Bangladesh Field Dataset Adds 11-Class Visual Evidence for Brinjal Diagnosis
A new brinjal image dataset assembled in Bangladesh gives researchers a structured way to train and test models that distinguish leaf disease, fruit damage, pest infestation, nutritional deficiency, and healthy plant tissue. The collection centers on Solanum melongena, a crop where visual diagnosis often determines whether a grower can intervene early enough to limit yield loss. In practical terms, the dataset turns field observations into a machine-readable resource, capturing the messy reality of farm conditions rather than the controlled perfection of studio images.
The release contains two versions built from the same underlying material. The RawDataset includes 2,206 original images distributed across 11 classes, while the AugmentedDataset expands the collection to 8,800 images, with 800 images per class. That second version creates a balanced class structure for machine learning experiments, a useful correction when one category would otherwise dominate model training and distort performance. The images were collected using multiple smartphone devices in field settings across Bangladesh, which matters because real-world crop diagnostics must cope with variable lighting, angles, background clutter, and partial occlusion.
From Unstructured Farm Images to Trainable Classes
The 11 labeled categories cover a broad spectrum of agronomic conditions: Aphids, Cercospora Leaf Spot, Defect Fruit, Early Blight, Healthy Fruit, Healthy Leaf, Magnesium Deficiency, Mosaic Virus, Phomopsis Blight, Phytophthora Blight, and Powdery Mildew. That spread makes the dataset more than a narrow disease archive. It captures insect injury, fungal and oomycete pathology, viral symptoms, nutrient stress, and normal tissue in a single framework, which is exactly the kind of multimodal visual burden that field classifiers must learn to separate.
The distinction between a diseased leaf and a healthy one is only part of the challenge. Several of these classes can produce overlapping visual signatures at early stages, and nutrient deficiency can mimic infectious disease when chlorosis, mottling, or texture change first appear. A model trained only on pristine or overly curated photographs will often fail when deployed in open-air agriculture. By preserving images collected under natural conditions, the dataset creates a harder, more realistic benchmark for computer vision systems.
Each image is standardized to 800 × 800 pixels and stored as a .jpg file. Standardization simplifies preprocessing and gives downstream models a consistent input size without erasing the field-based character of the photographs. The parent folder, BrinjalDataset, contains both the raw and augmented collections in parallel subfolders, each organized into the same 11 class directories. This structure supports direct comparisons between training on naturally imbalanced data and training on a balanced, expanded version.
Why Balanced Data Matters for Crop Recognition Models
In image classification, imbalance can quietly skew a model’s behavior. If one disease appears far more often than another, a classifier may learn to favor the most common label and still appear accurate on aggregate metrics while missing rarer but agronomically important conditions. The augmented set addresses that problem by equalizing the class counts at 800 images per class. For developers building deep learning systems, that balance can improve convergence, stabilize training, and reduce bias toward majority categories.
There is a second advantage as well. Balanced classes make error analysis cleaner. When a model confuses Phomopsis blight with Phytophthora blight, or Magnesium Deficiency with Mosaic Virus, the failure is easier to study because it is not buried inside a skewed label distribution. That matters for field deployment, where the cost of misclassification may be a delayed treatment decision, an unnecessary spray, or missed support for a nutrient problem that would respond to a different intervention altogether.
The dataset’s field collection method also increases its relevance. Smartphone-captured images reflect the tools most farmers and extension workers actually use. A diagnostic system built from those images is more likely to generalize to the conditions where brinjal is grown, sold, and inspected, especially in settings where access to laboratory testing is limited or slow.
Computer Vision Use Cases Extend Beyond Disease Labels
Researchers can use the dataset for more than simple disease classification. The description points to applications in plant disease and symptom recognition, pest and nutritional disorder recognition, fruit condition classification, and broader agricultural image analysis. That range opens the door to models that do not merely identify a named disease, but also decide whether a fruit is defective, whether a leaf is healthy, or whether the visible change is likely linked to insect pressure or mineral insufficiency.
The practical value is straightforward. Early visual recognition can support triage in crop management, allowing attention to move toward the plants that need it most. In regions where field scouting is labor-intensive and specialist access is uneven, an automated classifier can serve as a first-pass filter rather than a final authority. It can flag suspicious specimens for human review, help standardize extension workflows, and support research teams comparing model architectures on a real agricultural problem.
Because the images were manually labeled and organized into class-specific folders, the dataset also provides a reference structure for supervised learning studies. That makes it suitable for benchmarking convolutional neural networks and related deep learning approaches against a common ground truth. The absence of any associated publication or supplementary data keeps the dataset’s current role focused on image analysis itself rather than on a wider experimental package.
What the Release Contributes to Agricultural AI Research
For developers and agronomists, the most useful part of this release is not its size alone. It is the combination of field realism, class diversity, and a parallel balanced version that can support both exploratory work and cleaner model evaluation. The raw collection preserves natural imbalance, which is useful for studying how algorithms behave under authentic conditions. The augmented version offers a balanced counterpoint for training and comparison.
The dataset is distributed under Creative Commons Attribution 4.0 International, which should encourage reuse in academic and applied projects. No grants or funding data are listed, and no associated publication is attached, so the dataset currently stands on its own as a resource rather than as an extension of a larger trial or validation paper. For the moment, its value lies in what it enables: repeatable experiments, benchmark comparisons, and better visual classifiers for a crop that faces overlapping disease, pest, and nutrient threats in the field.
In a domain where diagnosis can depend on the trained eye of a local expert, the release gives machine learning researchers something more concrete to work with: 2,206 original field images, 8,800 balanced augmented images, and 11 clearly defined categories that reflect the complexity of real brinjal production. The next challenge is less about collecting more pictures than about proving which models can read them reliably when the crop is in the ground, not the lab.