Differential Box-Counting Tool¶

This work is licensed under a Creative Commons Attribution 4.0 International License.

The Differential Box-Counting (DBC) method extends traditional box-counting to estimate the fractal dimension of grayscale images, making it suitable for analyzing continuous data like canopy height models, terrain surfaces, and biological images.

Method Overview¶

Traditional Box-Counting¶

For binary images, box-counting covers the image with boxes of size \( \varepsilon \) and counts boxes \( N(\varepsilon) \) containing the structure:

\[ D_B = \lim_{\varepsilon \to 0} \frac{\log N(\varepsilon)}{\log(1/\varepsilon)} \]

Differential Box-Counting¶

For grayscale images, DBC partitions both the spatial domain and the intensity axis. For each grid position, it counts the number of boxes needed to cover the intensity range:

\[ n_r(i,j) = \text{ceil}\left(\frac{I_{\max}(i,j) - I_{\min}(i,j)}{s}\right) + 1 \]

where: - \( I_{\max}(i,j) \) and \( I_{\min}(i,j) \) are the maximum and minimum intensities in the box - \( s \) is the box size in the intensity dimension

The total count is:

\[ N_r = \sum_{i,j} n_r(i,j) \]

Using the Tool¶

With FracLac (ImageJ)¶

Install FracLac: Download from FracLac website
Prepare Image:
Convert to 8-bit or 16-bit grayscale
Crop to region of interest
Run DBC Analysis:
Plugins → Fractal Analysis → FracLac → Differential Box Count
Set box size progression (typically power series with factor 0.1)
Run analysis
Interpret Results:
Examine the log-log plot of \( N_r \) vs \( 1/r \)
Linear region indicates fractal scaling
Slope gives fractal dimension

With Python¶

import numpy as np
from scipy import ndimage

def differential_box_count(image, box_sizes):
    """
    Calculate fractal dimension using differential box counting.

    Parameters:
    -----------
    image : 2D numpy array
        Grayscale image (values 0-255)
    box_sizes : list
        List of box sizes to use

    Returns:
    --------
    dimension : float
        Estimated fractal dimension
    """
    counts = []

    for size in box_sizes:
        # Grid dimensions
        h, w = image.shape
        n_h = h // size
        n_w = w // size

        total_count = 0
        for i in range(n_h):
            for j in range(n_w):
                # Extract box
                box = image[i*size:(i+1)*size, j*size:(j+1)*size]

                # Count boxes needed to cover intensity range
                i_max = np.max(box)
                i_min = np.min(box)
                n_r = np.ceil((i_max - i_min) / size) + 1
                total_count += n_r

        counts.append(total_count)

    # Linear regression on log-log data
    log_sizes = np.log(1 / np.array(box_sizes))
    log_counts = np.log(counts)

    slope, _ = np.polyfit(log_sizes, log_counts, 1)

    return slope

Applications¶

Canopy Height Models¶

Aerial LiDAR-derived canopy height models (CHM) exhibit fractal properties. Expected dimensions:

Forest Type	Expected \( D_M \)	Notes
Lowland Rainforest	1.3-1.4	Complex vertical structure
Mixed Conifer	1.4-1.5	Intermediate complexity
Pine Plantation	1.5-1.6	More uniform structure

Biological Tissues¶

Medical imaging applications:

Bone trabecular structure: \( D \approx 2.2-2.4 \)
Tumor vasculature: \( D \approx 1.6-1.9 \)
Neural branching: \( D \approx 1.3-1.5 \)

Terrain Analysis¶

Digital elevation models:

Smooth terrain: \( D \approx 2.1-2.2 \)
Moderate terrain: \( D \approx 2.2-2.4 \)
Rugged mountains: \( D \approx 2.4-2.6 \)

Best Practices¶

Image Preparation¶

Resolution: Use highest available resolution
Normalization: Scale intensities to 0-255 range
Noise reduction: Apply gentle smoothing if needed
Border effects: Crop edges or use padding

Box Size Selection¶

Minimum: At least 3-4 pixels
Maximum: No more than ¼ of image dimension
Progression: Power series (factor 1.5-2.0) or exponential

Quality Assessment¶

\( r^2 \) value: Should exceed 0.99 for reliable estimate
Coefficient of variation: Low CV indicates stable estimate
Visual inspection: Check log-log plot linearity

Interpreting Results¶

Mass Dimension vs Box Dimension¶

The differential box-counting method estimates the mass dimension \( d_M \), which may differ from the standard box-counting dimension \( D_B \):

For self-similar fractals: \( d_M = D_B \)
For self-affine fractals: \( d_M \) reflects the scaling of mass/intensity

MST Predictions¶

Metabolic Scaling Theory predicts \( d_M = 3/2 = 1.5 \) for vascular networks when viewed in 2D projection. Empirical values typically fall in the range \( 1.45-1.55 \).

References¶

Sarkar, N., & Chaudhuri, B. B. (1994). An efficient differential box-counting approach to compute fractal dimension of image. IEEE Transactions on Systems, Man, and Cybernetics, 24(1), 115-120.
West, G. B., Brown, J. H., & Enquist, B. J. (1999). The fourth dimension of life: fractal geometry and allometric scaling of organisms. Science, 284(5420), 1677-1679.