20 Décembre – Thesis defense - Florent Abdelghafour
10 h Amphi Jean-Paul Dom - laboratory IMS / building A31 (Talence campus)
Potential of on-board color imaging for the detection and mapping of fungal diseases affecting grapevine.
Vine downy mildew is a plant pathology of fungal origin that is particularly worrying for the wine industry. The objective of this thesis is to study the potential of on-board colour imaging to estimate the health status of vineyards affected by downy mildew at an intra-plot scale. The proposed solution aims to assist epidemiological monitoring networks in estimating health risks and recommending chemical control plans. In practice, the image processing chain built is dedicated to the detection, counting and measurement of symptomatic tissues of downy mildew. This chain is designed to process images acquired directly in the plot under wine-growing working conditions.
The processing chain is based on joint structure-colour representations and probabilistic models of the tissue classes present in the vines studied. It operates in three steps: formulating descriptors to extract the characteristic and discriminating properties of each class; modelling the statistical distributions of these descriptors in each of the classes; and assigning each pixel to a class according to its likelihood regarding the models. The descriptors combine the local structure tensor (LST) with colorimetric statistics calculated in the vicinity of the pixel under consideration. To take into account the specific nature of the LST, the descriptors are transformed to be represented in the log-Euclidean space. In this space, it becomes possible to model the classes of tissues of interest with distributions of multivariate Gaussian mixtures of structure-color features. Finally, the classification is carried out by a Maximum A Posteriori (MAP) decision.
This processing chain is initially applied to images of healthy vines. This involves segmenting an image into organ classes (foliage, clusters or inflorescences and stems). The classifications obtained are very efficient. In addition, the processing chain is robust to the adjustment of the main hyper-parameters.
In a second step, the processing chain is adapted to analyse images showing symptoms of mildew as well as confounding factors such as necrosis, discoloration, deficiencies, mechanical wounds. The decision method is based on a reconstruction of symptoms using a region growing process. The criteria used are based on the structure-colour representations and probabilistic models already defined. The new treatment line allows reliable detection of downy mildew symptoms and estimation of the surface area of affected tissues.