Alessandro Benfenati, Paola Causin, Roberto Oberti

Università degli Studi di Milano

Crop protection from diseases through applications of plant protection products is crucial to secure worldwide food production. Nevertheless, sustainable management of plant diseases is an open challenge with a major role in the economic and environmental impact of agricultural activities. A primary contribution is expected to come from precision crop protection approaches, with treatments tailored to spatial and time-specific needs of the crop, in contrast to the current practice of applying treatments uniformly to fields.

A fundamental requirement for the full implementation of precision crop protection systems is the capability of automatically detect symptoms of disease at early stages to timely target the treatments on emerging infection spots and preventing their establishment and following epidemic expansion. In this respect, computer vision has an inherent great potential since symptoms of crop diseases very often cause a signature on plant organs which can be recognized by adequate image-analysis techniques. While computer vision approaches traditionally relied on human experts for the selection of the most relevant features to discriminate diseased from healthy plants, recent developments in machine learning disclosed new possibilities toward the automatic identification of relevant features.

Thanks to the unprecedented performance in image-recognition problems, Deep Learning (DL) methods have vigorously entered the domain of plant disease detection. In our recent research work, we explored DL approaches for automatic recognition of powdery mildew on cucumber leaves (see Fig.1), with focus on unsupervised techniques to overcome the need of large training sets of manually labelled images. Specifically, autoencoder (AE) networks were implemented for unsupervised detection of disease symptoms via an anomaly detection concept. Anomaly detection is the process of detecting data instances that deviate from a given set of samples. The idea we pursued is that an AE tailored to encode and decode a specific kind of data – in our case healthy leaves-, fails in encoding and decoding correctly other kinds of data – diseased leaves-, revealing an anomaly.

Fig. 1. RGB images of four samples of cucumber leaves from the dataset acquired in the greenhouse of University of Milano and used for training and evaluation of the anomaly detection approach. The leaves show different symptoms of the fungal disease known as powdery mildew, from very mild (top left) to severe, in clockwise order.

In a second research work, we explored an approach based on generative DL techniques to automatically generate (virtually unlimited) collections of synthetic images of plant leaves to be used for training. This effort aims to overcome the need for the enormous amounts of training images examples to avoid overfitting phenomena. Standard image augmentation methods, usually consisting in simple color and geometric transformations such as random rotations, translations, scaling or deformations of the original images, provide limited richness of the augmented dataset. We took inspiration from the synthesis of eye retina images for medical applications. Indeed, the fundus of the eye shares several characteristics with leaves: a fine network of hierarchically organized blood vessels (as the leaf veins) superposed to a colored background (as the tissue of the leaf blade). The proposed Leaf-to-Leaf (L2L) algorithm is organized in two steps: first it uses an AE architecture to generate synthetic leaf skeletons then it performs a translation via a Pix2pix framework, which uses conditional generator adversarial networks (cGANs) to reproduce the specific color distribution of the leaf blade, preserving leaf shape and venation pattern. Fig.2 shows an instance of the skeleton and the relative final colorized leaf obtained via such an approach.

Fig.2. Process of generation of a synthetic grapevine leaf. A: vein skeleton; B: fully synthetic colorized blade.

Related bibliography

[1] Benfenati A, Bolzi D, Causin P, Oberti R (2022) A deep learning generative model approach for image synthesis of plant leaves. PLoS ONE 17(11): e0276972.

[2] Benfenati A, Causin P, Oberti R, Stefanello G (2023) Unsupervised deep learning techniques for automatic detection of plant diseases: reducing the need of manual labelling of plant images, to appear on Journal of Mathematics in Industry (arXiv preprint arXiv:2112.11242)


%d bloggers like this: