Skip to Main content Skip to Navigation
Master Thesis

Analyse comparative de la sensibilité de méthodes pour l’étude de la prédiction de la fraction verte au travers de la résolution et de la diversité d’échelles spatiales

Abstract : High throughput phenotyping is developing rapidly for varietal selection and precision farming applications. The green fraction, the fraction of green pixels in an image, is one of the most useful features for tracking vegetation development that can be extracted from high-resolution images taken from a range of systems. However, this range, as varied as it may be, results in a diversity of image resolution and feature sizes, which undeniably implies an alteration in the accuracy of the estimated green fraction, due to the mixed pixel fraction, and the fact that relevant information on the objects under study varies at different spatial scales. A first study phase consisted in estimating the average size of the elements composing the image through variograms, then artificially simulating a degradation of the spatial resolution of the images. The objectives of the proposed study are to evaluate and compare the performance of different types of approaches in order to estimate the relative sensitivity of green fraction prediction to the resolution and diversity of spatial scales. Methods based on deep learning have been shown to be the most effective at high resolution. For moderate degradation, pixel-based methods take over and seem to minimize information loss. Finally, a strongly degraded resolution implies the use of regression methods to directly estimate the green fraction. A second study focused on the learning of scale-dependent patterns is also presented, it results that an insertion of object size diversity allowed better results thanks to a better representation of the object as a whole, for deep learning, and conversely for pixel-based methods.
Document type :
Master Thesis
Complete list of metadatas

https://dumas.ccsd.cnrs.fr/dumas-02968772
Contributor : Agrocampus Ouest <>
Submitted on : Friday, October 16, 2020 - 9:36:18 AM
Last modification on : Sunday, February 21, 2021 - 3:22:30 AM

File

serouart2020_Serouart_Mario_Sc...
Files produced by the author(s)

Licence


Distributed under a Creative Commons Attribution - NonCommercial - NoDerivatives 4.0 International License

Identifiers

  • HAL Id : dumas-02968772, version 1

Collections

Citation

Mario Serouart. Analyse comparative de la sensibilité de méthodes pour l’étude de la prédiction de la fraction verte au travers de la résolution et de la diversité d’échelles spatiales. Sciences du Vivant [q-bio]. 2020. ⟨dumas-02968772⟩

Share

Metrics

Record views

3

Files downloads

3