Skip to Main content Skip to Navigation

Spectro-imagerie et apprentissage profond : application à la détection de maladies de plantes

Abstract : This thesis is the result of a collaboration between the LIRIS and LARIS laboratories and Carbon Bee, a French company focused on developing digital technology for agriculture. Carbon Bee develops a camera coupled with a deep learning algorithm in order to conduct spot spraying of crop protection products. This camera contains several sensors which allow for acquisitions in different wavelength ranges. It includes in particular an infrared sensor along with a snapshot hyperspectral spectrometer seldom studied until now : the Computed Tomography Imaging Spectrometer (CTIS). This sensor allows for a fast acquisition of rich spectral information. However, it is necessary to post-process this information via a reconstruction algorithm to make it understandable to the human eye. In this work, we have taken interest in the optimal use of these sensors for a case study with a high agronomic impact : the detection of apple scab. We focused at first on the analysis of the CTIS signal in the context of a binary classification between images of healthy and diseased leaves. We developed a procedure which allowed to bypass the reconstruction algorithm by training a neural network directly on raw CTIS images, an approach known as compressed learning. Using a novel neural architecture allowed us to achieve a classification performance higher than the one obtained following the classical reconstruction pipeline, while substantially reducing the related training and inference times. This study led to the development of several novel image simulators which allowed to compensate for the low number of annotated images, an oft-encountered hurdle in deep learning studies, especially when working with a new imaging system. While the work we have conducted on the CTIS images was carried out at the leaf scale, we afterward focused on a more demanding context, closer to the industrial challenges faced by Carbon Bee.We strove to improve scab detection at a pixel level in infrared images of leaf canopies; what ismore, with a limited quantity of annotated data. For this purpose, we developed several image simulators inspired by the latest trends in the plant sciences domain. In particular, we designed a canopy image simulator whose images enabled us to considerably reduce the number of annotated images necessary to conduct a segmentation in this context. Finally, the presence of several sensors in the camera paved the way to the combination of the information that they gathered, a process known as data fusion.We have explored several pathways within this framework.
Complete list of metadata
Contributor : DOUARRE Clément Connect in order to contact the contributor
Submitted on : Wednesday, June 8, 2022 - 9:43:49 AM
Last modification on : Thursday, June 9, 2022 - 3:37:07 AM


Files produced by the author(s)


  • HAL Id : tel-03690297, version 1


Clément Douarre. Spectro-imagerie et apprentissage profond : application à la détection de maladies de plantes. Vision par ordinateur et reconnaissance de formes [cs.CV]. Université Lumière Lyon 2, 2021. Français. ⟨tel-03690297⟩



Record views


Files downloads