Visual Intelligence Seminar #103: From Satellite Data to Reliable Spatial Intelligence

Presented by Amanda Berg, Senior Staff R&D Scientist at Vantor Sweden AB / Adjunct Associate Professor at Linköping University

Abstract

Earth-observation systems must transform massive and heterogeneous satellite data into information that is timely, reliable, and operationally useful. In this seminar, I will present two complementary perspectives on this challenge. I will begin with an overview of my work at Vantor, focusing on the processing of large geospatial cells with machine-learning methods such as semantic segmentation. I will discuss practical challenges related to scale, spatial context, multimodal data, computational efficiency, and the integration of research algorithms into production systems.

I will then present our work within the remote sensing unit at CVL focusing on how remote-sensing models can remain reliable across sensors, geographic regions, seasons, and application domains. Recent projects have addressed wildfire detection and predictability, the value of multispectral and cross-sensor information, and building-centric multimodal reasoning for disaster assessment.

These two perspectives have a common theme: reliable remote sensing intelligence requires more than strong benchmark performance. It also requires robustness to sensor and domain shifts, effective use of complementary observations, spatially grounded reasoning, and algorithms that can scale from research datasets to large operational geospatial environments.