Centre de documentation
scientifique

Contribution to object extraction in cartography : A novel deep learning-based solution to recognise, segment and post-process the road transport network as a continuous geospatial element in high-resolution aerial orthoimagery / Calimanut-Ionut Cira (2022)  

Public Titre : Contribution to object extraction in cartography : A novel deep learning-based solution to recognise, segment and post-process the road transport network as a continuous geospatial element in high-resolution aerial orthoimagery Type de document : Thèse/HDR Auteurs : Calimanut-Ionut Cira, Auteur Editeur : Madrid [Espagne] : Universidad politécnica de Madrid Année de publication : 2022 Importance : 227 p. Format : 21 x 30 cm Note générale : bibliographie Thèse de Doctorat en Topographie, Géodésie et cartographie, Universidad politécnica de Madrid Langues : Anglais (eng) Descripteur : [Vedettes matières IGN] Traitement d'image optique [Termes IGN] analyse d'image orientée objet [Termes IGN] classification par réseau neuronal convolutif [Termes IGN] extraction du réseau routier [Termes IGN] image aérienne [Termes IGN] orthoimage [Termes IGN] réseau antagoniste génératif [Termes IGN] réseau neuronal artificiel [Termes IGN] route [Termes IGN] segmentation sémantique Index. décimale : THESE Thèses et HDR Résumé : (auteur) Remote sensing imagery combined with deep learning strategies is often regarded as an ideal solution for interpreting scenes and monitoring infrastructures with remarkable performance levels. Remote sensing experts have been actively using deep neural networks to solve object extraction tasks in high-resolution aerial imagery by means of supervised operations. However, the extraction operation is imperfect, due to the nature of remotely sensed images (noise, obstructions, etc.), the limitations of sensing resolution, or the occlusions often present in the scenes. The road network plays an important part in transportation and, nowadays, one of the main related challenges is keeping the existent cartographic support up to date. This task can be considered very challenging due to the complex nature of the geospatial object (continuous, with irregular geometry, and significant differences in width). We also need to take into account that secondary roads represent the largest part of the road transport network, but due to the absence of clearly defined edges, and the different spectral signatures of the materials used for pavement, monitoring, and mapping them represents a great effort for public administration, and their extraction is often omitted altogether. We believe that recent advancements in machine vision can enable a successful extraction of the road structures from high-resolution, remotely sensed imagery and a greater automation of the road mapping operation. In this PhD thesis, we leverage recent computer vision advances and propose a deep learning-based end-to-end solution, capable of efficiently extracting the surface area of roads at a large scale. The novel approach is based on a disjoint execution of three different image processing operations (recognition, semantic segmentation, and post-processing with conditional generative learning) within a common framework. We focused on improving the state-of-the-art results for each of the mentioned components, before incorporating the resulting models into the proposed solution architecture. For the recognition operation, we proposed two framework candidates based on convolutional neural networks to classify roads in openly available aerial orthoimages divided in tiles of 256×256 pixels, with a spatial resolution of 0.5 m. The frameworks are based on ensemble learning and transfer learning and combine weak classifiers to leverage the strengths of different state-of-the-art models that we heavily modified for computational efficiency. We evaluated their performance on unseen test data and compared the results with those obtained by the state-of-the-art convolutional neural networks trained for the same task, observing improvements in performance metrics of 2-3%. Secondly, we implemented hybrid semantic segmentation models (where the default backbones are replaced by neural network specialised in image segmentation) and trained them with high-resolution remote sensing imagery and their correspondent ground-truth masks. Our models achieved mean increases in performance metrics of 2.7-3.5%, when compared to the original state-of-the-art semantic segmentation architectures trained from scratch for the same task. The best-performing model was integrated on a web platform that handles the evaluation of large areas, the association of the semantic predictions with geographical coordinates, the conversion of the tiles’ format, and the generation of GeoTIFF results (compatible with geospatial databases). Thirdly, the road surface area extraction task is generally carried out via semantic segmentation over remotely sensed imagery—however, this supervised learning task can be considered very costly because it requires remote sensing images labelled at pixel level and the results are not always satisfactory (presence of discontinuities, overlooked connection points, or isolated road segments). We consider that unsupervised learning (not requiring labelled data) can be employed for post-processing the geometries of geospatial objects extracted via semantic segmentation. For this reason, we also approached the post-processing of the road surface areas obtained with the best performing segmentation model to improve the initial segmentation predictions. In this line, we proposed two post-processing operations based on conditional generative learning for deep inpainting and image-to-image translation operations and trained the networks to learn the distribution of the road network present in official cartography, using a novel dataset covering representative areas of Spain. The first proposed conditional Generative Adversarial Network (cGAN) model was trained for deep inpainting operation and obtained improvements in performance metrics of maximum 1.3%. The second cGAN model was trained for image-to-image translation, is based on a popular model heavily modified for computational efficiency (a 92.4% decrease in the number of parameters in the generator network and a 61.3% decrease in the discriminator network), and achieved a maximum increase of 11.6% in performance metrics. We also conducted a qualitative comparison to visually assess the effectiveness of the generative operations and observed great improvements with respect to the initial semantic segmentation predictions. Lastly, we proposed an end-to-end processing strategy that combines image classification, semantic segmentation, and post-processing operations to extract containing road surface area extraction from high-resolution aerial orthophotography. The training of the model components was carried out on a large-scale dataset containing more than 537,500 tiles, covering approximately 20,800 km2 of the Spanish territory, manually tagged at pixel level. The consecutive execution of the resulting deep learning models delivered higher quality results when compared to state-of-the-art implementations trained for the same task. The versatility and flexibility of the solution given by the disjointed execution of the three separate sub-operations proved its effectiveness and economic efficiency and enables the integration of a web application that alleviates the manipulation of geospatial data, while allowing for an easy integration of future models and algorithms. Resuming, applying the proposed models resulted from this PhD thesis translates to operations aimed to check if the latest existing aerial orthoimages contains the studied continuous geospatial element, to obtain an approximation of its surface area using supervised learning and to improve the initial segmentation results with post-processing methods based on conditional generative learning. The results obtained with the proposed end-to-end-solution presented in this PhD thesis improve the state-of-the-art in the field of road extraction with deep learning techniques and prove the appropriateness of applying the proposed extraction workflow for a more robust and more efficient extraction operation of the road transport network. We strongly believe that the processing strategy can be applied to enhance other similar extraction tasks of continuous geospatial elements (such as the mapping of riverbeds, or railroads), or serve as a base for developing additional extraction workflows of geospatial objects from remote sensing images. Note de contenu : 1- Introduction 2- Methodology 3- Theoretical framework 4- Litterature review 5- Road recognition: A framework based on nestion of convolutional neuronal networks and transfer learning to regognise road elements 6- Road segmentation: An approach based on hybrid semantic segmentation models to extract the surface area of rod elements from aerial orthoimagery 7- Post-processing of semantic segmentation predictions I: A conditional generative adversial network to improve the extraction of road surface areas via deep inpainting operations 8- Post-processing of semantic segmentation predictions II: A lightweight conditional generative adversial network to improve the extraction of road surface areas via image-to-image translation 9- An end-to-end road extraction solution based on regonition, segmentation, and post-processing operations for a large-scale mapping of the road transport network from aerial orthophotography 10- Conclusions Numéro de notice : 24069 Affiliation des auteurs : non IGN Thématique : IMAGERIE Nature : Thèse étrangère Note de thèse : Thèse de Doctorat : Topographie, Géodésie et cartographie : Universidad politécnica de Madrid : 2022 DOI : 10.20868/UPM.thesis.70152 En ligne : https://doi.org/10.20868/UPM.thesis.70152 Format de la ressource électronique : URL Permalink : https://documentation.ensg.eu/index.php?lvl=notice_display&id=102113