RoseLeafVision: A Comprehensive Multi-Class Image Dataset for Accurate Rose Leaf Disease Detection

Description
Rose plant species are among the most widely used flowering plants, establishing themselves in the marketplace due to their aesthetic appeal, fragrance, as well as importance in the economy. They are also vulnerable to several diseases that can easily infect the leaves of the plants, reducing their appeal in the marketplace. In this regard, the RoseLeafVision dataset has been established.

Dataset Content
This dataset was collected from Golap Gram, Savar, Bangladesh, in July to September, 2025, where there are a total of 2,458 raw images of rose leaves collected in natural lighting conditions, although some were collected in a white background for clarity. The dataset has been categorized into five groups:
1. Black Spot: 335
2. Downy Mildew: 316
3. Dry Leaf: 712
4. Healthy Leaf: 668
5. Insect Hole: 427

After Augmentation
The data was then augmented to a total of 12,991 pictures using rotation, shearing, zoom, flipping, as well as brightness and contrast changes. These augmentation techniques were necessary to ensure equal representation of the five classes, as the models were trained. This was in a bid to counter biased models in the classification strategies. The next step, Augmented
1. Black Spot: 2567
2. Downy Mildew: 2564
3. Dry Leaf: 2641
4. Healthy Leaf: 2634
5. Insect Hole: 2585

Geographical Location
All rose leaf images were collected from Golap Gram, Savar, Dhaka, a well-known rose cultivation zone in Bangladesh that naturally provides diverse leaf conditions.
Location: Golap Gram, Savar, Dhaka
Latitude: 23°48'5.808''N
Longitude: 90°19'30.4908''E

Preprocessing Steps
1. Otsu Thresholding: separates the leaf region from the background.
2. Morphological Operations: This skips the noise and smooths the segmented leaf area.
3. Brightness Adjustment: Enhances the overall look of leaf, making images much more visible to understand.
4. CLAHE: Enhances the local contrast, more visible edge, emphasizing the disease patterns.
5. Normalization: It involves scaling pixel values to be compatible with a selected model.

Augmentation Details
1. Rotation (±30°): Random rotation helps to make the model robust against different angle of leaf orientations.
2. Shear Transformation (0.15): It ensure slight angular distortion, allowing the model to handle shape variations.
3. Zoom (±20%): Random zoom-in and zoom-out enhance scale invariance.
4. Horizontal Flip: Flips the image left to right to increase variety in the dataset and reduce directional bias.
5. Vertical Flip: top-to-bottom flipping of images in order to enhance model generalization.
6. Fill Mode Nearest: fills no-value pixels appearing due to transformation with the nearest pixel.

Use of the Dataset
1. Perform development and benchmarking of the AI models on automated rose leaf disease detection.
2. IoT-based plant monitoring system to improve smart agriculture solutions.
3. Support research in precision agriculture field.
