מושגי ליבה
A novel Geographic Heterogeneity-Aware Federated Learning (GeoFed) framework to address privacy-preserving collaborative remote sensing semantic segmentation by alleviating class-distribution heterogeneity and object-appearance heterogeneity.
תקציר
The paper proposes a Geographic Heterogeneity-Aware Federated Learning (GeoFed) framework to address the challenges of privacy-preserving collaborative remote sensing semantic segmentation (RSS).
Key highlights:
- Remote sensing data from different institutions often exhibit strong geographic heterogeneity, including class-distribution heterogeneity and object-appearance heterogeneity. This poses challenges for effectively applying federated learning (FL) in RSS.
- The GeoFed framework consists of three main components:
- Global Feature Extension (GFE) and Tail Regeneration (TR) modules to alleviate class-distribution heterogeneity.
- Essential Feature Mining (EFM) strategy to alleviate object-appearance heterogeneity.
- An overall loss function that integrates these components.
- Extensive experiments on three public datasets (FBP, CASID, Inria) demonstrate that GeoFed consistently outperforms state-of-the-art methods in privacy-preserving collaborative RSS.
סטטיסטיקה
The paper reports the following key metrics:
On the FBP dataset, the proposed GeoFed method achieves an average mIoU of 66.54%, outperforming the baseline FedAvg method by 3.51%.
On the CASID dataset, GeoFed achieves an average mIoU of 55.75%, outperforming the previous state-of-the-art method Elastic by 0.37%.
On the Inria dataset, GeoFed achieves an average IoU of 52.49%, outperforming the previous state-of-the-art method Elastic by 2.51%.
ציטוטים
"Remote sensing semantic segmentation (RSS) is an essential task in Earth Observation missions. Due to data privacy concerns, high-quality remote sensing images with annotations cannot be well shared among institutions, making it difficult to fully utilize RSS data to train a generalized model."
"Remote sensing images in various institutions often exhibit strong geographical heterogeneity. More specifically, it is reflected in terms of class-distribution heterogeneity and object-appearance heterogeneity."
"Considering the aforementioned issues, we propose a novel Geographic Heterogeneity-Aware Federated Learning (GeoFed) framework to address privacy-preserving RSS."