Estimering af volumen vha. Sentinel-data

Session

Nørderi

Abstract

Præsentationen handler om ombygningen af en regressionsmodel for estimering af volumen af menneskeskabte strukturer ud fra data fra Sentinel-satellitterne. Desuden betragtes effekten af satellitternes kredsløbsretning og sammenhæng med SAR (coherence) med modellen.

Data fra Sentinel 1- og 2-satellitterne er indsamlet i foråret 2020 over Region Midtjylland. Data bliver brugt til at træne modellen. Valideringsdata består af bygningsdata fra GeoDanmark samt de nationale terrænkort til volumenbestemmelsen.

Modellen kan integreres i beslutningsunderstøttende systemer til brug for befolkningsestimater, forandringsudpegning og evalueringssystemer i forbindelse med katastrofer.

Over store dele af det afrikanske kontinent er Sentinel 1-data kun tilgængelig ved en enkelt kredsløbsretning. Betydningen for nøjagtigheden af volumenestimaterne ved kun at have adgang til den ene kredsløbsretning i Afrika bliver afprøvet vha. datasættene fra Danmark.

Målgruppe

Målgruppen er deltagere ved Virtuelle Kortdage 2020, der interesserer sig for jordobservation, data fra Copernicus-programmet, samt udviklingsstudier i Afrika.

Yderligere uddybning af abstract

This study investigates a regression method for estimating the volume of human-made structures from the Sentinel satellites and the effect of orbital direction and coherence calculations on the accuracy. Sentinel 1 and 2 data sensed during 2020 over the Central Denmark Region is used to train the model. Ground truth is the GeoDanmark building dataset with volumes calculated using the Danish national point cloud dataset. The volume estimates can be integrated into decision support systems for population estimates, change detections, and building and valuation systems.

Over large swaths of the African continent, Sentinel 1 data is only available for one orbital direction. The impact of having access to only one orbital direction on the accuracy of structural volume estimates is tested. It is important to many urbanisation studies, which rely on earth observation data, to know if there is a significant decrease in model prediction accuracy due to the single orbital direction available in the Global South. The model is trained using Danish data as it is possible to generate a high-quality national human-made structural volumes dataset. The dataset created for this study also contains population numbers, allowing further investigation into the correlation with population density.

Estimates are made using Sentinel 1 data and a combined approach with Sentinel 2 data. VIIRS data is included in the tests as a baseline. As Sentinel 2 is dependent on mostly cloud-free imagery to make valid predictions, it is not ideal for quick assessments in disaster management scenarios, especially in areas which are often covered in clouds, such as Denmark during the winter months. Transfer learning of the models trained in the study can help future studies of urban spaces in the Global South, incorporating data from the Sentinel satellites and where only one orbital direction is available.