ViTs map algal blooms from Landsat/Sentinel-2 at 30m resolution
First vision transformer model for algal blooms beats spectral indices under clouds.
Monitoring coastal algal blooms demands frequent, high-resolution, and globally consistent imagery. Landsat-8/9 and Sentinel-2 A/B/C provide 30m multispectral data with near-global coverage every 2–3 days, but their limited spectral bands and lack of harmonized reflectance products make traditional bio-optical methods challenging. To overcome this, Thainara Lima and Vitor Martins introduced the first vision transformer (ViT)-based approach for floating algal bloom mapping using medium-resolution satellite imagery.
They created a globally distributed bloom patch dataset from coastlines worldwide and compared four transformer architectures (including Swin Transformer) against a standard convolutional baseline. Under cloud and glint stress, the Swin Transformer eliminated false positives that plagued spectral-index methods. Results showed omission and commission errors of 8–65%, and comparisons with MODIS products confirmed that higher spatial resolution (30m vs. 1km) detects fragmented blooms missed by coarse sensors. This work demonstrates deep learning as a reliable tool for consistent, operational algal bloom monitoring.
- First successful vision transformer application for algal bloom mapping using Landsat-8/9 and Sentinel-2 imagery at 30m resolution.
- Swin Transformer avoided false positives from clouds and sun glint that degrade traditional spectral-index approaches.
- Global bloom patch dataset trained models achieving 8–65% error rates, enabling detection of fragmented blooms unreachable by coarse sensors.
Why It Matters
Offers a scalable, accurate method to monitor harmful algal blooms from free satellite data, aiding coastal management.