Remote sensing diagnosis of Forest fire risk based on state-trend characteristics using machine learning models

Abstract

Forest fires rank among the primary natural disasters threatening the ecological security of the Greater Khingan Range region, making the precise diagnosis of fire risk factors crucial for fire prevention and control. This study proposes a remote sensing diagnostic framework for forest fire risk that integrates dual “trend–state” characteristics, aiming to address the challenges of factor redundancy and insufficient dynamic response in fire risk assessment. Multi-source datasets from 2010 to 2024, including remote sensing, meteorological, topographic, and human activity data, were integrated to construct an initial set of candidate factors. Dominant factors were identified through the combined use of Geodetector and Recursive Feature Elimination (RFE), and further categorized into static “state indicators” and dynamic “trend indicators” to establish a multidimensional risk assessment system. On this basis, Random Forest (RF) and Long Short-Term Memory (LSTM) models were employed to conduct fire risk prediction, enabling a comparative analysis of machine learning and deep learning performance. Research findings indicate: (1) Forest fire risk in the Greater Khingan Range is primarily dominated by the “moisture-heat-wind” climate combination, with annual precipitation ( q =0.75) being the most explanatory dominant factor; (2) The RF model demonstrated optimal performance in fire risk assessment ( AUC =0.94), significantly outperforming the LSTM model; (3) Spatially, high-risk zones are concentrated in low-elevation areas of the southeast, west, and south. Temporally, the pattern exhibits overall stability alongside periodic fluctuations and extreme risk events. This research provides a high-precision dynamic diagnostic framework for fire prevention and control in cold-temperate forests, verifying the key role of remote sensing time-series features in risk warning.

Topics & Concepts