Identification of Smoke Peatland Fire using Remote Sensing Data based on Machine Learning

Smoke information serves as a crucial marker for detecting peatland fires. Practically, smoke identification utilizing remote sensing satellites, based on visual interpretation techniques, proves inefficient in processing time and high subjectivity. The application of machine learning technology for smoke detection remains limited in the tropics, especially in peatland areas. This study aims to identify smoke from peatland fires using machine learning techniques. The dataset comprises Visible Infrared Imaging Radiometer Suite (VIIRS) images and the VIIRS's hotspots on September 11st 2019, coinciding with a major peat fire incident in Indonesia. Various machine learning techniques were tested, encompassing Random Forest (RF), Classification and Regression Trees (CART), Support Vector Machine (SVM), Naive Bayes (NB), and Gradient Tree Boost (GTB). Object classification includes thin smoke, thick smoke, clouds/smoke, clouds, vegetation, water body, and bare land. The accuracy assessment involved both qualitative assessment based on true-color images and quantitative evaluation through the 70:30 sample splitting accuracy assessment method. The analysis of spectral distance for the seven object types reveals that band 5, 10, and 12 exhibit the highest value. Successfully identification of peatland fire smoke is achieved via supervised machine learning, particularly logic-based algorithms (RF, CART, GTB) and support vector machine methods (SVM), while statistical method (NB) yield comparatively less success. Qualitative validation using true-color VIIRS image indicates strong alignment between thick smoke and the RF all-bands approach. Quantitative validation, based on accuracy assessment with 1531 samples, establishes SVM as the most accurate method, boasting an overall accuracy of 0.93, followed by GTB at 0.91, RF at 0.90, and CART at 0.88.