[Paper Review] A Review on Machine Learning Algorithms for Dust Aerosol Detection using Satellite Data

Nurul Rafi, Pablo Rivas|arXiv (Cornell University)|Apr 15, 2024

Atmospheric aerosols and cloudsEnvironmental Science3 citations

TL;DR

This paper reviews machine learning (ML) algorithms for detecting dust aerosols using satellite data, emphasizing multispectral remote sensing from MODIS, CALIPSO, and VIIRS. It demonstrates that while traditional linear band combinations are effective, advanced ML models like LP-SVR and PNN significantly improve detection accuracy (AUC up to 0.7349) and reduce false positives, enabling better dust storm monitoring and climate modeling.

ABSTRACT

Dust storms are associated with certain respiratory illnesses across different areas in the world. Researchers have devoted time and resources to study the elements surrounding dust storm phenomena. This paper reviews the efforts of those who have investigated dust aerosols using sensors onboard of satellites using machine learning-based approaches. We have reviewed the most common issues revolving dust aerosol modeling using different datasets and different sensors from a historical perspective. Our findings suggest that multi-spectral approaches based on linear and non-linear combinations of spectral bands are some of the most successful for visualization and quantitative analysis; however, when researchers have leveraged machine learning, performance has been improved and new opportunities to solve unique problems arise.

Motivation & Objective

• To evaluate the effectiveness of machine learning models in detecting dust aerosols using multispectral satellite imagery.
• To identify the most accurate and efficient ML algorithms for dust detection across diverse datasets and sensors.
• To compare physical-based methods with ML-based approaches in terms of accuracy, processing time, and generalization across different dust events.
• To explore the potential of deep learning and hybrid models for future dust aerosol monitoring and prediction.
• To assess the impact of feature engineering and spectral band combinations on model performance in dust detection tasks.

Proposed method

• Utilized multispectral satellite data from MODIS (Terra/Aqua), CALIPSO, and VIIRS, focusing on spectral bands sensitive to dust properties.
• Applied machine learning models including PNN, FFNN, SVM, RF, MARS, and SVR (including LP-SVR) for classification and regression of dust aerosol presence.
• Employed feature extraction techniques such as band ratios (e.g., NDDI), brightness temperature differences, and spectral indices to enhance model input.
• Used statistical metrics like AUC, TSS (True Skill Statistic), RMSE, and precision–recall to evaluate model performance across multiple datasets.
• Leveraged high-performance computing (HPCF) to reduce processing time from 30 hours to 30 minutes for large-scale data processing.
• Integrated bias correction techniques using SVM and neural networks to align MODIS AOD with ground-truth AERONET measurements.

Experimental results

Research questions

• RQ1 How do machine learning models compare to traditional physical methods in detecting dust aerosols from satellite data?
• RQ2 Which ML algorithm achieves the highest accuracy and AUC in classifying dust events across different satellite sensors?
• RQ3 What is the impact of feature selection and spectral band combinations on model performance in dust detection?
• RQ4 How do processing time and scalability vary across different ML models when applied to large-scale satellite datasets?
• RQ5 What are the most promising future directions for deep learning in dust aerosol detection, such as attention mechanisms or semi-supervised learning?

Key findings

• LP-SVR achieved the highest AUC (0.7349) and accuracy (0.9104), outperforming other models including PNN and FFNN in dust detection tasks.
• Random Forest (RF) demonstrated the highest TSS (True Skill Statistic) value, indicating superior performance in minimizing both commission and omission errors.
• PNN and LP-SVR showed comparable output performance, with LP-SVR achieving a 0.8295 precision and 0.9104 accuracy on a 75 million feature vector dataset.
• The use of HPCF reduced processing time from 30 hours to 30 minutes, significantly improving scalability for large-scale dust monitoring.
• SVM outperformed neural networks in bias correction between MODIS AOD and AERONET AOD due to its ability to operate in high-dimensional spaces via kernel mapping.
• GRA (Group Method of Data Handling) was found to be the most accurate model for predicting the Dust Storm Index (DSI) in Iran’s arid regions, with high AUC and low RMSE.