13 Février – Thesis defense - Florian Rançon
10 h Amphi JP. Dom - Laboratory IMS (building A31) / Talence campus
Color and hyperspectral imagery for detection and characterization of grapevine wood diseases.
Grapevine wood diseases in the vineyard are responsible for significant economic losses in the wine industry. These diseases of fungal origin are caracterised by a degradation of the wooded part of the plant material and by the erratic appearance of characteristic symptoms on the leaf part. This thesis is dedicated to the study of these diseases (mainly esca disase) using two imaging sensors and proximal sensing.
The issue of visible symptom detection is first addressed using an RGB color sensor to acquire an image for each plant automatically or semi-automatically. The recognition of symptoms is approached in two stages, firstly by considering the classification at leaf-scale and then the detection at the plant-scale. The particularity of this study is the inclusion of confounding factors in the classification problem, taking advantage of the shape information of esca symptoms to differentiate them from other disorders and diseases. For this purpose, a comparison between SIFT approaches and recent transfer learning approaches is then conducted. The results then lead us to consider a simple deep learning architecture (RetinaNet) for the detection of the symptoms on the images, making it possible to estimate a level of disease severity for each vineplant.
The second sensor used, a hyperspectral camera covering the spectrum from 500 nm to 1300 nm, tries to tackle a more experimental problem, namely the spectral behavior of the diseased plants which may lead to early detection of diseased plants without foliar symptoms. An experimental protocol and a database of spectra are then formed for the occasion. The dimensionality reduction methods make it possible to exploit the hyperspectral information or even to isolate the wavelengths associated with each class. However, the data do not allow, for the measured wavelength range and in the field acquisition conditions, to perform early detection of the disease on the plant without symptoms.
The differences and similarities between each of these two applications, in terms of database constitution, algorithms, difficulties and application potential in real conditions are discussed throughout the manuscript.