Title:

Fire regimes at the arid fringe: A 16-year remote sensing perspective (2000–2016) on the controls of fire activity in Namibia from spatial predictive models

Publication Year:
2018
Abstract:

Dry-season fires affect the grassland and savanna ecosystems in Namibia and other regions around the globe. Whereas climate, especially precipitation, has been found to constrain fire activity in (semi-)arid regions through productivity, the feedbacks with human systems lack generalization. Here, we assess the biophysical and humanrelated controls of fire activity in Namibia based on a 16-year record (2000–2016) of the MODIS Burned Area product (MCD45A1). The two derived parameters of fire  activity include burned area (positive continuous) and the number of fire occurrences (zero-inflated counts), and are individually investigated at a 0.1°-resolution by means of five common statistical and machine-learning techniques. We evaluate performance and consistency among the models using the adjusted coefficient of determination and the root mean square error, which is obtained from 5-repeated 10-fold cross-validation. A comparable measure of predictor importance among the models is assessed by means of a permutation-based approach. As spatial autocorrelation is present for both parameters of fire activity, we consider this with a spatial cross-validation setup, where k-Means clusters of geographic coordinates are used to derive the test partitions. We find complex machine-learning techniques generally improve the predictions of both parameters of fire activity. Our results confirm the exceptional importance of mean annual precipitation for fire activity across Namibia and highlight human impacts as an additional control of fuel availability. Apart from an increase of burned area and fire occurrences at a mean annual precipitation of approximately 400 mm, population and livestock densities strongly limit fire activity in the bestperforming Random Forest models. The largest burned areas are found with moderate green-up rates of vegetation, which we attribute to the presence of open landscapes. The consideration of spatial autocorrelation generally decreases model performances but the relative decreases are higher for the models of burned area, which we attribute to the increased spatial autocorrelation present with this response variable. Resultantly, we recommend accounting for spatial autocorrelation in the evaluation of spatial ecological models and the assessment of predictor importance. Although Namibia's land use practices denote a special case, our model may be of relevance to other regions located at the arid fringe of fire-affected ecosystems and those with projected future aridification. Keywords: Fire ecology, Fuel limitation, Machine-learning, Dry savanna, Spatial autocorrelation, Model comparison.

Publication Title:

Ecological Indicators

Volume:
91
Pages:
324-337
Item Type:
Journal Article
Language:
en

EIS custom tag descriptions