Skip to content

Baltimore Dataset Analysis Example

This example demonstrates a complete PMF analysis workflow using the Baltimore dataset, one of the sample datasets included with Easy PMF.

Overview

The Baltimore dataset contains atmospheric particulate matter measurements with concentrations and uncertainties for various chemical species. This example will walk you through:

  1. Loading and preparing the Baltimore dataset
  2. Running PMF analysis with different factor numbers
  3. Interpreting the results
  4. Visualizing factor profiles and contributions

Dataset Information

The Baltimore dataset includes: - File: data/Dataset-Baltimore_con.txt (concentrations) - Uncertainty File: data/Dataset-Baltimore_unc.txt (uncertainties) - Species: Multiple chemical species including metals, ions, and organic compounds - Samples: Time series of atmospheric measurements

Step-by-Step Analysis

1. Loading the Data

from easy_pmf import PMF
import pandas as pd

# Initialize PMF with Baltimore dataset
pmf = PMF()

# Load concentration and uncertainty data
pmf.load_data(
    concentration_file="data/Dataset-Baltimore_con.txt",
    uncertainty_file="data/Dataset-Baltimore_unc.txt"
)

# Display basic information about the dataset
print(f"Dataset shape: {pmf.concentration_data.shape}")
print(f"Species: {list(pmf.concentration_data.columns)}")
print(f"Sample period: {len(pmf.concentration_data)} samples")

2. Data Preparation

# Check for missing values and data quality
pmf.check_data_quality()

# Prepare data for PMF analysis
pmf.prepare_data(
    remove_missing_threshold=0.5,  # Remove species with >50% missing data
    replace_missing_method="median",  # Replace missing values with median
    apply_uncertainty_scaling=True
)

3. Running PMF Analysis

# Test different numbers of factors
factor_range = range(3, 8)
results = {}

for n_factors in factor_range:
    print(f"Running PMF with {n_factors} factors...")

    result = pmf.run_pmf(
        n_factors=n_factors,
        n_runs=20,  # Multiple runs for stability
        random_seed=42
    )

    results[n_factors] = result

    # Print model fit statistics
    print(f"Q/Qexp: {result.q_qexp:.2f}")
    print(f"Q robust: {result.q_robust:.2f}")

4. Selecting Optimal Number of Factors

# Analyze Q/Qexp values to select optimal factor number
q_qexp_values = {n: results[n].q_qexp for n in factor_range}

print("Q/Qexp values by factor number:")
for n_factors, q_qexp in q_qexp_values.items():
    print(f"{n_factors} factors: {q_qexp:.2f}")

# Select the optimal number (typically where Q/Qexp stabilizes around 1)
optimal_factors = 5  # Example selection
optimal_result = results[optimal_factors]

5. Interpreting Results

# Get factor profiles (source signatures)
factor_profiles = optimal_result.factor_profiles
print("\\nFactor Profiles (top species by contribution):")

for factor_idx in range(optimal_factors):
    print(f"\\nFactor {factor_idx + 1}:")
    # Sort species by their contribution to this factor
    factor_contributions = factor_profiles.iloc[:, factor_idx].sort_values(ascending=False)
    top_species = factor_contributions.head(5)

    for species, contribution in top_species.items():
        print(f"  {species}: {contribution:.2f}")

# Get factor contributions (time series)
factor_contributions = optimal_result.factor_contributions
print(f"\\nFactor contributions shape: {factor_contributions.shape}")

6. Visualization

import matplotlib.pyplot as plt

# Create comprehensive visualizations
pmf.create_all_plots(
    result=optimal_result,
    dataset_name="Baltimore",
    output_dir="output"
)

# The following plots will be generated:
# - Factor profiles heatmap
# - Factor contributions heatmap
# - Individual factor profile plots
# - Individual factor contribution plots
# - Top species by factor
# - Factor correlation analysis

Expected Results

For the Baltimore dataset, you typically expect to identify factors representing:

  1. Traffic/Mobile Sources: High in elemental carbon, organic carbon
  2. Secondary Sulfate: Dominated by sulfate, ammonium
  3. Crustal/Dust: High in aluminum, silicon, calcium
  4. Industrial Sources: Metals like zinc, lead, copper
  5. Biomass Burning: Potassium, organic carbon

Interpreting Factor Profiles

  • High values indicate species strongly associated with that source
  • Low values suggest minimal contribution from that source
  • Chemical ratios help identify source types (e.g., OC/EC ratios for traffic)

Interpreting Factor Contributions

  • Time series patterns reveal when sources are most active
  • Seasonal variations may indicate temperature-dependent sources
  • Episodic peaks could represent specific events or meteorological conditions

Quality Assessment

Key indicators of a good PMF solution:

  • Q/Qexp ≈ 1: Model fits the data appropriately
  • Stable factors: Results consistent across multiple runs
  • Interpretable profiles: Factors match known source signatures
  • Reasonable contributions: Time series make physical sense

Next Steps

After completing this analysis, you might want to:

  1. Sensitivity analysis: Test different data preparation methods
  2. Bootstrap analysis: Assess uncertainty in factor profiles
  3. Seasonal analysis: Separate data by seasons
  4. Comparison with other sites: Compare with Baton Rouge or St. Louis datasets

Files Generated

This analysis creates several output files in the output/ directory:

  • Baltimore_factor_profiles.csv: Numerical factor profile data
  • Baltimore_factor_contributions.csv: Time series contribution data
  • Baltimore_*.png: Various visualization plots
  • Baltimore_summary.png: Overview plot with key statistics

Troubleshooting

Common issues and solutions:

  • High Q/Qexp values: Try more factors or check data quality
  • Unstable factors: Increase number of runs or check for outliers
  • Uninterpretable factors: Consider different factor numbers or data preparation
  • Missing data warnings: Review uncertainty treatment and missing value handling

For more advanced techniques, see Custom Workflows.