Mars cloud clustering

name: mars-cloud-clustering description: Clustering Mars cloud annotations using DBSCAN with a custom distance metric and evaluating against expert labels. Use this skill when processing citizen science data that needs to be matched to ground truth points.

Mars Cloud Clustering Logic

This skill covers the implementation of DBSCAN with a custom distance metric and the evaluation of clusters against expert annotations.

Custom Distance Metric

DBSCAN should use a custom distance metric defined as: d(a, b) = sqrt((w * Δx)² + ((2 - w) * Δy)²) where w is the shape_weight.

Implementation in Python:

import numpy as np
from sklearn.cluster import DBSCAN

def custom_metric(a, b, w):
    dx = a[0] - b[0]
    dy = a[1] - b[1]
    return np.sqrt((w * dx)**2 + ((2 - w) * dy)**2)

# For use in DBSCAN (precompute or use a callable)
# metric=lambda u, v: custom_metric(u, v, w)

Evaluation Workflow

For each image:

1. Clustering: Run DBSCAN on citizen science points.
2. Centroids: Compute the mean (x, y) for each cluster found.
3. Greedy Matching:
   • Match centroids to expert points using standard Euclidean distance.
   • Use a greedy approach: match the closest pair first, then the next closest, etc.
   • Max distance threshold: 100 pixels.
4. Metrics:
   • True Positives (TP): Number of matches.
   • False Positives (FP): Number of clusters - TP.
   • False Negatives (FN): Number of expert points - TP.
   • F1 Score: 2 * TP / (2 * TP + FP + FN)
   • Delta: Average Euclidean distance of matched pairs.

Averaging Across Images

• Include all expert images in the F1 average.
• If no points/clusters/matches, F1 = 0.0 and delta = NaN.
• avg_F1 = mean(all_F1s)
• avg_delta = mean(all_deltas excluding NaN)