Docs
Landcover Point Sampling (human in the loop FL)

Landcover Point Sampling (human in the loop FL)

Documentation for the Landcover Point Sampling system in LandOS.

Overview

The Landcover Point Sampling system is a human in the loop approach to landcover classification that provides a structured interface for recording, retrieving, and managing land cover observations tied to specific geographic locations.

Sample points collected through this system serve as training and validation data for the Forest Loop Random Forest classifier. Point quality and spatial distribution directly affect land cover classification accuracy. Remote sensing and landscape expertise is required to tage training points.


Data Sources

Sample points may originate from multiple sources, each identified by a source label:

SourceDescription
userLabels entered manually by a platform user
dwPoints derived from Dynamic World land cover classifications
esaPoints derived from ESA WorldCover classifications
mapbiomasPoints derived from MapBiomas annual land cover maps
hansenPoints derived from Hansen Global Forest Change data
aefPoints derived from AlphaEarth Foundations embeddings
jrcPoints derived from JRC global land cover products

Land Cover Classes

Points are labeled using the ten-class Forest Loop land cover scheme:

Forest, Water, Built, Degraded, Savannah, Clouds, Degraded Forest, Bare, Grassland, Shrubland.


Best Practices

Spatial distribution: Training accuracy improves when labeled points are spatially distributed across all land cover classes present in the project area. Avoid clustering points within a single land cover type.

Temporal alignment: Sample points should be labeled based on satellite imagery from the same time period as the Forest Loop classification being trained. Labeling against imagery from a different year introduces temporal inconsistency.

Global points: Points flagged as global are included in Forest Loop classifications across all polygons that overlap the H3 cell. Use the global flag only for points with high label confidence, as errors in globally shared points will propagate across projects.