FloraBD: A Field-Collected Multi-Category Smartphone Image Dataset of Ornamental, Medicinal, and Wild Flowering Plants from Bangladesh for Computer Vision and Biodiversity Research

FloraBD is a large-scale, field-collected smartphone image dataset comprising 22,511 RGB images distributed across 45 flowering plant species captured from urban nurseries in Aftabnagar and Staff Quarter, Dhaka, Bangladesh. The dataset spans three ecologically distinct categories — ornamental, medicinal, and wild flowering species — making it the first publicly available multi-category flower image dataset specifically representing the botanical diversity of Bangladesh.
All images were captured using a CMF Phone 2 Pro smartphone (50 MP) under natural daylight conditions without artificial lighting or controlled backgrounds. Images were preprocessed to a uniform resolution of 1024×1024 pixels using center-crop resizing with LANCZOS resampling and stored in JPEG format at quality level 85. A manual quality review was conducted to remove blurred and low-quality frames prior to finalization.
Dataset quality was characterized through a comprehensive statistical analysis suite. The mean Laplacian sharpness score of 94.06 confirms well-focused imagery, while the mean colorfulness index of 55.84 reflects the visual richness of the collection. The mean Signal-to-Noise Ratio of 6.76 dB is consistent with natural field photography conditions. The intra-class diversity coefficient of 0.14 confirms that images within each class represent genuine visual variation rather than repetitive near-duplicate frames. Inter-class separability analysis using color histogram features yielded a Silhouette Score of -0.019, confirming the fine-grained nature of the dataset and the need for deep convolutional feature representations for reliable classification.
Class-wise image counts range from 231 images (Costus Speciosus) to 749 images (Arabian Jasmine), with a mean of 500.2 images per class. The dataset is organized into 45 class-named folders following a consistent serialized naming convention compatible with TensorFlow, PyTorch, and Keras data loaders.