Convnext improves sentinel 1 sar image analysis with less data and compute
Data-Efficient ConvNeXt Pretraining for High-Resolution Sentinel-1 GRD SAR Representation Learning
Computer Vision and Pattern Recognition
Summary
It is hard to get large labeled datasets for satellite images from Sentinel-1, which makes teaching computers to understand these images difficult. The authors designed SparK-SAR, a new training method that uses a special type of neural network called ConvNeXt, which processes images more efficiently with fewer resources. Their approach achieves better results than existing methods on several satellite image tasks, even with less data and smaller model sizes. They also studied how much data is needed for good performance and found different tasks benefit differently from more data.
What this means in practice
- •For remote sensing engineers: Improve land-surface analysis from Sentinel-1 SAR data with a data-efficient pretrained ConvNeXt model that reduces compute and memory use.
- •For earth observation platform developers: Embed a compact SAR image representation model pretrained with SparK-SAR to enhance several Copernicus task performances using limited labeled data.
Authors
Yuki Natsume, Sang-Ho Yun, Chai Kiat Yeo
Abstract
Labeled Sentinel-1 Ground Range Detected (S1GRD) datasets remain scarce, limiting supervised deep learning for land-surface remote sensing, while self-supervised pretraining for this domain has been dominated by Vision Transformer (ViT)-based masked image modeling requiring substantial data and compute. We propose SparK-SAR, a sparse convolutional masked pretraining framework built on SparK and ConvNeXt, designed to exploit convolutional inductive biases for S1GRD imagery under constrained compute. SparK-SAR makes three core technical contributions: (i) ConvNeXt-RS, a domain-adapted encoder that halves the stem stride (4x to 2x) to produce 14x14 feature maps at stride 16x, providing 4x more spatial tokens at identical parameter count; (ii) a lightweight ConvNeXt-512 reconstruction decoder with PixelShuffle upsampling that reduces pretraining VRAM by 7.5-8.5% per GPU relative to a UNet-128 baseline while matching its accuracy; and (iii) an empirical characterization of FP16 mixed-precision pretraining instability in the standard SparK UNet-768 decoder on S1GRD data, complemented by a finding that finetuning with layer-wise learning rate decay (LLRD) improves downstream performance of pretrained encoders. On four Copernicus-Bench S1GRD tasks, our finetuned 27.8M-parameter ConvNeXt-RS encoder improves over a non-pretrained baseline by +4.7, +5.0, +5.8, and +1.4 pp on EuroSAT-S1, BigEarthNet-S1, DFC2020-S1, and Flood-S1 respectively, and outperforms the Copernicus-FM ViT (139M parameters, frozen) on three of four tasks at roughly 5x fewer parameters. A pretraining data-scaling study reveals task-dependent data efficiency: classification saturates with as little as 10-25% of the SSL4EO-S12 v1.1 corpus, while dense prediction continues to improve significantly up to 25-50%. The study also includes a CKA representation analysis and a geographic de-collision experiment.