Fire images classification using high order statistical features

Abstract

Wildfires and forest fires have devastated millions of hectares of forest across the world over the years. Computer vision-based fire classification, which classifies fire pixels from non-fire pixels in image or video datasets, has gained popularity as a result of recent innovations. A conventional machine learning-based approach or a deep learning-based approach can be used to distinguish fire pixels from an image or video. Deep learning is currently the most prominent method for detecting forest fires. Although deep learning algorithms can handle large volumes of data, typically ignore the differences in complexity among training samples, limiting the performance of training models. Moreover, in real-world fire scenarios, deep learning techniques with little data and features underperform. The present study utilizes a machine learning-based approach for extracting features of higher-order statistical methods from pre-processed images from publicly available datasets: Corsican and FLAME, and a private dataset: Firefront Gestosa. It should be noted that handling multidimensional data to train a classifier in machine learning applications is complex. This issue is addressed through feature selection, which eliminates duplicate or irrelevant data that has an effect on the model's performance. A greedy feature selection criterion is adopted in this study to select the most significant features for classification while reducing computational costs. The Support Vector Machine (SVM) is a conventional machine classifier that works on discriminative features input obtained using the MIFS, feature selection technique. The SVM uses a Radial Basis Function (RBF) kernel to classify fire and non-fire pixels, and the model's performance is assessed using assessment metrics like overall accuracy, sensitivity, specificity, precision, recall, F-measure, and G-mean.