Medellín - Minas - Doctorado en Ingeniería - Sistemas · 2022
Automatic generation of GIS vector Layers from orthomosaics using deep learning
This thesis presents a three methods pipeline for extraction of point, line, and polygon vector objects from orthomosaics using a deep generative model as an alternative to the default semantic segmentation approach. The first method consists of two workflows, the vector ground truth is acquired by manual digitalization of certain objects or from Open Street Maps. Raster layers input are spectral and geometrically augmented, both inputs are then tessellated and paired into image-masks that pass through an imbalance checking step. Balanced dataset is then random split into a final dataset. Conditional and unpaired generative models are compared and pix2pix is chosen by its better results on image to mask translation. Results of the chosen model on different datasets and configurations are reported on the mIoU metric. A batch size of 10 and datasets of 1000 image-masks pairs of 512x512 pixels, with overlapping augmentation showed the best quantitative results. Height of objects from the DSM, and VARI index contribute to decrease variance of discriminator and generator losses. Producing synthetic data is the horsepower of generative models, so a double image to mask translation is used to improve resultant masks in terms of continuity and uniform width. Double image to mask translation model is trained with a dataset of equal size masks of 1 meter called primitive masks, that are obtained by a buffer distance parameter. This cleaning procedure showed to improve resultant masks, that are then converted to vector and measured by quantity, length, or area against vector ground truth, using a proposed metric for map creation called “The average geometry similarity (AGS)”.
Texto completo 138 páginas con texto
Leer la tesis completa Ficha en el repositorio
Contenido
- Introductionp. 8
- Problem Descriptionp. 11
- Research Questionsp. 14
- Objectivesp. 14
- Contribution and Academic Productsp. 14
- An Overview of GIS Vector Layer Extraction from Orthomosaicsp. 17
- State of the Artp. 17
- Manual Digitalization and Object Attributionp. 20
- Semi-automatic Methods for Object Extraction from Aerial Imagesp. 21
- Deep Learning for Image Semantic Segmentationp. 23
- Metrics for Semantic Segmentationp. 28
- Deep Generative Modelsp. 29
- Modelp. 64
- Paired Data Methodp. 37
- Generative Method for Image to Mask Translationp. 38
- Post-processing Methodp. 40
- A Method for Producing Paired Datasetsp. 42
- A Workflow for Producing Paired Datasetsp. 42
- Experiments and Resultsp. 51
- Conclusions and Future Workp. 58
- A Generative Method for Objects Extraction from Orthomosaicsp. 59
- Conditional and Unpaired Image to Image Translationp. 33
- Experiments and Resultsp. 51
- Conclusions and Future Workp. 58
- A Post-processing Methodp. 86
- Primitive Masksp. 86
- Double Image to Mask Translation and Double Paired Datasetsp. 87
- Experiments and Resultsp. 51
- Conclusions and Future Workp. 58
- Validation of the Proposed Methodologyp. 95
- Proposed Performance Metricp. 95
- Experiments and Resultsp. 51
- Conclusions and Future Workp. 58
- Conclusionsp. 99
- Bibliographyp. 101
- Annex 1. Python Scriptsp. 110
- Annex 2. Remote Sensing Datasets for Semantic Segmentation of Orthomosaicsp. 128