In their pursuit to address the significant challenges posed by pests and diseases in agriculture, the authors of the Rust and Leaf Miner in Coffee Crop dataset have developed an algorithm designed for the detection of rust (Hemileia vastatrix) and leaf miner (Leucoptera coffeella) within coffee leaves (Coffea arabica). Their solution extends to quantifying disease severity, facilitated through a mobile application functioning as a high-level interface for model inferences.

Within the realm of coffee cultivation, numerous phytosanitary issues abound, including bacteria, fungi, viruses, nematodes, and pests. Among these, rust has emerged as a particularly impactful concern since its inception in East Africa in 1869. Rust inflicts severe harm, leading to the premature defoliation of coffee trees and substantial production declines in subsequent years. Reports indicate production reductions of up to 34%. Similarly, leaf miner infestations, which began in Africa around the mid-1850s, have now become the most prevalent pest globally. Leaf miners damage the leaf epidermis, causing malnourishment and desiccation, resulting in production losses of up to 40%.

To combat these phytopathologies, the authors have turned to artificial intelligence, a field that has witnessed remarkable advancements in recent years. These advancements, particularly in image recognition, owe their success not only to more robust hardware and abundant data but also to the development of new and improved algorithms, including convolutional neural networks (CNNs). CNNs are particularly well-suited for tasks involving image analysis as they can learn and generalize features from images effectively.

Data Acquisition

The dataset for this study was meticulously assembled through the random collection of coffee leaves afflicted with rust and leaf miners from a farm located in Brazil. These leaves were then photographed in a laboratory against a white background using a smartphone camera operating in the RGB color system, which is the standard color space adopted by digital cameras. The RGB color system has demonstrated satisfactory performance in machine-learning applications when compared to other color spaces.

The resulting dataset comprises 285 samples of rust-infested leaves and 257 samples of leaves affected by the leaf miner. Initially, the images were captured at a resolution of 4000 x 2250 pixels. However, to align with the standards set by Liu et al. for Single Shot MultiBox Detector (SSD) and to optimize training time and computational resources, the images were resized to dimensions of 300 x 300 pixels and 512 x 512 pixels.

Creating a labeled dataset poses one of the central challenges in object detection techniques. It necessitates the manual annotation of object locations within images, often involving the delineation of bounding boxes. This process is not only resource-intensive but also time-consuming and costly, particularly when numerous objects per image need to be labeled, specialized knowledge is required, or the dataset is extensive.

Data Labeling

The authors describe that the initially collected photos underwent manual labeling using LabelImage software. However, to streamline this labor-intensive process and expedite future dataset expansion, an algorithm was devised for automatic image labeling, adopting a semi-supervised approach. Consequently, for each new unlabeled image, a pre-trained model, fueled by a limited number of strongly-labeled images, inferred the objects of interest.

The automatically generated labels underwent a meticulous review process due to inherent inaccuracies in detecting objects within certain images. Unaddressed inaccuracies could potentially diminish model accuracy. Nevertheless, this revised labeling process, albeit more rapid and cost-effective, continues to demand less human effort compared to conventional approaches.