EVIE

Ensemble Visualization and Imaging Environment

Coming soon. EVIE quantifies shape and texture in ensembles of curves extracted from imaging data. The software is being prepared for public release; the worked example below is real output from it, run today.

Shape and texture in the solar corona

A worked example run end to end through EVIE — segmentation, separable shape tensors, principal geodesic analysis, and a two-sample maximum-mean-discrepancy test. Every number and every curve on this page came out of the pipeline. Nothing here is illustrative.

Imagery courtesy of NASA/SDO and the AIA, EVE, and HMI science teams — sdo.gsfc.nasa.gov. Three AIA 193 Å frames, 2048 × 2048, grouped by observation date.

The question

Each frame segments into hundreds of closed level-set boundaries — the outlines of coronal structure. One image is therefore already an ensemble of shapes. Group the frames by date and a question follows naturally: does the distribution of coronal loop shapes change from one day to the next?

A separable shape tensor splits each closed curve into two independent factors — a point on the Grassmannian Gr(2, n) carrying undulation (how the outline wiggles), and a symmetric-positive-definite matrix carrying anisotropic scaling (how big and how elongated it is). The test can then ask about each factor separately, and say which kind of difference it found.

Segmented boundaries

The traced level sets, in image pixel coordinates. Colour marks the observation date. These are the actual curves the pipeline consumed.

What the shape tensors say

Grassmannian PGA spectrum

Singular values σk of the undulation factor, log scale. The decay says how many modes carry the ensemble's shape variation.

Distance from the intrinsic mean

Each point is one boundary: dGr (undulation) against dSPD (scaling), coloured by date. The two clouds overlap.

Fréchet / Karcher mean

The intrinsic mean shape (solid) over the standardized ensemble (faint). Computed on the manifold, not by averaging coordinates.

Decision landscape

Test outcome across the analysis grid — reduced dimension r against landmark count nre. Circle = fail to reject, × = reject.

The answer, and the interesting part

A single run at one smoothing scale is a weak claim, so the pipeline sweeps the scale and re-runs the whole grid at each one. That is where this example stops being flat:

Read this before drawing a solar-physics conclusion. The two ensembles are imbalanced — boundaries from two frames on the first date against from a single frame on the second. The surviving boundary count also changes with smoothing scale, because smoothing decides which small structures clear the minimum-arc-length threshold. A decision that appears at one scale and not another is therefore evidence about scale sensitivity of the statistic, which is exactly what the sweep exists to expose — not, on this evidence, a measured change in the corona. Reporting it the other way round is the mistake the sweep is designed to prevent.

Reproducing this

The dataset ships with the code as data/sun_temporal/, and this is one command: 

evie-mmd-g2image --data-path data/sun_temporal \
    --r 5 10 20 --n-re 200 500 --scale 5 --min-arc 200 \
    --scales 3 5 7 --alpha 0.05 --shuff 200

EVIE also runs as an interactive terminal interface (python -m evie) and as a local dashboard, and mirrors the same pipeline in MATLAB/Octave.