공간 교차검증을 적용한 딥러닝 기반 산불 민감성 분석 - 경상북도 안동시와 의성군을 중심으로 -

초록

The large-scale wildfire that occurred in the Yeongnam region in March 2025 demonstrated that damage can be concentrated in specific areas where multiple risk factors such as weather, topography, vegetation, and human activities interact. However, wildfire occurrences exhibit strong spatial clustering, which can lead to overestimation of prediction performance and reduce map reliability due to information leakage caused by spatial autocorrelation when random partition-based validation is applied. To address this limitation, this study constructed a deep neural network-based wildfire susceptibility map at 60m resolution for Andong-si and Uiseonggun in Gyeongsangbuk-do, and evaluated its spatial generalization performance using spatial cross-validation. The results indicated that, even under spatially separated validation conditions, the model maintained its discriminative power in distinguishing wildfire affected and unaffected areas. High susceptibility areas exhibited a clustering pattern centered around slopes, mountainous regions, and transition zones to lowlands. These high susceptibility patterns generally corresponded with NBR (Normalized Burn Ratio)-based damage distributions, whereas threshold based indicators showed limited performance due to severe class imbalance and strict spatial separation conditions. This study demonstrates the potential of susceptibility maps as evidence-based tools for wildfire prevention and resource allocation prioritization, emphasizing the necessity of incorporating spatial autocorrelation into validation designs for deep learning-based wildfire prediction.

키워드