Hyperspectral Classification of Blood-Like Substances Using Machine Learning Methods Combined with Genetic Algorithms...

Area of Science:

• Computer Science
• Remote Sensing
• Forensic Science

Background:

• Hyperspectral image classification presents challenges due to high dimensionality and spectral variability.
• Accurate classification is crucial in fields like forensic science for identifying substances.
• Optimizing classifier models and selecting relevant spectral bands are key to improving performance.

Purpose of the Study:

• To apply genetic algorithms (GAs) for optimizing model selection and band selection in hyperspectral image classification.
• To evaluate the performance of GA-optimized classifiers against traditional grid search methods.
• To assess the impact of data similarity between training, validation, and test sets on classifier performance.

Main Methods:

• Utilized a forensic dataset comprising seven hyperspectral images with blood and similar substances.
• Implemented genetic algorithms (GAs) for simultaneous model and band selection.
• Compared GA-based optimization with a classic grid search (GS) approach.
• Tested classifiers in two scenarios: same-image training/testing and different-image training/testing.

Main Results:

• GA-based optimization successfully reduced the number of spectral bands required for classification.
• GA-optimized classifiers demonstrated superior performance compared to GS-based models.
• Classifier accuracy was significantly influenced by the similarity of data used for model optimization and testing.
• The importance of a representative validation set for GA optimization was highlighted.

Conclusions:

• Genetic algorithms offer an effective approach for optimizing hyperspectral image classification models and band selection.
• GA-based methods provide a more efficient and accurate alternative to traditional grid search.
• Careful consideration of data similarity during the validation phase is critical for successful GA application in hyperspectral imaging.