Webinar Announcement: Developing a Deep Learning Approach to Map Pan-Arctic Infrastructure from Maxar Satellite Imagery

Date: 
19 January 2023

Webinar Announcement
Developing a Deep Learning Approach to Map Pan-Arctic Infrastructure from Maxar Satellite Imagery

Permafrost Discovery Gateway Webinar Series
Thursday, 26 January 2023
9:00 a.m. Alaska Time

For more information, go to:
https://arcticdata.io/catalog/portals/permafrost/Stay-Connected#upcoming...


The Permafrost Discovery Gateway announces their upcoming webinar, Developing a Deep Learning Approach to Map Pan-Arctic Infrastructure from Maxar Satellite Imagery, featuring Elias Manos (University of Connecticut). This webinar will take Thursday, 26 January 2023 at 9:00 a.m. Alaska Time.

Webinar Abstract:

Comprehensive and up-to-date analysis-ready geospatial data on pan-Arctic infrastructure is lacking, hampering risk assessment efforts that attempt to quantify the socioeconomic impacts of permafrost thaw-related natural hazards on the built environment. A recent study addresses this data gap by producing the first pan-Arctic satellite-based record of infrastructure and anthropogenic impacts within 100 km of Arctic coasts at a 10 m spatial resolution, mapping infrastructure from Sentinel-1 and Sentinel-2 imagery using machine learning and deep learning models. In this ongoing study, we attempt to complement and improve upon this data product by developing a deep learning framework to map pan-Arctic infrastructure at a sub-meter spatial resolution using Maxar commercial satellite imagery, which presents a number of unique challenges. Semantic complexity of objects at sub-meter spatial resolution requires a plausible classification scheme that generalizes across the thematic and geographic variability in Arctic infrastructure. The amount of time required to create a circumpolar training dataset requires the integration of numerous open-source geospatial datasets to speed up the process. Model training and testing sites must be carefully selected in order to account for variables including settlement type, structure size and shape, density of building distribution, rooftop design and material, and natural environmental factors. Early stages of the study show promising performance of a U-Net++ model trained to detect various buildings types, roads, airport runways, gravel pads, pipelines, and storage tanks in rural, medium-density, urban, and industrial settings across Alaska, Russia, and Canada.

For more webinar connection information, go to:
https://arcticdata.io/catalog/portals/permafrost/Stay-Connected#upcoming...