BRACOL - A Brazilian Arabica Coffee Leaf images dataset to identification and quantification of coffee diseases and pests

Published: 6 November 2019| Version 1 | DOI: 10.17632/yy2k5y8mxg.1
Renato Krohling,
José esgario,
Jose A. Ventura


The dataset was developed with the purpose to evaluate deep learning algorithms for segmentation and classification. it contains images of arabica coffee leaves affected by the main biotic stresses that affect the coffee tree: leaf miner, leaf rust, brown leaf spot, and cercospora leaf spot. The images were obtained using different smartphones (ASUS Zenfone 2, Xiaomi Redmi 5A, Xiaomi S2, Galaxy S8, and iPhone 6S). The leaves were collected at different times of the year in Santa Maria of Marechal Floreano in the mountains regions of the state of Espirito Santo, Brazil. The photos were taken from the abaxial (lower) side of the leaves under partially controlled conditions and placed on a white background. The acquisition of the images was done without much criterion to make the dataset more heterogeneous. A total of 1747 images of arabica coffee leaves were collected, including healthy leaves and diseased leaves, affected by one or more types of biotic stresses. The process of biotic stresses recognition for dataset labeling was assisted by an expert and performed with the captured images. From the obtained photos were generated two datasets. A dataset with the original images of the entire leaves and a second one containing only symptoms images. Details of each dataset are described in the following. Leaf dataset: It consists of the original images of the entire leaves. The images were labeled in relation to the predominant biotic stress of each leaf and its severity. Stress severity was calculated using the symptom and leaf segmentation mask using automatic image processing methods presented in Manso et al. (2019). For certain severity ranges, labels were assigned as follows: healthy (< 0:1%), very low (0.1% - 5%), low (5% - 10%), high (10% - 15%) and very high (> 15%). Symptom dataset: This dataset was created by cropping the isolated symptoms from the original images in a way that only single stress was present in each image. A total of 2147 symptom images were cropped. Each dataset is divided in training, validation and test. GL Manso, H Knidel, RA Krohling, JA Ventura (2019), A smartphone application to detection and classification of coffee leaf miner and coffee leaf rust. arXiv preprint arXiv:1904.00742



Universidade Federal do Espirito Santo


Agricultural Science, Computer Science, Artificial Intelligence, Plant Pathology, Deep Learning