Publications

Fusco, EJ; Finn, JT; Abatzoglou, JT; Balch, JK; Dadashi, S; Bradley, BA (2019). Detection rates and biases of fire observations from MODIS and agency reports in the conterminous United States. REMOTE SENSING OF ENVIRONMENT, 220, 30-40.

Abstract
With growing concern about the impacts of fires on ecosystems and economies, satellite products are increasingly being used to understand fire regimes. Concurrently, where available, agency records of fires have also been used to assess fire regimes. Yet, it remains unclear if these independent datasets measure the same fires, which raises concerns about the interpretation and benchmarking of models derived from these products. Here, we present a novel product intercomparison of the MODIS burned area and active fire products across the conterminous United States using nearly 250,000 agency reported wildfires as reference data to model consistencies and inconsistencies between all three datasets. We compared agency reported wildfires from the Fire Program Analysis fire occurrence database to the MODIS products to identify which fires were detected vs. omitted by MODIS products relative to agency fire records, and by agency fire records relative to MODIS. We created generalized linear models as a function of fire attributes (e.g. size) and environmental variables (e.g. cloud cover) to predict MODIS detection of agency wildfires, and anthropogenic variables (e.g. agriculture) to predict agency detection of MODIS fires. We modeled fire detection probability separately for MODIS burned area and active fire products, and for the eastern and western U.S. Overall, we found that MODIS product detection rates ranged from 3.5% to 23.4% of all documented agency wildfires > 1 ha, and that likelihood of detection increased with fire size. Agency detection rates ranged from 23.5% to 48% of MODIS burned area and active fires. Under ideal conditions, the MODIS active fire product had a 50% probability of detecting a wildfire that grew to at least 10 ha (eastern U.S.) 78 ha (western U.S.), while the burned area product had a 50% probability of detecting a wildfire that grew to at least 169 ha (eastern U.S.) 234 ha (western U.S.). Cloud cover and leaf area index were significant predictors of MODIS fire detection, while state boundaries were significant predictors of agency fire detection. This analysis presents an important assessment of the fire attributes and ground conditions that influence MODIS fire detection relative to extensive and increasingly used ground-based wildfire records. The large discrepancy in records of fire occurrence between MODIS and agency fire datasets highlights the need for this type of analysis into the types of fires likely to be included in each database.

DOI:
10.1016/j.rse.2018.10.028

ISSN:
0034-4257