BanglaRiceLeaf: A benchmark dataset for automated rice leaf disease detection and health classification in Bangladesh.

Abstract

Rice leaf diseases pose a major challenge to crop health and agricultural productivity, particularly when timely and accurate diagnosis is required under natural field conditions. The development of automated disease recognition systems depends heavily on the availability of large, well-annotated image datasets. However, many existing rice leaf disease datasets are limited in terms of environmental variability, disease representation, and real-field imaging conditions. To address this gap, this paper presents BanglaRiceLeaf, an original rice leaf image dataset collected and curated by the authors from the experimental fields of the Bangladesh Rice Research Institute (BRRI), Gazipur, Bangladesh, between July 2023 and July 2024. The dataset contains 4152 images belonging to five classes: Bacterial Leaf Blight, Bacterial Leaf Streak, Sheath Blight, Leaf Blast, and Healthy Leaf. The images were acquired from two rice varieties, BR11 and BRRI dhan34, under natural field conditions across varying illumination environments in order to reflect practical disease recognition scenarios. All images were manually annotated by trained annotators under expert supervision. The dataset is systematically organized and publicly released to support reproducible research in rice disease classification. In addition to dataset presentation, benchmark experiments using Xception, NASNetMobile, and InceptionV3 are provided to demonstrate its applicability for deep learning-based disease recognition. BanglaRiceLeaf is expected to serve as a useful resource for plant disease analysis, comparative model evaluation, and future research in precision agriculture and agricultural computer vision.

Plant phenotyping relevance

イネ葉の病徴・健全状態を画像で分類する公開ベンチマークデータセットであり、データ収集・注釈・ベンチマーク評価が中心です。

Code and data availability

The paper's core asset is the BanglaRiceLeaf rice leaf disease image dataset (4152 field images, five classes), publicly released on Harvard Dataverse with DOI 10.7910/DVN/XAOBYW. No author analysis code or trained model checkpoints are stated as publicly available.