Data-Efficient ConvNeXt Pretraining for High-Resolution Sentinel-1 GRD SAR Representation Learning

Name: Data-Efficient ConvNeXt Pretraining for High-Resolution Sentinel-1 GRD SAR Representation Learning
Uploaded: 2026-10-20T12:32:34.604779+00:00
Duration: 773 s
Description: SparK-SAR is a sparse convolutional masked pretraining framework using ConvNeXt to learn from limited Sentinel-1 Ground Range Detected (S1GRD) data efficiently. It introduces a domain-adapted encoder and lightweight decoder that significantly outperform frozen Vision Transformer models with fewer parameters, showing that classification performance saturates quickly with less data.

SparK-SAR is a sparse convolutional masked pretraining framework using ConvNeXt to learn from limited Sentinel-1 Ground Range Detected (S1GRD) data efficiently. It introduces a domain-adapted encoder and lightweight decoder that significantly outperform frozen Vision Transformer models with fewer parameters, showing that classification performance saturates quickly with less data.

Open the full episode