This study examines if it is possible to use pretrained neural models to make a detailed map in a set of case studies in Sweden. The models are taken from ESRI Living Atlas and are trained with data from the USA. The models are given orthophotos as input from three different parts of Sweden to make predictions from. The model results are combined to extract the best predictions of the presence of different object types and land cover classes, such as forests, open areas, lakes, roads and buildings. To enhance the map’s topographic details, a height model is employed, and certain features are added to mimic a conventional topographic map. A brief theoretical section elucidates the fundamentals of deep learning. Additionally, specific technical terms are defined and their application within the context of model usage and successful architectural designs is discussed. The performance metrics of the neural models are explained and subsequently applied in validating the results. The resulting map is validated versus an existing map provided by Lantmäteriet. The results are analyzed, and some shortcomings and advantages are highlighted and discussed. The final result shows that it is possible to make a detailed map based on automatic processes by combining the results of pretrained neural models. The map have some deviations relative the existing maps, the overall accuracy was 0.94, but the deviations seems smaller if you visually compared it with the ortofoto. The maps quality is at least of such good quality that it could serve a visual guide when making yourself a picture of chosen terrain in Sweden.
| Identifer | oai:union.ndltd.org:UPSALLA1/oai:DiVA.org:kau-100134 |
| Date | January 2024 |
| Creators | Algesten, Erik |
| Publisher | Karlstads universitet, Institutionen för miljö- och livsvetenskaper (from 2013) |
| Source Sets | DiVA Archive at Upsalla University |
| Language | Swedish |
| Detected Language | English |
| Type | Student thesis, info:eu-repo/semantics/bachelorThesis, text |
| Format | application/pdf |
| Rights | info:eu-repo/semantics/openAccess |
Page generated in 0.002 seconds