Development of a deep-learning based system for disease symptoms detection over crop leaves images

Abstract

Family farming represents a critical segment of Brazilian agriculture, involving more than 5 million properties and generating 74% of rural jobs in the country. Yield losses caused by crop diseases and pests can be devastating for small-scale producers. However, successful disease control requires correct identification, which challenges smallholders, who often lack technical assistance. The present work proposes a system that alerts smallholder farmers and phytopathology experts about possible crop disease outbreaks, enabling a faster diagnosis and intervention. To this extent, we detect disease symptoms in images of plant leaves taken by farmers directly in the field using a mobile phone app developed for this purpose. The implemented module is part of a service platform connecting producers and experts, designed in partnership with phytopathology professionals from the Federal Rural University of Pernambuco (UFRPE). The work uses deep learning and Convolutional Neural Networks to perform the image classification. The classification experiments were applied over a dataset composed of leaf images of grape crops cultivated in the state of Pernambuco, whose image collection was also part of the present work. Therefore, pictures taken under field conditions sometimes present low quality, which decreases the classification performance. Thus, we also classify the images regarding their quality, to exclude challenging images from disease detection and reduce the number of erroneously classified images entering the database. The multi-label technique is employed in this scenario, enabling a single neural network model to classify whether a leaf picture reveals crop disease symptoms and whether they present good enough quality to do so reliably. The multi-label mechanism is also a promising approach to include additional picture properties in the future, like disease agents. The developed classification system achieves a recall value of 97.6% for symptom detection and a precision value of 94.8% for image quality classification.