are responsible for substantial annual yield losses, which has driven research into machine learning (ML) and deep learning (DL) methods for rice leaf disease (RLD) detection and classification. Using an SVM-based classifier, Prajapati et al. [18] reported 93.33% training accuracy, 73.33% test accuracy, and up to 88.57% with 10-fold cross-validation. Ramesh and Vydeki[20] developed a field-optimized Deep Learning (DL) model, achieving 98.9% accuracy for rice blast and over 92% for other common diseases. Hossain et al.[11] developed a lightweight CNN trained on 4,199 images for disease classification. The model achieved high performance (97.35% validation accuracy, 97.82% test accuracy, 0.99 AUC) and exceeded 93% accuracy in binary classification across five diseases, confirming its reliability and deployment readiness.Stephen et al.[24] improved feature learning in CNNs by integrating self-attention mechanisms. The most effective variant, ResNet34 with self-attention, achieved 98.54% accuracy in multi-class classification, successfully distinguishing diseases such as Brown Spot, Hispa, and Leaf Blast. Extending to other crops, Kaur et al. [12] combined EfficientNet-B7 with logistic regression-based feature down-sampling for grapevine disease detection, attaining 98.7% accuracy after 92 epochs on the PlantVillage dataset. Zhou et al.[28] proposed a combined FCM-KM clustering and Faster R-CNN framework for rice leaf disease (RLD) identification, achieving class-wise 96.71% accuracies for rice blast, 97.53% for bacterial blight, and 98.26% for blight, based on a dataset of 3,010 pictures. Abasi et al. [1] introduced a CNN-based model that outperformed transfer learning baselines (InceptionV3, EfficientNet-B2), achieving 91.4% accuracy with minimal overfitting. Chakrabarty et al. [5] employed a transformer-based ensemble combining an optimized BEiT model with pretrained CNNs, surpassing ViT, Xception, InceptionV3, DenseNet169, and ResNet50, yielding precision of 0.97, recall of 0.96, and an F1-score of 0.97. Recent work in plant disease classification has progressed, yet most models still rely on categorical cross-entropy, which fails to reduce intra-class variance or enforce strong inter-class margins—critical needs in fine-grained tasks with subtle visual differences. We address this by introducing a dual-loss framework that combines Center Loss and ArcFace with three high-performing pretrained backbones (InceptionNetV3, DenseNet201, EfficientNetB0), enhancing feature discriminability without major architectural changes and enabling scalable, real-world agricultural deployment. I Methodology Figure 2: Schematic representation of the proposed framework for plant disease classification. I-A Dataset This study employs the publicly accessible Rice Leaf Disease (RLD) dataset from the Roboflow platform [19], licensed under C BY 4.0 for ethical academic and practical use. The dataset comprises six categories: Bacterial Leaf Blight, Brown Spot, Leaf Blast, Leaf Scald, Narrow Brown Spot, and Healthy Leaves. It captures a wide range of disease symptoms under diverse environmental conditions, highlighting both intra-class variability and subtle inter-class similarities (see Fig. 1), which are characteristic challenges in rice leaf disease classification. For robust evaluation, it is stratified into training, validation, and test subsets, with class-wise distributions summarized in Table I. TABLE I: Dataset Image Distribution by Category and Split Category Train Validation Test Bacterial Leaf Blight 439 73 35 Brown Spot 522 75 46 Healthy 557 100 47 Leaf Blast 522 77 51 Leaf Scald 412 88 35 Narrow Brown Spot 521 76 36 Total 2973 489 250 I-B Data Augmentation To improve the generalization capability of the model and minimize the risk of overfitting, a structured data augmentation pipeline was implemented to the training set. All input images were converted into 299×299299× 299 pixels to match the input specifications of the pretrained backbone model. The following transformations were then applied sequentially using the torchvision.transforms library: I-B1 Random Horizontal Flip Utilized a probability of 0.5 to simulate the natural variability in leaf orientation. I-B2 Random Rotation Each image was rotated randomly within a range of ±15∘± 15 , enhancing rotational invariance. I-B3 Color Jitter Random adjustments of brightness, contrast, and saturation were applied using a factor of 0.1 to increase robustness against lighting and color variations. I-B4 Random Affine Transformation Minor spatial translations (up to 10% of image dimensions) were introduced to support spatial invariance without altering the rotation. I-B5 Normalization The images were converted to tensors and normalized according to ImageNet statistics, with mean values [0.485,0.456,0.406][0.485,0.456,0.406] and standard deviations [0.229,0.224,0.225][0.229,0.224,0.225], ensuring compatibility with the weights of the backbone. For validation and testing, a minimal preprocessing pipeline was employed to maintain consistency during evaluation. Images were converted into 299×299299× 299 pixels, converted to tensors, and normalized using the same statistical parameters as in the training phase, without any stochastic augmentations. I-C Proposed Framework The proposed approach employs a transfer learning-based convolutional architecture for fine-grained classification of rice leaf diseases. As illustrated in Fig. 2, the framework consists of three key stages: deep feature extraction, embedding transformation, and dual-loss supervised classification. We denote the input image as in∈ℝ3×H×WX_in ^3× H× W, where H=W=299H=W=299. The image is first passed through a pretrained convolutional backbone (e.g., InceptionNetV3, DenseNet201, EfficientNetB0), denoted as Backbone(⋅)Backbone(·), to obtain a deep feature map: feat=Backbone(in)∈ℝC×H′×W′.X_feat=Backbone(X_in) ^C× H × W . (1) To reduce spatial dimensions, an adaptive average pooling operation is applied: f=AvgPool(feat)∈ℝC.x_f=AvgPool(X_feat) ^C. (2) This pooled feature vector is then projected into an embedding space via a two-layer feature transformation module (MLP): 1 _1 =ReLU(W1f+1),1∈ℝ512, =ReLU(W_1x_f+b_1), _1 ^512, (3) ~1 h_1 =Dropout(1;p=0.5), =Dropout(h_1;p=5), ei e_i =ReLU(W2~1+2),ei∈ℝD, =ReLU(W_2 h_1+b_2), e_i ^D, where D=256D=256 is the final embedding dimension, and W1,W2W_1,W_2, 1,2b_1,b_2 are learnable parameters. The final classification logits are obtained via a linear classifier: zi=Wc⊤ei+c,zi∈ℝC,z_i=W_c e_i+b_c, z_i ^C, (4) where Wc∈ℝD×CW_c ^D× C and c∈ℝCb_c ^C are the classifier weights and biases, and C=6C=6 denoting the count of classes. I-D Loss Function Formulation To enforce both inter-class separability and intra-class compactness, we adopt a dual loss objective consisting of ArcFace Loss and Center Loss. The overall training objective is: ℒtotal=ℒArc+α⋅ℒCenterL_total=L_Arc+α·L_Center (5) where α∈ℝ+α ^+ are weighting coefficients. ArcFace Loss. ArcFace introduces an additive angular margin to encourage angular decision boundaries in the embedding space. First, the embeddings and classifier weights are normalized: ‖ei‖=‖k‖=1,∀i,k.\|e_i\|=\|w_k\|=1, ∀ i,k. The cosine similarity between the embedding eie_i and class weight kw_k is: zi,k=cos(θi,k)=ei⊤k.z_i,k= ( _i,k)=e_i w_k. For the ground-truth class yiy_i, the angle is modified by a margin m: zi,k′=cos(θi,k+m),if k=yicos(θi,k),otherwisez _i,k= cases ( _i,k+m),&if k=y_i\\ ( _i,k),&otherwise cases (6) The logits are then scaled by a factor s, and cross-entropy is applied: ℒarc=−1N∑i=1Nloges⋅zi,yi′∑k=1Ces⋅zi,k′L_arc=- 1N _i=1^N e^s· z _i,y_i _k=1^Ce^s· z _i,k (7) where N is the batch size. Center Loss. Center Loss seeks to reduce the intra-class variance by learning a center k∈ℝDc_k ^D for each class k, and imposing penalties on the disparity between feature embeddings and their respective centroids: ℒcenter=1N∑i=1N‖ei−yi‖22.L_center= 1N _i=1^N \|e_i-c_y_i \|_2^2. (8) The centers kc_k are updated dynamically during training to reflect the mean embeddings of their respective classes. Algorithm 1 Training Procedure with Dual-Loss Supervision 1: Input: Training dataset =(i,yi)i=1ND=\(X_i,y_i)\_i=1^N, pretrained backbone Backbone(⋅)Backbone(·), learning rate η, center-loss weights α 2: Output: Trained model parameters Θ 3: for each mini-batch (i,yi)i=1m⊂\(X_i,y_i)\_i=1^m do 4: Feature Extraction: i=Backbone(i),i∈ℝm×C×H′×W′F_i=Backbone(X_i), _i ^m× C× H × W 5: Global Pooling: f,i=AvgPool(i),f,i∈ℝm×Cx_f,i=AvgPool(F_i), _f,i ^m× C 6: Embedding Generation via MLP: ei,zi=MLP(Xf,i),ei∈ℝN×D;i∈ℝm×Ce_i,z_i=MLP(X_f,i), e_i ^N× D; _i ^m× C 7: Loss Computation: ℒtotal=ℒarc(i,yi)+α⋅ℒcenter(i,yi)L_total=L_arc(z_i,y_i)+α·L_center(e_i,y_i) 8: Parameter Update: Θ←Θ−η⋅∇Θℒtotal ← -η· _ L_total 9: end for IV Experiments IV-A Implementation Details All experiments were implemented using PyTorch 2.5.1 and Python 3.9.21. Training was performed in a GPU-accelerated Kaggle cloud environment running on an NVIDIA T4 GPU. AdamW optimizer was used to optimize the model at a learning rate of 1×10−41× 10^-4, trained over 30 epochs, and a batch size of N=32. Dropout regularization and L2 weight decay were used to prevent overfitting. The ArcFace module was configured using a scale factor s=30s=30 and angular margin m=0.5m=0.5, while the Center Loss component was weighted with α=0.5α=0.5 to balance its contribution. The complete training configuration is summarized in Table I. TABLE I: Training Configuration Parameter Value Optimizer AdamW (learning rate = 0.0001) Loss Functions ArcFace, Center Loss Batch Size 32 Epochs 30 Regularization Dropout (0.5), L2 regularization (λ=0.0001λ=0.0001) Scale Factor (s) 30 Angular Margin (m) 0.5 Center Loss Weight (α) 0.5 IV-B Evaluation Metrics To comprehensively assess model performance, we employed standard classification metrics including Accuracy, Precision, Recall, and F1-Score. The computation of these metrics was carried out utilizing the confusion matrix, which tabulates the distribution of true versus predicted labels. The formulas for these metrics are: Accuracy=TP+TNTP+TN+FP+FNAccuracy= TP+TNTP+TN+FP+FN (9) Precision=TPTP+FPPrecision= TPTP+FP (10) Recall=TPTP+FNRecall= TPTP+FN (11) F1-Score=2×Precision×RecallPrecision+RecallF1-Score=2× Precision×RecallPrecision+Recall (12) where TP=True Positive,TN=True Negative,FP=False Positive,FN=False Negative.TP=True Positive,\ TN=True Negative,\ FP=False Positive,\ FN=False Negative. TABLE I: Comparison of the proposed method with state-of-the-art models on rice leaf disease classification. Authour Year Dataset Method Accuracy (%) Catal Reise and Turk et al.[4] 2024 Wheat Leaf Dataset RegNetY080 97.64 Abasi et al.[1] 2023 Rice Leaf Images Custom CNN 91.40 Ahad et al. [2] 2023 Rice Leaf Dataset (UCI) DEX (Densenet121, EfficientNetB7, and Xception) 98.00 Sankareshwaran et al. [22] 2023 Rice Leaf Dataset CAHA-AXRNet 98.10 Naresh kumar et al.[17] 2025 Rice Leaf Dataset FVBC model 97.6 Proposed framework - Rice Leaf Dataset InceptionNetV3+Dual Loss 99.6 V Result & Discussion TABLE IV: Performance of pretrained models on Roboflow RLD using Cross Entropy vs. ArcFace + Center Loss. Model Loss Function Precision (%) Recall (%) F1 Score (%) Accuracy (%) Experiment InceptionNetV3 Cross Entropy 99.31 99.31 99.31 99.20 a ArcFace + Center Loss 99.65 99.67 99.66 99.60 b DenseNet201 Cross Entropy 99.29 98.33 99.28 98.00 c ArcFace + Center Loss 99.31 99.31 99.31 99.20 d EfficientNetB0 Cross Entropy 98.19 98.13 98.14 98.00 e ArcFace + Center Loss 99.31 99.35 99.31 99.20 f V-A Performance comparison of Pretrained Models To assess the effectiveness of the proposed loss function combination, experiments were carried out using three pretrained models: InceptionNetV3, DenseNet201, and EfficientNetB0. Each model was fine-tuned using two different loss function settings: standard Cross Entropy, and a combined ArcFace with Center Loss. The evaluation metrics such as Precision, Recall, F1 Score, and Accuracy, are summarized in Table IV. All models demonstrated high performance, with values exceeding 99% across all metrics. Overall, the integration of ArcFace and Center Loss either improved or preserved model performance when compared to the baseline Cross Entropy loss. InceptionNetV3 exhibited a consistent improvement with the combined loss, achieving higher precision (99.65% vs. 99.31%) and recall (99.67% vs. 99.31%). These improvements suggest enhanced intra-class compactness and inter-class separability enabled by the proposed loss formulation. Similarly, DenseNet201 showed a notable increase in recall (99.31% vs. 98.33%) and accuracy (99.20% vs. 98.00%), indicating that the model demonstrates improved generalization to previously unseen samples when trained with ArcFace and Center Loss. EfficientNetB0 also benefited substantially, with accuracy increasing from 98.0% to 99.2% using the proposed loss combination, significantly outperforming its Cross Entropy baseline. Taken together, these findings highlight the synergistic effect of combining ArcFace and Center Loss in optimizing both feature embeddings and classification boundaries, thereby enhancing feature discrimination and generalization across different backbone architectures. Figure 3: Performance comparison of pretrained models with Cross Entropy vs. ArcFace + Center Loss. V-B Ablation Study To analyze the influence of various loss functions on the model’s performance, we performed an ablation study using the InceptionNetV3 architecture. Table V summarizes the results obtained with four different loss configurations: standard Cross-Entropy (CE) Loss, Center Loss, ArcFace Loss, and a combination of ArcFace + Center Loss. From the results, it is evident that the the joint use of ArcFace and Center Loss attains the highest performance across all evaluation metrics, with 99.65% precision, 99.67% recall, 99.66% F1-score, and 99.60% accuracy. This indicates that combining the strengths of both losses improves feature discriminability and class separation. While the CE Loss performs well with a balanced performance (99.31% accuracy), the Center Loss alone significantly underperforms. This suggests that Center Loss by itself is insufficient for effective training, potentially due to its weak gradient signals or the need for an auxiliary classification loss. TABLE V: Performance of ablation experiment on Roboflow RLD dataset. No. ℒCEL_CE ℒCenterL_Center ℒArcL_Arc Precision Recall F1-Score Accuracy 1 ✓ 99.31 99.31 99.31 99.20 2 ✓ 19.50 38.56 25.07 38.80 3 ✓ 98.68 98.59 98.62 98.40 4 ✓ ✓ 99.65 99.67 99.66 99.60 ArcFace Loss alone delivers strong results (98.40% accuracy), confirming its ability to enhance inter-class separability through angular margin optimization. However, fusing it with Center Loss leads to further gains, validating the complementary nature of the two losses in guiding feature embedding. To show significance our proposed framework we perform another ablation experiment on another public available Kaggle Rice Leaf Dataset (RLD) dataset [23] and result present in Table VI. TABLE VI: Performance of ablation experiment on Kaggle RLD dataset. No. ℒCEL_CE ℒCenterL_Center ℒArcL_Arc Precision Recall F1-Score Accuracy 1 ✓ 97.39 97.41 97.38 97.35 2 ✓ 4.01 16.3 6.43 16.67 3 ✓ 95.37 95.08 95.16 95.08 4 ✓ ✓ 98.17 98.15 98.13 98.11 V-C Comparison with previous study We evaluate the efficacy of the proposed method against various state-of-the-art (SOTA) models reported in recent literature. As demonstrated in Table I, the proposed approach utilizing InceptionNetV3 combined with ArcFace Loss and Center Loss achieves a perfect accuracy of 99.6% on the Rice Leaf Dataset [21], surpassing existing methods. Traditional CNN-based approaches such as Abasi et al. [1] achieved modest accuracy (91.40%) due to limited feature discriminability. Notably, the closest competitor, the CAHA-AXRNet model by Sankareshwaran et al. [22], achieved an accuracy of 98.10%. More recent models, including RegNetY080 [4], FVBC [17], and DEX (DenseNet121, EfficientNetB7, and Xception) ensemble [2] showed significant (97–98%), also demonstrated strong performance [8], yet they fall short compared to our method [25]. This substantial improvement highlights the effectiveness of incorporating advanced loss functions like ArcFace [26] and Center Loss [9], which enhance the discriminative capacity of the feature embeddings. Furthermore, the use of InceptionNetV3 as the backbone provides an efficient and powerful feature extraction mechanism suitable for leaf disease classification. VI Conclusion This paper proposes a framework for rice leaf disease (RLD) classification that addresses intra-class compactness and inter-class separability by combining ArcFace and Center Loss. ArcFace enhances inter-class discrimination with angular margins, while Center Loss improves intra-class consistency. This dual-loss synergy proves particularly advantageous in scenarios characterized by high class similarity and limited labeled data. This is, to our knowledge, the first study to implement this dual-loss combination for RLD classification. Our best fine-tuned pretrained model achieved 99.6% accuracy on the target dataset, setting a new benchmark in this domain. However, evaluation was limited to only two datasets with few disease types, affecting generalizability. Future work includes expanding dataset diversity, incorporating multi-source and cross-seasonal data, and evaluating model robustness under diverse environmental conditions. Deployment on edge and cloud systems will support real-time, affordable farmer solutions, while integrating XAI will enhance transparency and trust in automated predictions. Beyond rice, the proposed approach also shows strong potential for broader crop disease detection tasks.This work advances precision agriculture by integrating advanced deep learning techniques with specific domain challenges to develop a realistic, dependable, and scalable solution for plant disease categorization. References [1] A. K. Abasi, S. N. Makhadmeh, O. A. Alomari, M. Tubishat, and H. J. Mohammed (2023-10) Enhancing rice leaf disease classification: a customized convolutional neural network approach. Sustainability 15 (20), p. 15039. External Links: ISSN 2071-1050, Link, Document Cited by: §I, TABLE I, §V-C. [2] M. T. Ahad, Y. Li, B. Song, and T. Bhuiyan (2023-09) Comparison of cnn-based deep learning architectures for rice diseases classification. Artificial Intelligence in Agriculture 9, p. 22–35. External Links: ISSN 2589-7217, Link, Document Cited by: TABLE I, §V-C. [3] M. Baite, S. Raghu, S. Prabhukarthikeyan, U. Keerthana, N. Jambhulkar, and P. Rath (2020) Disease incidence and yield loss in rice due to grain discolouration. Journal of Plant Diseases and Protection 127, p. 9–13. External Links: Document Cited by: §I. [4] H. Catal Reis and V. Turk (2024-02) Integrated deep learning and ensemble learning model for deep feature-based wheat disease detection. Microchemical Journal 197, p. 109790. External Links: ISSN 0026-265X, Link, Document Cited by: TABLE I, §V-C. [5] A. Chakrabarty, S. T. Ahmed, Md. F. U. Islam, S. M. Aziz, and S. S. Maidin (2024-09) An interpretable fusion model integrating lightweight cnn and transformer architectures for rice leaf disease identification. Ecological Informatics 82, p. 102718. External Links: ISSN 1574-9541, Link, Document Cited by: §I. [6] J. Deng, J. Guo, and S. Zafeiriou (2018) ArcFace: additive angular margin loss for deep face recognition. 2019 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), p. 4685–4694. External Links: Document Cited by: §I. [7] K. O. M. Goh, Md. K. Morol, Md. J. Hossen, Md. A. Al-Jubair, R. I. Rabbi, and N. Fahad (2025-03) The next chapter in wound analysis: introducing a hybrid model for improved segmentation with the help of deep convolutional neural network. Journal of Advanced Research in Applied Sciences and Engineering Technology 63 (1), p. 225–239. External Links: ISSN 2462-1943, Link, Document Cited by: §I. [8] B. M. T. Haque and Md. A. Rahman (2021) ERP modernization outcomes in cloud migration: a meta-analysis of performance and total cost of ownership (tco) across enterprise implementations. International Journal of Scientific Interdisciplinary Research 2 (2), p. 168–203. External Links: Document, Link Cited by: §V-C. [9] M. J. Hasan, S. H. Shifat, J. Matubber, R. Hossain, M. A. Rahman, B. M. T. Haque, and M. J. Hossen (2026) An in-depth exploration of machine learning methods for mental health state detection: a systematic review and analysis. Frontiers in Digital Health 7. External Links: Document, Link, ISSN 2673-253X Cited by: §V-C. [10] E. Hoffer and N. Ailon (2015) Deep metric learning using triplet network. In Similarity-Based Pattern Recognition, p. 84–92. External Links: ISBN 9783319242613, ISSN 1611-3349, Link, Document Cited by: §I. [11] S. Md. M. Hossain, Md. M. M. Tanjil, M. A. B. Ali, M. Z. Islam, Md. S. Islam, S. Mobassirin, I. H. Sarker, and S. M. R. Islam (2020) Rice leaf diseases recognition using convolutional neural networks. In Advanced Data Mining and Applications, p. 299–314. External Links: ISBN 9783030653903, ISSN 1611-3349, Link, Document Cited by: §I. [12] P. Kaur, S. Harnal, R. Tiwari, S. Upadhyay, S. Bhatia, A. A. Mashat, and A. Alabdali (2022) Recognition of leaf disease using hybrid convolutional neural network by applying feature reduction. Sensors (Basel, Switzerland) 22. External Links: Document Cited by: §I. [13] A. Kumar and B. Bhowmik (2023) Rice cultivation and its disease classification in precision agriculture. 2023 International Conference on Artificial Intelligence and Smart Communication (AISC), p. 200–205. External Links: Document Cited by: §I. [14] G. Modak, S. S. Das, Md. A. Islam Miraj, and Md. K. Morol (2022-03) A deep learning framework to reconstruct face under mask. In 2022 7th International Conference on Data Science and Machine Learning Applications (CDMA), External Links: Link, Document Cited by: §I. [15] Md. K. Morol, Md. S. J. Rokon, I. B. Hasan, A. M. Saif, R. H. Khan, and S. S. Das (2022-03) Food recipe recommendation based on ingredients detection using deep learning. In Proceedings of the 2nd International Conference on Computing Advancements, ICCA 2022, p. 191–198. External Links: Link, Document Cited by: §I. [16] R. Mukherjee, A. Ghosh, C. Chakraborty, J. N. De, and D. P. Mishra (2025-01) Rice leaf disease identification and classification using machine learning techniques: a comprehensive review. Engineering Applications of Artificial Intelligence 139, p. 109639. External Links: ISSN 0952-1976, Link, Document Cited by: §I. [17] B. Naresh kumar and S. Sakthivel (2025-03) Rice leaf disease classification using a fusion vision approach. Scientific Reports 15 (1). External Links: ISSN 2045-2322, Link, Document Cited by: §I, TABLE I, §V-C. [18] H. B. Prajapati, J. P. Shah, and V. K. Dabhi (2017) Detection and classification of rice plant diseases. Intelligent Decision Technologies 11, p. 357 – 373. External Links: Document Cited by: §I. [19] Project (2023-03) Rice leaf disease detection dataset. Roboflow. Note: https://universe.roboflow.com/project-khcjh/rice-leaf-disease-detectionAccessed: 2025-06-04 External Links: Link Cited by: §I-A. [20] S. Ramesh and D. Vydeki (2020-06) Recognition and classification of paddy leaf diseases using optimized deep neural network with jaya algorithm. Information Processing in Agriculture 7 (2), p. 249–260. External Links: ISSN 2214-3173, Link, Document Cited by: §I. [21] Md. A. H. Reaj, M. R. Abir, Md. A. Rahman, et al. (2025-11-17) XFE-bert: the way to the interpretable financial text analysis. Research Square. Note: Preprint, Version 1 External Links: Document, Link Cited by: §V-C. [22] S. P. Sankareshwaran, G. Jayaraman, P. Muthukumar, and A. Krishnan (2023-08) Optimizing rice plant disease detection with crossover boosted artificial hummingbird algorithm based ax-retinanet. Environmental Monitoring and Assessment 195 (9). External Links: ISSN 1573-2959, Link, Document Cited by: TABLE I, §V-C. [23] D. I. D. Saputra (2022) Rice leaf’s disease dataset. Note: https://w.kaggle.com/datasets/dedeikhsandwisaputra/rice-leafs-disease-datasetAccessed on 25th June 2025 Cited by: §V-B. [24] A. Stephen, A. Punitha, and A. Chandrasekar (2022-11) Designing self attention-based resnet architecture for rice leaf disease classification. Neural Computing and Applications 35 (9), p. 6737–6751. External Links: ISSN 1433-3058, Link, Document Cited by: §I. [25] B. M. Taslimul, T. A. I. Foysal, Md. I. Hossan, A. A. Noman, Md. A. Rahman, A. Ahmed, and Md. J. Hossen (2026-01) The role of ai and automation in it risk governance and enterprise systems. International Journal of Applied Mathematics 39 (1s), p. 151–163. External Links: Document, Link Cited by: §V-C. [26] B. M. Taslimul, Md. I. Hossan, A. A. Noman, Md. A. Rahman, A. Ahmed, D. Rahman, G. Md. Mohiuddin, and T. H. Liew (2026) Enterprise risk management in erp implementation: challenges, strategies and recent trends in ai – a mini review. International Journal of Applied Mathematics 39 (1s). Cited by: §V-C. [27] Y. Wen, K. Zhang, Z. Li, and Y. Qiao (2016) A discriminative feature learning approach for deep face recognition. In Computer Vision – ECCV 2016, p. 499–515. External Links: ISBN 9783319464787, ISSN 1611-3349, Link, Document Cited by: §I. [28] C. Zhou, Y. Zhong, S. Zhou, J. Song, and W. Xiang (2023-05) Rice leaf disease identification by residual-distilled transformer. Engineering Applications of Artificial Intelligence 121, p. 106020. External Links: ISSN 0952-1976, Link, Document Cited by: §I.