A deep learning approach for detection and localization of leaf anomalies

Instances are the increasing demand for food due to the growth of the world population [7], as well as the impact of practices which are detrimental for the ecosystem [10]. In these contexts, precision agriculture has recently attracted a lot of interest since playing a significant role in the development of advanced techniques that optimize the soil productivity in a sustainable way [39, 20]. The general goal is to preserve the stability of the ecosystem while fostering the reuse of the soil for future produce. The strong interest in this new way of conceiving agriculture justifies the spread of innovative start-ups, of services devoted to eco-friendly practices, and of software solutions which allow farmers to accurately estimate yields on a simple smartphone or tablet (see, for instance, [2, 3, 4, 5]). In particular, the availability of higher-quality measurements, offered by advanced in field-sensors as well as by satellite or drone data, supported the proposal of breakthrough software solutions using modern deep learning algorithms [15, 19, 9, 36, 30]. In this paper, we focus on the detection of possible diseases in crops. Anomaly detection in plants represents a pivotal procedure in agriculture since an early detection of the disease enables a timely intervention to prevent the anomaly from spreading to the rest of the plant. Additionally, a precise disease localization allows confining the use of pesticides and other treatments to small strategically selected areas of the plant.