LandTrendr Temporal Segmentation

Temporal segmentation of annual spectral time series using the LandTrendr algorithm (Kennedy et al. 2010). Identifies breakpoints in per-pixel trajectories to detect disturbance events (abrupt drops), recovery trends (gradual rises), and periods of stability. The algorithm fits piecewise linear segments to yearly composite values (e.g. NBR, NDVI), balancing goodness-of-fit against model complexity. Outputs include fitted values, vertex flags, RMSE, and per-segment metadata (magnitude, duration, rate of change). Two backends are available: a pure Python/NumPy implementation that runs on any array source (COGs, Zarr, GEE exports), and a GEE backend that wraps ee.Algorithms.TemporalSegmentation.LandTrendr.

ID	analysis_landtrendr
Domain	geo
Category	analysis
Type	raster-to-raster
Version	1.0.0

view yaml ◦view github ◦

Parameters

max_segments: integer · default 6 · Maximum number of segments to fit. The trajectory will have at most max_segments + 1 vertices (including endpoints). Higher values capture more detail but risk overfitting.
spike_threshold: number · default 0.9 · Threshold for dampening ephemeral spikes (0 to 1). 1.0 = no filtering. Lower values apply more aggressive spike removal. Spikes are single-year deviations that reverse in the following year.
vertex_count_overshoot: integer · default 3 · Number of extra candidate vertices beyond max_segments + 1 identified in the initial pass. These are later culled by angle-based removal.
recovery_threshold: number · default 0.25 · Controls maximum allowable recovery rate. A segment cannot fully recover in fewer than 1/recovery_threshold years. E.g. 0.25 means recovery cannot complete in fewer than 4 years. 1.0 disables the constraint.
p_value_threshold: number · default 0.01 · P-value threshold for model acceptance during iterative simplification.
best_model_proportion: number · default 1.25 · During model selection, a more complex model is preferred if its score is within this proportion of the best (simplest acceptable) score. Lower values favor simpler models.
min_observations_needed: integer · default 6 · Minimum number of valid (non-NaN) observations required to run segmentation. Pixels with fewer valid years are returned unfitted.
prevent_one_year_recovery: boolean · default true · Disallow recovery segments with a duration of exactly one year. Prevents overfitting in environments where such rapid recovery is not ecologically realistic.