Jonas Klotz

FedX: Explanation-Guided Pruning for Communication-Efficient Federated Learning in Remote Sensing

B. Büyüktaş^†, Jonas Klotz^†, B. Demir

^† Equal contribution

IEEE JSTARS · 2026

Federated learning for remote sensing suffers from high communication costs. We propose explanation-guided model pruning to reduce bandwidth while preserving accuracy.

On the Effectiveness of Methods and Metrics for Explainable AI in Remote Sensing Image Scene Classification

Jonas Klotz, T. Burgert, B. Demir

IEEE JSTARS · 2025

We systematically evaluate XAI methods and their evaluation metrics for scene classification in remote sensing, revealing critical gaps between method capability and metric design.

Communication-Efficient Federated Learning Based on Explanation-Guided Pruning for Remote Sensing Image Classification

Jonas Klotz, B. Büyüktaş, B. Demir

IGARSS · 2025

Using gradient-based explanations to identify and prune redundant parameters, we reduce federated communication overhead for remote sensing classifiers without accuracy loss.

A Label Propagation Strategy for CutMix in Multi-Label Remote Sensing Image Classification

T. Burgert, K. N. Clasen, Jonas Klotz, T. Siebert, B. Demir

IEEE JSTARS · 2025

We introduce a label-aware mixing strategy that improves multi-label classification robustness in satellite imagery by propagating soft labels through augmented training pairs.

Land Sealing Dataset

Data Science for Social Good initiative building a dataset to monitor German land use by digitizing Bauleitpläne (building development plans) from geoportals using OCR and NLP.

Chess Vision

Algorithm for detecting chessboards and recognizing chess pieces from photos, using line-based grid detection and CNNs (Xception, ResNet, MobileNet). Designed for a mobile app to help players analyze games.