Application of machine learning models for asbestos cement detection in multispectral images

Introduction: the detection of asbestos cement has predominantly been carried out using hyperspectral data. The challenge lies in improving detection efficiency without compromising accuracy, especially in contexts where only multispectral images are available. Objectives: the main objective of this study is to apply machine learning models for the detection of asbestos cement in multispectral images. Methodology: the research was conducted using a four-phase adaptation of the CRISP-DM methodology, which included the following steps: Dataset creation: A dataset was compiled and prepared from a reference multispectral image of the city of Cartagena. Evaluation of relevant bands: The most relevant spectral bands for detecting asbestos cement were selected. Model training and evaluation: Four machine learning models were trained and evaluated: kNN, decision trees, support vector machines (SVM), and logistic regression. Deployment of the best model: Finally, the best-performing model was deployed on the reference image to assess its performance in a real-world scenario.Results: of the four models evaluated, the decision tree model demonstrated the highest efficiency and consistency on both the training and testing sets, achieving a value of 0.93 for both Precision and Recall metrics. In contrast, the logistic regression model performed the worst, with values of 0.829 and 0.875 for Precision and Recall, respectively. Conclusions: the proposed decision tree model proved to be an effective and consistent tool for detecting asbestos cement in multispectral images, with a good balance between precision and recall. This model provides a viable solution for automatic material detection in contexts where only multispectral images are available and can serve as a reference in both academic and industrial settings for integration into material detection systems.