Deep Learning and Remote Sensing for Forest-Fire Monitoring and Management in China: A Systematic Review and Evidence Synthesis

Abstract

Remote sensing and deep learning are increasingly used across the forest-fire management lifecycle in China, but the evidence differs substantially in task, sensor, model, and validation design. We screened 146 candidate records published during 2015–2025 using predefined criteria requiring a China-specific empirical component, forest/wildland-fire relevance, substantive remote-sensing input, substantive deep learning, and primary empirical evaluation. Thirty-four studies met all criteria, and 82.4% were published during 2023–2025. Active-fire detection and monitoring represented 38.2% of the evidence, followed by pre-fire risk and preparedness (32.4%), post-fire assessment (23.5%), and spread/behaviour modelling (5.9%). Sensor use was strongly task-dependent: UAVs supported fine-scale detection and fuel assessment, geostationary satellites supported frequent active-fire monitoring, and multisource satellite/geospatial inputs supported risk and post-fire analysis. Operational evidence remained limited: 11.8% of studies reported multi-region or multi-country validation, 8.8% runtime or latency, 11.8% false-alarm information, 5.9% interpretability, and none reported explicit uncertainty in the study-level evidence extracted for this review. The evidence does not support one universally superior architecture. Progress will depend more on task-matched sensing, independent geographic and event validation, calibrated uncertainty, false-alarm control, and clearer links between model outputs and management decisions. These priorities are relevant beyond China wherever remote-sensing fire systems must transfer across heterogeneous landscapes, sensors, and operating conditions.