Characterizing fire history on military land using machine learning and landsat imagery

Background In the past several decades, United States wildland fire occurrences have increased due to anthropogenic activities, shifts in precipitation and temperature patterns, and long-term fire suppression policies. Detailed records of local fire histories are needed to further understand ignition sources and the interaction between human activity, weather patterns and fire occurrences. Aim To estimate local fire histories, we delineate burned area on military installations over decadal time series (1984–2023) of Landsat imagery using random forest and boosted regression tree algorithms. Methods We trained and tested each model with 10 images from a manually delineated burned area dataset and applied them to Landsat images acquired from 1984 to 2023. We validated the model’s yearly summaries with the remaining manually delineated burned area dataset and compared success rates through confusion matrices, omission/commission error, sensitivity and specificity. Key results The mean accuracy for the random forest models across all four installations was 0.941, while the mean accuracy of boosted regression models was 0.935. There was no significant difference between random forest and boosted regression model performance. Conclusions We present a methodology which can be utilized by other Army personnel and local land managers to develop fire histories for local-scale management units, particularly those geared towards national defense institutions.