Contributions of deep learning methods for automatic recognition ofland use and land cover patterns and objects by remote sensing in tropical environments

Apports des méthodes d'apprentissage profond pour la reconnaissance automatique des modes d'occupation des sols et d'objets par télédétection en milieu tropical

Guillaume Rousset (1)

1 UNC - Université de la Nouvelle-Calédonie

The New Caledonian landscape is changing rapidly with the development of new mining projects, the intensification of urbanization and the impacts of extreme climatic events such as cyclones. With the democratization and accumulation of satellite data and the advent of artificial intelligenc methods, the implementation of automatic detection methods is becoming an essential tool for documenting and monitoring these changes to the scale of a territory in a regular, rapid and objective manner. Among these methods, deep learning has shown effective results on co mplex problems, especiall on image processing using dense convolutional neural networks. Considering the constraints related to the processing of satellite imagery and the problems related to learning algorithms, the objective of the thesis is multiple: to contribute to the adaptation of deep learning techniques to remote sensing problems on several key points of the processing chain; to estimate the performance of these techniques compared to the methods commonly used in the field of remote sensing; and to dev elop automatic detection methods to deliver reliable indices to any exploitation of satellite imagery. This thesis focused on three applications: 1) land cover and land use detection on very high resolution data; 2) the detection of land cover in New Caledonia at an annual frequency on high resolution data; 3) and the detection of palm trees in the Pará region of Brazil using computer simulated data. For the first application, a reference dataset based on SPOT 6 satellite data was manually created and made available to the scientific community in order to compare land cover class detection techniques in tropical island environments. Dense neural networks show better performance, especially in the context of land cover detection which requires a higher level of conceptualization of the environment. For the second application, an automatic land cover detection chain, based on a dense neural network fed by Sentinel-2 data, has been realized. This coverage is compared to coverage obtained by semi-automatic methods in the South Province of New Caledonia. The model performs equally well ove a few test areas, but additional field data are required to confirm the reliability over the whole New Caledonian territory. Finally, for the last application, the originality of the research work consisted of testing the contribution of synthetic sa tellite images in the learning base. For this purpose, images of the palm tree were constructed from a radiative transfer model. The use of these synthetic images in addition to the Pleiades images significantly improved the overall accuracy of the models.

Mots clés: Deep Learning; Remote Seruing; Land Cover; Land Use; Machine Learning; Simulated Data; Object Detection

Mots clés: Cartographie numérique des sols -- Nouvelle-Calédonie; Apprentissage profond -- Nouvelle-Calédonie; Imagerie satellitaire -- Nouvelle-Calédonie; Apprentissage en profondeur; Utilisation à distance; Couverture terrestre; L'utilisation des terres; Classification; Apprentissage automatique; Données simplifiées; Détection d'objet

Domaines: Traitement des images [eess.IV]; Apprentissage [cs.LG]; Ingénierie de l'environnement

Guillaume Rousset. Apports des méthodes d'apprentissage profond pour la reconnaissance automatique des modes d'occupation des sols et d'objets par télédétection en milieu tropical. Traitement des images [eess.IV]. Université de la Nouvelle-Calédonie, 2021. Français. ⟨NNT : 2021NCAL0006⟩. ⟨tel-03725622⟩