SugarcaneLD-BD: Bangladeshi Leaf-Disease Dataset

• SugarcaneLD-BD is a curated, Bangladesh-origin dataset comprising 638 field-acquired images across five classes of sugarcane diseases.
• It offers regional realism and expert-validated annotations, addressing gaps in traditional plant disease datasets with diverse agroecological conditions.
• Integrated into a 17-class merged benchmark, it underpins lightweight CNN models like SugarcaneShuffleNet to achieve high classification performance in real-world scenarios.

SugarcaneLD-BD is a curated dataset for sugarcane leaf-disease classification introduced in the paper “SugarcaneShuffleNet: A Very Fast, Lightweight Convolutional Neural Network for Diagnosis of 15 Sugarcane Leaf Diseases.” It contains 638 curated images across five classes, including four major sugarcane diseases and one healthy class, collected in Bangladesh under diverse field conditions and verified by expert pathologists. In the paper’s contribution structure, SugarcaneLD-BD is the dataset component, while SugarcaneShuffleNet is the model contribution and SugarcaneAI is the deployment component (Arman et al., 23 Aug 2025).

1. Dataset identity and rationale

SugarcaneLD-BD was introduced to address the scarcity of public, field-representative sugarcane disease data. The paper states that existing general plant datasets such as PlantVillage and PlantDoc do not include sugarcane leaf diseases, and that existing sugarcane datasets are limited by geography, class coverage, or labeling consistency. The dataset is positioned as a Bangladesh-specific contribution intended to provide regional realism, expert-verified labels, and field variability in background, viewpoint, and lighting (Arman et al., 23 Aug 2025).

The authors explicitly argue that disease appearance depends on local agroecological conditions. On that basis, SugarcaneLD-BD is presented not merely as an additional image collection, but as a dataset meant to capture Bangladesh-relevant phenotypes that may not be represented well by Indian datasets alone. The paper further states that SugarcaneLD-BD can be merged with existing datasets to produce a larger and more representative corpus, which is exactly how it is used in the reported experiments (Arman et al., 23 Aug 2025).

A common misunderstanding is to treat SugarcaneLD-BD as a fully benchmarked standalone dataset with its own train/validation/test protocol and headline metrics. The paper does not do that. Instead, it introduces SugarcaneLD-BD as the local-data foundation of a broader benchmarking and deployment pipeline, and all reported model results are on a merged multi-source corpus rather than on SugarcaneLD-BD in isolation (Arman et al., 23 Aug 2025).

2. Acquisition, geography, and annotation

SugarcaneLD-BD was collected from four sugarcane-producing locations in Bangladesh, selected to reflect different agroecological conditions, farming practices, and disease prevalence (Arman et al., 23 Aug 2025).

Location	Coordinates	Agroecological zone
BSRI research field, Ishwardi, Pabna	24.130000° N, 89.060000° E	Active Ganges Floodplain (AEZ 10)
BSRI Regional Station, Gazipur	23.999941° N, 90.420273° E	Middle Meghna River Floodplain (AEZ 16)
Farmer fields in Narsingdi	23.920700° N, 90.718800° E	Old Meghna Estuarine Floodplain (AEZ 15)
Farmer fields in Natore	24.411000° N, 89.012000° E	High Ganges River Floodplain (AEZ 11)

The acquisition period was September and October 2023. Two smartphones were used: Realme 8 and Xiaomi Redmi Note 11. The paper reports the following native image specifications: Realme 8: 2608 × 4624, RGB, focal length 4.71 mm, aperture f/1.8, exposure 1/104 s, ISO 100; Xiaomi Redmi Note 11: 2296 × 4080, RGB, focal length 4.25 mm, aperture f/1.8, exposure 1/120 s, ISO 58. Images were captured under varying lighting conditions, at different times of day, from multiple angles, and with diverse backgrounds, emphasizing field realism rather than controlled capture (Arman et al., 23 Aug 2025).

All images were then resized to 224 × 224 pixels. The stated rationale is standard CNN input compatibility, simplification of the input pipeline, and reduced computational complexity. No other preprocessing steps are reported for SugarcaneLD-BD itself beyond resizing (Arman et al., 23 Aug 2025).

The annotation regime is described as expert-led: images were carefully labeled by qualified plant pathologists, the dataset was expert-annotated, and labels were verified by expert pathologists. At the same time, the paper does not specify the number of experts, whether annotation was independent or consensus-based, whether disagreements were adjudicated, or whether laboratory confirmation was used. This leaves the validation protocol identifiable in principle but under-specified in procedure (Arman et al., 23 Aug 2025).

3. Class taxonomy and dataset composition

SugarcaneLD-BD contains five classes total: four disease classes and one healthy class. The four disease classes described in detail are Red Rot, Ring Spot, Red Leaf Spot, and Eye Spot (Arman et al., 23 Aug 2025).

Class	Description in the paper	Count explicitly reported for SugarcaneLD-BD
Red Rot	Caused by Colletotrichum falcatum; described as the most damaging sugarcane disease in Bangladesh	Not isolated
Ring Spot	Caused by Leptosphaeria sacchari; lesions with reddish-brown borders	83
Red Leaf Spot	Caused by Dimeriella sacchari; small red spots merging into larger lesions	43
Eye Spot	Caused by Bipolaris sacchari; long lesions with red centers and a narrow yellow chlorotic ring	75
Healthy	Healthy class mentioned in the composition table	Not isolated

The dataset total of 638 images is stated multiple times. However, the paper only explicitly isolates per-class counts for Eye Spot (75), Red Leaf Spot (43), and Ring Spot (83). It does not explicitly isolate SugarcaneLD-BD-specific counts for Red Rot and Healthy. The paper notes that the implied combined remainder for Healthy + SugarcaneLD-BD Red Rot is 437 images, but does not break that remainder down (Arman et al., 23 Aug 2025).

This partial transparency in class counts is important for interpretation. SugarcaneLD-BD is clearly defined taxonomically, but its fully enumerated per-class distribution is not reported in the same level of detail for all classes. A second point of clarification is that the paper does not report a standalone train/validation/test split for SugarcaneLD-BD, does not report a standalone augmentation protocol for it, and does not report standalone benchmark results on it (Arman et al., 23 Aug 2025).

4. Incorporation into the merged 17-class benchmark

Experimentally, SugarcaneLD-BD is merged with two Indian sugarcane disease datasets: the Thite et al. dataset from India with 11 classes and 6748 images, and the Daphal and Koli dataset from India with 5 classes and 2521 images. The paper states that the initial merged corpus had 17 distinct classes and 9,908 images total. It also notes a minor arithmetic inconsistency, because 6748 + 2521 + 638 = 9907, not 9908, although the official figure given repeatedly in the source text is 9,908 (Arman et al., 23 Aug 2025).

The merged class list includes Banded Chlorosis, Brown Rust, Brown Spot, Dried Leaves, Eye Spot, Grassy Shoot, Healthy, Mosaic, Pokkah Boeng, Red Rot / RedRot, Red Leaf Spot, Ring Spot, Rust, Sett Rot, Smut, Viral Disease, and Yellow Leaf. The paper indicates conceptual class harmonization across sources, but does not provide a canonical source-to-target mapping table. It also leaves some naming ambiguity, particularly the coexistence of Brown Rust and Rust, and the alternating use of Red Rot and RedRot (Arman et al., 23 Aug 2025).

A substantial curation step follows the merge. The paper reports that exact duplicates were removed using MD5 hashing, and near-duplicates were removed using a 5-bit difference threshold. The results are reported as follows: 1,264 exact duplicates removed, 1,607 near-duplicates removed, and a final deduplicated dataset of 7,037 images. Files were renamed into a standardized format such as ClassName_0001.jpg (Arman et al., 23 Aug 2025).

The final deduplicated corpus remains strongly imbalanced. The paper reports that the imbalance ratio improved from 26:1 to 3.8:1 after training-set augmentation. For the combined dataset, the split is 80% train and 20% test with stratified sampling, yielding 5623 pre-augmentation training images and 1414 test images. Augmentation is applied only to the training set, with class-dependent multiplicities; for example, Eye Spot, Red Leaf Spot, and Ring Spot each receive six augmentations per training image, while larger classes such as Brown Spot, Healthy, RedRot, and Yellow Leaf receive none. The final augmented training set contains 11,313 images, including 5,691 additional generated training images (Arman et al., 23 Aug 2025).

5. Benchmarking, SugarcaneShuffleNet, and deployment ecology

The paper’s central benchmark results are not reported for SugarcaneLD-BD alone, but for the combined 17-class deduplicated dataset that includes SugarcaneLD-BD. Within that setting, SugarcaneShuffleNet is presented as an optimized ShuffleNet-based lightweight CNN chosen for edge deployment because of its favorable balance between accuracy and efficiency. Its best hyperparameters were identified with Optuna using the Tree-structured Parzen Estimator, with 20 trials and 25 epochs per trial, and full training used CrossEntropyLoss, batch size 32, CosineAnnealingLR, 100 epochs, and early stopping with patience 10 (Arman et al., 23 Aug 2025).

On the merged benchmark, SugarcaneShuffleNet achieves 98.02% accuracy, 0.98 precision, 0.98 recall, and 0.98 F1, with 2.19 M parameters, 152.43 MMACs, 9.26 MB model size, and 4.14 ms average inference time per image. The same table shows that MnasNet reaches 98.51% accuracy, but the paper selects SugarcaneShuffleNet as the preferred deployment trade-off because MnasNet requires more parameters, memory, and computation (Arman et al., 23 Aug 2025).

The per-class F1 scores are particularly informative for the classes originating from SugarcaneLD-BD. On the combined benchmark, SugarcaneShuffleNet scores 0.97 for Eye Spot, 0.89 for Red Leaf Spot, 1.00 for Ring Spot, 0.98 for Red Rot, and 0.99 for Healthy. This shows that the hardest SugarcaneLD-BD-derived class in the merged benchmark is Red Leaf Spot, while Ring Spot and Healthy are nearly saturated under the reported evaluation protocol (Arman et al., 23 Aug 2025).

The same paper integrates the model into SugarcaneAI, a Progressive Web Application. The application accepts uploaded or real-time images, returns the top predicted class, confidence score, top five predicted classes, and a Grad-CAM heatmap, and can request recommendations from the Gemini API. The authors state that Grad-CAM heatmaps focus on meaningful disease features such as chlorotic bands, necrotic rings, rust pustules, and symptomatic red-rot regions. These deployment details matter because SugarcaneLD-BD is not treated in isolation; it functions as the data substrate for a full data-model-application stack (Arman et al., 23 Aug 2025).

6. Position within the sugarcane disease-data landscape

SugarcaneLD-BD occupies a specific niche within sugarcane disease datasets: it is a Bangladesh-origin, expert-annotated, field-acquired RGB leaf dataset intended for disease classification under variable real-world capture conditions. It is neither the first public sugarcane disease image dataset nor a universal sugarcane health benchmark. Earlier work such as “SugarcaneNet” uses a different public benchmark, the Sugarcane Leaf Disease Dataset from Mendeley Data, with 2569 images from Maharashtra, India, split into Healthy, Mosaic, RedRot, Rust, and Yellow disease, and that paper explicitly does not call its dataset “SugarcaneLD-BD” (Talukder et al., 2024).

SugarcaneLD-BD is also distinct from remote-sensing disease datasets and from hyperspectral screening studies. The broader sugarcane monitoring literature includes satellite-based multispectral detection of asymptomatic Ratoon Stunting Disease, where SVM-RBF achieved classification accuracy between 85.64% and 96.55% depending on variety using Sentinel-2-derived vegetation indices (Waters et al., 2024). A review of sugarcane health monitoring further emphasizes that satellite disease literature remains sparse and confounded by variety, crop age, soil background, viewing angle, weather, and atmospheric effects, even though the approach is considered promising (Waters et al., 2024). In that context, SugarcaneLD-BD represents a leaf-level RGB benchmark rather than a canopy-scale or satellite-scale disease-monitoring system.

Its principal strengths, as stated in the paper, are regional novelty, field realism, multi-zone agroecological diversity, expert validation, and public availability on Mendeley Data and Kaggle. Its principal limitations are equally explicit: small size at 638 images, severe imbalance for rare classes, incomplete annotation-protocol detail, no standalone split or standalone benchmark results, and no formal cross-region, cross-device, or cross-lighting generalization analysis (Arman et al., 23 Aug 2025).

A reasonable synthesis is that SugarcaneLD-BD is best understood as a curated, Bangladesh-specific leaf-disease dataset whose importance lies less in standalone scale than in the way it injects new geography, new phenotypes, and field-acquired variability into a merged 17-class sugarcane disease benchmark. In the source paper, its scientific role is foundational rather than self-sufficient: it fills a public data gap, sharpens regional relevance, and supports the training and deployment of lightweight sugarcane disease classifiers under low-resource conditions (Arman et al., 23 Aug 2025).