SugarcaneShuffleNet: Lightweight Disease Diagnosis
- The paper demonstrates a fine-tuned ShuffleNet model using transfer learning and Bayesian hyperparameter optimization, achieving 98.02% accuracy on a 17-class sugarcane disease diagnostic task.
- SugarcaneShuffleNet is a lightweight CNN with a 9.26 MB model size and an average inference time of 4.14 ms, ideal for deployment on low-resource devices in agricultural settings.
- The integrated pipeline combines the curated SugarcaneLD-BD dataset, Grad-CAM interpretability, and the SugarcaneAI application to deliver practical and explainable disease diagnosis for farmers.
Searching arXiv for the named paper and the ShuffleNet-related context paper. SugarcaneShuffleNet is a lightweight convolutional neural network for sugarcane leaf-disease diagnosis, introduced together with the SugarcaneLD-BD dataset and the SugarcaneAI Progressive Web Application. In the cited work, it is presented as an optimized ShuffleNet-based model for rapid on-device inference under low-resource conditions, with transfer learning and Bayesian hyperparameter optimization used to adapt the backbone to sugarcane pathology. On the combined 17-class test set, the reported model size is 9.26 MB, with 98.02% accuracy, 0.98 precision, 0.98 recall, 0.98 F1-score, 0.3951 test loss, and an average inference time of 4.14 ms per image (Arman et al., 23 Aug 2025).
1. Problem domain and system definition
SugarcaneShuffleNet is situated in the problem setting of field-realistic plant-disease diagnosis, where classification must remain fast, accurate, and interpretable despite constrained hardware, variable image acquisition conditions, and uneven class frequencies. The motivating constraints are explicitly low-resource deployment, real-world variability in lighting and background clutter, class imbalance with rare diseases, and the need for trust and interpretability in farmer-facing use (Arman et al., 23 Aug 2025).
Within that setting, the model is not described as a generic benchmark CNN, but as the model component of a broader diagnostic pipeline. The same study introduces SugarcaneLD-BD, a curated Bangladeshi field dataset; a larger merged sugarcane corpus assembled from three sources; SugarcaneShuffleNet as the optimized lightweight classifier; and SugarcaneAI as the field-deployment interface. This framing is important because the model’s significance in the paper derives not only from classification accuracy, but from its placement in an end-to-end diagnostic workflow spanning data collection, model optimization, deployment, explanation, and recommendation generation (Arman et al., 23 Aug 2025).
Architecturally, SugarcaneShuffleNet is described as a fine-tuned ShuffleNet backbone optimized for sugarcane disease classification. The justification for the ShuffleNet choice follows the standard efficient-CNN rationale: point-wise group convolutions reduce computation, and channel shuffle restores information flow across groups that would otherwise be bottlenecked by grouped processing. In the study’s presentation, these properties make ShuffleNet suitable for mobile and edge inference, and therefore suitable for agricultural deployment where GPU access and cloud connectivity may be limited (Arman et al., 23 Aug 2025).
2. Data resources and corpus engineering
The paper’s first dataset contribution is SugarcaneLD-BD, a field-collected dataset from Bangladesh designed to reflect realistic disease appearance. Images were collected from four locations: the BSRI research field in Ishwardi (Pabna), the BSRI Regional Station in Gazipur, farmer fields in Narsingdi, and farmer fields in Natore. The collection interval was September and October 2023. Acquisition used two smartphones, Realme 8 and Xiaomi Redmi Note 11, under varying illumination, different times of day, multiple viewpoints or angles, and diverse backgrounds. All labels were verified by qualified plant pathologists, and the original images were resized to 224 × 224 (Arman et al., 23 Aug 2025).
SugarcaneLD-BD contains 638 images across five classes: Red Rot, Ring Spot, Red Leaf Spot, Eye Spot, and Healthy. The study characterizes these as four major diseases plus a healthy class. The relatively small size is explicitly acknowledged, but the dataset is presented as carefully curated and representative of field conditions rather than laboratory-only imagery (Arman et al., 23 Aug 2025).
To enlarge class coverage and improve representativeness, the authors combined SugarcaneLD-BD with two Indian datasets: Thite et al., with 6,748 images across 11 classes, and Daphal and Koli, with 2,521 images across 5 classes. The resulting combined corpus contained 9,908 images spanning 17 classes: Banded Chlorosis, Brown Rust, Brown Spot, Dried Leaves, Eye Spot, Grassy Shoot, Healthy, Mosaic, Pokkah Boeng, Red Rot, Red Leaf Spot, Ring Spot, Rust, Sett Rot, Smut, Viral Disease, and Yellow Leaf (Arman et al., 23 Aug 2025).
The merged corpus was heavily imbalanced. The reported large classes include Brown Spot with 1,722 images and Yellow Leaf with 1,699, whereas Red Leaf Spot had 43, Eye Spot 75, and Ring Spot 83. The initial imbalance ratio was about 26:1. To address this, the authors applied a multi-stage cleaning and balancing procedure. Exact duplicates were removed using MD5 hashing, near duplicates were removed using a 5-bit difference threshold, and filenames were standardized. After deduplication, the dataset size fell from 9,908 to 7,037 images, with 1,264 exact duplicates and 1,607 near duplicates removed. Class-dependent augmentation was then used: classes with fewer than 100 images were augmented up to 600, while larger classes were capped at 1200 or 1.5× their original count. Augmentation produced 5,691 additional training images. The final split was 80/20 stratified, with augmentation applied only to training data, yielding 11,313 training images and 1,414 test images, and reducing the imbalance ratio from 26:1 to 3.8:1 (Arman et al., 23 Aug 2025).
3. Architecture, transfer learning, and hyperparameter optimization
SugarcaneShuffleNet is presented as a deployment-oriented instantiation of ShuffleNet rather than as a newly invented macro-architecture. The study fine-tunes a pretrained ShuffleNet backbone using transfer learning and Bayesian hyperparameter optimization. This distinction matters: the novelty claimed in the paper lies in optimization for sugarcane disease classification, the dataset pipeline, and the deployment stack, rather than in proposing a new generic CNN primitive (Arman et al., 23 Aug 2025).
The optimization framework uses Optuna with a Tree-structured Parzen Estimator sampler. Each architecture underwent 20 trials, each trial trained for 25 epochs, with validation loss as the optimization objective. The search space included learning rate from to on a log scale, optimizer choice between Adam and AdamW, weight decay from to on a log scale, dropout rate 1 from 0.1 to 0.6, dropout rate 2 from 0.1 to 0.6, freeze ratio from 0.0 to 0.8, label smoothing from 0.0 to 0.2, and gradient clipping from 0.5 to 2.0 (Arman et al., 23 Aug 2025).
The final training configuration used CrossEntropyLoss, batch size 32, CosineAnnealingLR, 100 epochs of full training, early stopping with patience 10, and pretrained weights. For SugarcaneShuffleNet, the best reported hyperparameters were validation accuracy 98.01%, dropout1 0.480, dropout2 0.492, freeze ratio 0.453, gradient clip 1.702, label smoothing 0.052, learning rate , optimizer Adam, and weight decay (Arman et al., 23 Aug 2025).
The broader model comparison in the paper places ShuffleNet alongside MnasNet, EdgeNeXt, EfficientNet-Lite, MobileNet, and SqueezeNet. The rationale given for these comparators is that each represents a lightweight CNN family oriented toward high accuracy per parameter, low MACs, or mobile latency. In the authors’ characterization, SugarcaneShuffleNet emerges from this comparison as the selected compromise between predictive performance and deployment constraints rather than as the single absolute optimum on every isolated metric (Arman et al., 23 Aug 2025).
4. Evaluation protocol and empirical performance
The evaluation protocol emphasizes standard classification metrics as well as deployment-oriented efficiency metrics. The paper defines accuracy, precision, and recall as
and places particular emphasis on macro-F1 because some disease classes are rare and accuracy alone can therefore be misleading. Additional reported metrics include inference time per image, parameter count, model size, MACs/FLOPs, and qualitative suitability for ARM, Raspberry Pi, and TFLite-type deployment (Arman et al., 23 Aug 2025).
On the combined 17-class test set, SugarcaneShuffleNet achieved 98.02% accuracy, 0.98 precision, 0.98 recall, 0.98 F1-score, and 0.3951 test loss. Class-wise performance was strongest on Dried Leaves, Grassy Shoot, Pokkah Boeng, Ring Spot, Sett Rot, and Smut, each with F1 equal to 1.00. The hardest class was Red Leaf Spot, with F1 equal to 0.89; the paper connects this difficulty to the class’s rarity, with only 43 original samples, and to visual similarity with other lesions. The study further reports that most classes had F1 greater than 0.95, along with precision–recall curves in which most classes were near 1.0 average precision, ROC–AUC curves with nearly all classes close to 1.0 AUC, class-wise accuracy with 95% confidence intervals, and t-SNE feature visualization showing well-separated clusters (Arman et al., 23 Aug 2025).
The central comparative claim concerns the speed–accuracy–size trade-off. MnasNet achieved 98.51% accuracy, 0.98 precision, 0.98 recall, and 0.98 F1, with 324.19 MMac, 4.42 M parameters, 2.91 ms inference time, and 17.59 MB size. EdgeNeXt achieved 98.23% accuracy, 0.97 precision, 0.97 recall, and 0.97 F1, but required 971.41 MMac, 5.55 M parameters, 3.02 ms, and 22.02 MB. EfficientNet-Lite achieved 98.02% accuracy and 0.98 F1 with 373.91 MMac, 4.70 M parameters, 3.26 ms, and 18.65 MB. MobileNet was the fastest at 2.55 ms and lowest in MACs at 57.56 MMac, but its accuracy dropped to 96.53%. SqueezeNet was smallest at 7.00 MB but reached 97.31% accuracy and 542.16 MMac. ShuffleNet, the backbone family used for SugarcaneShuffleNet, achieved 98.02% accuracy, 0.98 precision, 0.98 recall, 0.98 F1, 152.43 MMac, 2.19 M parameters, 4.14 ms inference time, and 9.26 MB size (Arman et al., 23 Aug 2025).
The paper therefore identifies SugarcaneShuffleNet as the preferred practical compromise: near-top accuracy, a low parameter count of 2.19 M, modest size of 9.26 MB, reasonable inference time of 4.14 ms, and strong per-class F1 on the 17-class task. Relative to prior sugarcane studies, the article notes that EfficientNet-B7 reached 99.79% accuracy on 11 classes but exceeded 250 MB, SE-ViT Hybrid reached 89.57% on 5 classes, Hybrid CNN + GLCM reached 96.80% on 5 classes, and AMRCNN reached 86.53% on 5 classes. The comparison is framed to show that SugarcaneShuffleNet addresses a harder 17-class task while retaining practical deployment suitability (Arman et al., 23 Aug 2025).
5. Deployment architecture, interpretability, and advisory functions
SugarcaneShuffleNet is deployed within SugarcaneAI, described as a Progressive Web Application for field use. The application allows users either to upload a leaf image or capture one on site, runs inference on-device or edge-side, and returns the top predicted disease, a confidence score, and the top five candidate classes. This application-level integration is part of the paper’s substantive contribution, because it situates the CNN in an operational agronomic workflow rather than in an isolated benchmark setting (Arman et al., 23 Aug 2025).
Interpretability is implemented through Grad-CAM. The stated purpose is to verify whether the model is attending to chlorotic bands, necrotic rings, rust pustules, and symptomatic lesions rather than to background noise. The reported Grad-CAM findings indicate that heatmaps consistently highlight disease-relevant regions, including red rot symptom regions, ring spot lesions, chlorotic areas associated with eye spot, and other lesion structures. The paper presents this as important for both farmer trust and biological plausibility. The t-SNE feature-space visualizations complement this by showing class clustering, which the authors interpret as evidence of learned separability rather than memorization of incidental artifacts (Arman et al., 23 Aug 2025).
The application also includes an advisory component. After diagnosis, the user can request recommendations, and the app queries Gemini with the predicted disease name. The returned guidance is structured into three categories: the cause of the disease, immediate containment steps, and long-term control strategies. The paper describes this as LLM-generated management advice delivered through the app interface. In combination with Grad-CAM explanations, this makes SugarcaneAI not only a classifier front end but a practical decision-support layer built around SugarcaneShuffleNet (Arman et al., 23 Aug 2025).
6. Position within ShuffleNet and efficient-CNN research
SugarcaneShuffleNet belongs to the ShuffleNet lineage in the direct sense that the underlying model is a fine-tuned ShuffleNet backbone. In the study’s account, ShuffleNet is chosen because point-wise group convolutions reduce computation and channel shuffle restores cross-group information flow, producing a network with a small parameter footprint and low MACs suitable for mobile and edge inference. The authors further note compatibility with TensorFlow Lite, PyTorch Mobile, and TVM, as well as support for quantization and mixed-precision deployment, which reinforces its fit for practical low-resource systems (Arman et al., 23 Aug 2025).
A common source of ambiguity is the possible conflation of SugarcaneShuffleNet with other ShuffleNet-adjacent efficient-CNN proposals that modify the channel-mixing mechanism itself. In particular, the earlier paper “Shift-based Primitives for Efficient Convolutional Neural Networks” proposed channel shift as a faster alternative to ShuffleNet’s channel shuffle, together with address shift and shortcut shift, and built AddressNet and Enhanced AddressNet around pointer movement rather than memory-copy operations. That paper reported, for example, that channel shift was 12.7× faster than channel shuffle while achieving the same accuracy in the cited comparison, and it positioned the resulting architectures as ShuffleNet-like compact CNNs optimized for GPU inference (Zhong et al., 2018).
SugarcaneShuffleNet is distinct from that line of work. It is not named in the 2018 shift-primitives paper, and the disease-diagnosis model is not presented as an AddressNet or channel-shift architecture. Rather, it is a domain-adapted ShuffleNet used as the classification core of a sugarcane diagnostic system. The conceptual connection is therefore one of lineage and design space: both works are situated in compact CNN research focused on grouped computation, low latency, and mobile deployment, but they operate at different levels of contribution. The 2018 paper modifies efficient-CNN primitives themselves, whereas SugarcaneShuffleNet applies and optimizes a ShuffleNet backbone for a specific agricultural diagnostic task (Arman et al., 23 Aug 2025).
From that perspective, SugarcaneShuffleNet can be understood as a specialized, deployment-oriented ShuffleNet instantiation whose significance lies in the conjunction of dataset curation, imbalance-aware corpus construction, transfer-learned and Bayesian-optimized lightweight modeling, Grad-CAM-based interpretability, and field-facing application design. This suggests that its main research value is not only architectural efficiency in isolation, but the demonstration that a compact grouped-convolution network can be embedded in a realistic end-to-end pipeline for 17-class sugarcane disease classification under on-farm conditions (Arman et al., 23 Aug 2025).