Efficacy Analysis of Boosting Techniques for Road-Surface Detection Using Very High Resolution Trispectral Satellite Images

Abstract

Boosting techniques are machine-learning algorithms that combine the strengths of several weak learners to increase the accuracy of predictive models. They are invaluable for tasks involving objects, fraud, and analysis. Optimization algorithms iteratively construct models by concentrating on the errors made by earlier models, thereby increasing their overall accuracy and precision. Changing the weights of misclassified instances helps manage imbalanced data such as anomalous or fraudulent instances. Boosting techniques are resistant to noise, focusing on hard-to-classify cases and mitigating the impact of noise through weighting schemes to produce more dependable detection models. Boosting techniques can be used for many detection problems where the desired features are few compared to non-desired ones. They offer flexibility in model building, as they are not restricted to a particular type of base learner and can be used with decision trees, neural networks, and other algorithms. Updating the model and adding new data without retraining it through incremental learning is also possible. High-dimensional data can be handled effectively by boosting techniques, particularly for object detection in images and videos. In this study, the WHU dataset was used to detect road surfaces in an image. A smaller number of desired pixels than the others mimic an imbalanced binary classification problem. In this study, we tested adaptive boosting, gradient boosting, histogram gradient boosting, and light gradient boosting machines to analyze the efficacy of boosting techniques.