Rolling tail-dependence cubes vs center pixel (Sentinel-2 NDVI z-scores)

Tail-dependence statistics reveal whether extremes in one pixel align with extremes in another. This recipe computes bottom- and top-tail dependence against the center pixel, plus their difference, using NDVI z-score cubes as input.

Steps:

1. Load Sentinel-2 data and compute NDVI z-scores.
2. Optionally coarsen/stride for tractable rolling windows.
3. Call rolling_tail_dep_vs_center to compute bottom, top, and difference cubes.
4. Visualize the difference cube with Lexcube and review QA medians.

import cubedynamics as cd
from cubedynamics import pipe, verbs as v
from cubedynamics.stats.tails import rolling_tail_dep_vs_center
from cubedynamics.utils.chunking import coarsen_and_stride
from cubedynamics.viz.qa_plots import plot_median_over_space

s2 = cd.load_sentinel2_cube(
    lat=43.89,
    lon=-102.18,
    start="2023-06-01",
    end="2024-09-30",
    edge_size=1028,
    resolution=10,
    cloud_lt=40,
)

ndvi_z = (
    pipe(s2)
    | v.ndvi_from_s2()
    | v.zscore(dim="time")
)
ndvi_z_ds = coarsen_and_stride(ndvi_z, coarsen_factor=4, time_stride=2)

# Rolling tail-dependence cubes vs center pixel
bottom_tail_cube, top_tail_cube, diff_tail_cube = rolling_tail_dep_vs_center(
    ndvi_z_ds,
    window_days=90,
    min_t=5,
    b=0.5,
)

# Visualize the difference cube with Lexcube
diff_clip = diff_tail_cube.clip(-1, 1)
pipe(diff_clip) | v.show_cube_lexcube(
    title="Tail-dependence difference (bottom - top) vs center pixel",
    cmap="RdBu_r",
    vmin=-1,
    vmax=1,
)

# QA: median tail-dependence difference over space
ax = plot_median_over_space(
    diff_tail_cube,
    ylabel="Median bottom - top tail dep",
    title="Median tail-dependence difference (rolling 90 days)",
    ylim=(-1, 1),
)

Positive difference values highlight areas that tend to share low NDVI events with the center pixel more often than high NDVI events, which can be an early indicator of coordinated stress.