Workshop + Hackathon: Neural Networks for remote sensing data classification

This workshop is partially funded by the NHR and supported by ForestCARE.

Hackathon:  Classification of tree damages from RGB aerial imagery using deep neural networks

Hauke Kirchner (GWDG) and Selina Schwarz (KIT) will present the use of the High-Performance Cluster as well as the Training & Remote Sensing data set in short online presentations.

In the 60-day period until the Expert Workshop, attendees have the chance to compete for prizes in the value of 200€ as teams or individuals and try out their ideas for the classification of tree damages.

Expert Workshop:  Neural Networks for remote sensing data classification

Connect with other Remote Sensing Experts and Computer Scientists at our 2-day workshop in Berlin.

Enjoy the talks from fellow researchers:

Gencer Sümbül (TUB, to be confirmed)
Most of the deep learning (DL) based representation learning (IRL) methods require the availability of a set of high quantity and quality of annotated training remote sensing (RS) images, which can be time-consuming, complex, and costly to gather. To reduce labeling costs, publicly available thematic maps, automatic labeling procedures or crowdsourced data can be used. However, such approaches increase the risk of including label noise in training data. This potentially leads to overfitting, sub-optimal learning procedures, and thus inaccurate characterization of RS images. In this talk, a general overview of scientific problems related to DL-based IRL under noisy labels will be initially discussed. Then, our recent investigations that can address these issues will be presented, while particular attention will be given to our generative reasoning Integrated IRL approach that can learn deep features of RS images robust to label noise independently of the type of annotation, label noise, neural network architecture, loss function or learning task.

Margot Verhulst (KUL)
Margot’s work focuses on forest monitoring in Flanders (Belgium) by combining Sentinel-2 time series and data from the National Forest  Inventory. The trained models treat the temporal dimension of the data in different ways and help investigate the transferability of the models between different years. This includes dominant tree species classification on the plot level as well as the prediction of leaf type (conifer/broadleaf) fractions and tree species fractions. As an introduction to the expert workshop on July 27th, attendees can use the opportunity to get a first-hand impression of Deep Learning workflows and their practical application in the context of remote sensing.

Michael Reuss (TUM)
During his Master’s, Michael built a Deep Neural Network for disturbance detection in forests from satellite image time series. He scaled his method to be applied for nationwide surveying using a large series of LandSat data and will share his finds on how to maximize the efficiency of large

Selina Schwarz (KIT)

We will discuss the results of the hackathon and provide the opportunity to discuss your specific challenges and pieces of code in our coworking-space in a relaxed atmosphere.

There are limited places available for the expert workshop in July.