Remote sensing on ViCoS Lab

Foundation models for Earth observation

Mon, 01 Jan 0001 00:00:00 +0000

Foundation models for Earth observation

Earth-observation data spans many sensors, spatial resolutions, spectral ranges, and acquisition conditions. Rather than expecting a single model to cover this diversity, we investigate how specialized foundation models can share knowledge and provide transferable representations for a range of downstream tasks.

Our work focuses on self-supervised and contrastive representation learning for optical and multispectral imagery. We explore cross-modal distillation from optical vision foundation models, the combination of complementary teachers, and efficient adaptation to new sensors without discarding useful pretrained knowledge. The goal is to build representations that remain effective across tasks such as semantic segmentation, classification, and change detection while reducing the need for large annotated datasets.

Remote sensing change detection

Mon, 01 Jan 0001 00:00:00 +0000

Remote sensing change detection

Remote sensing change detection aims to identify meaningful changes between images of the same geographical area acquired at different times. It supports applications such as environmental monitoring, land-use analysis, disaster assessment, and infrastructure planning.

Our research examines both supervised and unsupervised approaches. We study how fundamental choices, including pretrained representations, model architecture, training strategy, and evaluation protocol, affect performance and generalization. We are also developing methods that learn without manually annotated change masks by generating diverse, data-driven perturbations in latent feature spaces. This makes it possible to address rare or complex changes and extend the methods to different sensing modalities, including optical and synthetic-aperture radar imagery.

Satellite image time series

Mon, 01 Jan 0001 00:00:00 +0000

Satellite image time series

We are using time series analysis on satellite multi-spectral data (Sentinel-2) to identify different crops in Slovenia. Our focus is on early identification of crops in relation to temporal and spatial transferability of the models.

Publications

Multi-Year Time Series Transfer Learning: Application of Early Crop Classification

Matej Račič, Krištof Oštir, Anže Zupanc and Luka Čehovin Zajc

Remote Sensing, MDPI, 2024
Application of Temporal Convolutional Neural Network for the Classification of Crops on SENTINEL-2 Time Series

Matej Račič, Krištof Oštir, Devis Peressutti, Anže Zupanc and Luka Čehovin Zajc

International Archives of the Photogrammetry, Remote Sensing & Spatial Information Sciences, 2020

Elevation model analysis

Mon, 01 Jan 0001 00:00:00 +0000

Elevation model analysis

We are using deep learning for semantic segmentation to analyze LIDAR data of Slovenian landscape and identify terraced landscapes. Methodological challenges include noisy labels and unbalanced classes. We are also interested in transferability of the models to other regions.

Publications

Zaznavanje terasiranih pokrajin kot semantična segmentacija digitalnega modela višin

Anže Glušič, Rok Ciglič and Luka Čehovin Zajc

ERK2021, 2021
Towards the deep learning recognition of cultivated terraces based on Lidar data: The case of Slovenia

Rok Ciglič, Anže Glušič, Lenart Štaut and Luka Čehovin Zajc

Moravian Geographical Reports, 2024