Dynamic World, Revisited — Lightweight, Temporally Consistent Land Cover at Kayrros

In 2022, I joined Kayrros to explore whether we could reproduce—and, for some production needs, improve on—Google’s Dynamic World land‑cover pipeline. The goal was not to beat Dynamic World outright, but to build a lightweight, temporally consistent model that behaved well across geographies, sensors, and seasons.

Internship at Kayrros, supervised by Aurélien De Truchis (2022).

Why Dynamic World (and why revisit it)?

Dynamic World provides a near real‑time, 10‑m land‑cover product across nine classes (Water, Trees, Grass, Crops, Shrub & Scrub, Flooded Vegetation, Built‑up Area, Bare Ground, Snow & Ice). It is remarkably useful, but we observed two recurring challenges in some deployments:

Temporal consistency across dates and seasons (flicker, sensitivity to atmospherics)
Robustness across biomes when training/operating at scale with multiple sensors

Our focus was to build a minimal model that stays stable over time while remaining fast and simple to operate in production.

Method at a glance

Data curation + temporal augmentation: feed multiple dates that share a refined label so the model learns invariances to seasonal/atmospheric effects.
Satellite‑agnostic inputs: Sentinel‑2 and Landsat families plus derived indices (e.g., NDVI, NDMI) and SRTM30 elevation for context.
Lightweight U‑Net variant: three downsampling stages with targeted enhancements (attention/APSP as needed) for speed/accuracy balance.

Temporal augmentation across dates with varying atmospherics and corresponding training confidences — Temporal augmentation: multiple observations of the same area (top) paired with training confidences (bottom) help the model learn invariance to atmospherics and phenology.

Seeing the model work

Below are examples comparing raw Sentinel‑2 imagery and the model’s land‑cover output around Mount Kenya.

Zoomed Sentinel‑2 crop near Mount Kenya — Zoom view — left: Sentinel‑2 crop (raw). See also the paired classified view below.

Zoomed model classification near Mount Kenya — Zoom view — right: model classification (Crops, Trees/Shrubs, etc.).

Land‑cover legend for the classification outputs — Legend — nine Dynamic World classes as used in this project.

Results across biomes

We tested across 14 major ecoregions (temperate to tropical). Final metrics:

Overall accuracy: 46.9%
Mean IoU: 31.8%
MCC: 75.1%

While slightly below Google Dynamic World on benchmark scores, the model delivered two practical advantages important in operations:

1) Improved temporal consistency across dates and seasons; 2) Portability across sensors and regions with a compact runtime.

From land cover to change signals

The model is particularly effective for dynamic indicators: deforestation alerts, agricultural expansion, and other environmental monitoring signals.

Example of forest loss detection with dating of change — Example: temporal analysis reveals forest loss (red) and dates of change, derived from stable land‑cover predictions.

Takeaways

Temporal augmentation and a restrained architecture go a long way for stability.
Sensor‑agnostic preprocessing and select indices add useful context without heavy engineering.
A lightweight, consistent baseline can unlock reliable monitoring at scale—even if headline benchmark numbers trail large, general models.

Acknowledgment: developed during my 2022 internship at Kayrros under the supervision of Aurélien De Truchis.