Integrated grassland fire danger assessment for Golden Gate Highlands National Park using geospatial techniques

Abstract

Grasslands are key to the earth’s system and provide crucial ecosystem services. However, the global degradation of grassland ecosystems is increasing at an alarming rate, and fire is regarded as one of the major drivers of degradation. Fire has been an integral evolution factor of grassland ecosystems, key in shaping and maintaining grassland’s ecological functioning. With global changes such as anthropogenic climate changes that alter fire regimes, fire is perceived as a destructive factor associated with negative impacts on the environment, society and economy. Moreover, there is numerous evident showing that over the past several decades and predicted to continue, average temperature area rising, predicted to continue, rainfall patterns are changing, resulting in more frequent and longer lasting drought across all ecosystems; protected and grassland areas are becoming progressively more challenging with threat to biodiversity and tourism operations and activities due to bush-encroachment, fauna and flora and species losing their natural inhabitant and competition, destroy tourism infrastructure and alter the attractiveness of tourist destination and loss of revenue and landscape ecosystem services.

Integrated fire risk assessment systems provide an integral approach to fire prevention and control and mitigate the negative impacts of fire. Fire danger assessment, as one of the components of integrated fire risk assessment, is regarded as an extremely challenging endeavour owing to the confluence of factors that influence fire ignition and fire behaviour. Most global fire danger systems and indices are based on meteorological data, do not consider fire causes and have limited spatio-temporal dimensions of fire drivers or switches; therefore, they are deemed outdated and inadequate in accuracy and validating estimation of fire danger.

Numerous methods have been employed to quantify the complex interrelation between fire switches and fire occurrence to estimate fire danger or susceptibility with different precision outcomes, thence, the prediction accuracy is still debated. However, most of these studies focused on northern hemisphere forest fires; therefore, grassfire research is limited, particularly in the African and Latin American grassland ecosystem. This study, therefore, developed an integrated grassland fire danger assessment model in a protected Afromontane grassland ecosystem, the Golden Gate Highland National Park (GGHNP) as the study area, taking into account all four fire drivers adapting the fire danger assessment component of the framework for fire risk assessment. Geospatial technology such as remote sensing (RS), geographical information system (GIS), Google Earth Engine (GEE) and Climate Engine (CE) were used to evaluate the spatial and temporal variation of the fire drivers.

Lightning is a well-known natural fire ignition source and has been used as the proxy for such, and is inadequately represented in fire ignition models due to lack of quantitative spatial pattern on flash density. The study developed the Lightning Hazard Map (LHM) from 11-year lightning strikes data (2007 – 2017) using GIS spatial analytic techniques (hotspot analysis, Global Moran I and spatial autocorrelation analysis). The results revealed December, 15:00 and 17:00 as the month and hours with the highest lightning activity, respectively. Clustering was observed throughout the year, except in July and the early to late morning hours at 06:00, 08:00 and 09:00 South African Standard Time (SAST). The clustering of lightning strikes at the park is at a distance of about 1.2 km. This connotes that strikes clustered with other strikes are not likely to strike an individual specific location from the centre of the cluster of strikes much beyond a circle with a radius of 1.2 km. While clusters starting at 2 km to 8 km. Forty-nine (49%) and 16% of the study area was found to be highly to extremely prone to lightning-caused fire.

Humans cause 90% of fires globally through land clearing and agricultural activities, however, most of the fire danger systems do not consider the temporal dimension of human-caused fire ignition. The study temporal modelled and evaluated the human behaviour as an agent of human-caused fires using distance from roads and tourist infrastructure (INFRA) as indicators using 11-year fire occurrence data. The results showed that INFRA and roads influenced all models at different times. The INFRA had influence during weekend days (Saturday and Friday) of both winter and spring seasons, while ROADS had influence during weekday (Monday to Friday) of both seasons.

The inaccuracy of most fire danger systems and indices often lies in the coarse spatial resolution data, particularly at the landscape scale. Using the GEE platform, this study developed a data fusion method termed “Blend-then-Index” (BI) by blending indices from Sentinel-2 and MODIS sensors to a 10 m spatial resolution to estimate the degree of grass curing (DoC). Grass curing is the transformation of live fuel to dead fuel component of the fuel bed and is one of the crucial input parameters for the grassland fire danger system or index, also inadequately represented in fire danger assessment systems due to sparse measurement systems which are too expensive and time-consuming, data scarcity and inaccessible. The results revealed September, February and May as the months with the highest, lowest and onset of degree of curing, respectively. Based on grass curing, over 80% of the study area was prone to fire danger. The results showed that over 90% of fire points fell under grass curing maps’ (GCMs) high and extreme fire danger. Moreover, the study revealed an inverse correlation between grass curing indexes (GCIs) and precipitation and soil moisture. Owing to this precision, the “BI” data fusion method was used in the study to measure other fuel factors.

Finally, the study developed prediction fire danger map models taking into consideration the relationship between the confluence of fire drivers and fire danger and 20 variables were selected. Namely: LHM, proximity from road, infrastructure, rivers, GCI, global vegetation moisture index (GVMI), vegetation condition index (VCI), bare soil index (BSI), soil bulk density, coarse fragments, soil moisture content, total plant available water holding capacity (TAWCP) elevation, aspect, slope, topographic positioning index (TPI), topographic ruggedness index (TRI), topographic water index (TWI), land surface temperature (LST) and wind speed. All the models showed that almost 85% of the study area is prone to fire, ranging from low to extremely dangerous. In comparing the best-performing models in predicting fire danger, the results revealed trade-off between statistical and machine learning methods. Decision trees (DT) showed high precision on model fit and success rate (ROC/AUC value of 0.93, 0.96 and 0.50); however, it was outperformed by weight of evidence (WoE) (0.83, 0.82 and 0.7) in predictive rates. Other models, frequency ratio (FR) (0.92, 0.95 and 0.66) for model-fit, success rate and predict rate validation, logistic regression (LR) (0.63, 0.64 and 0.59), random forest (RF) (0.91, 0.94 and 0.53), support vector machine (SVM) (0.63, 0.64 and 0.59). The WoE method showed robust performance in all three accuracy assessments (model fit, success and prediction rates). and therefore, based on that, the study suggests a hybrid approach of statistical and machine learning methods for improving the accuracy of fire danger prediction. Furthermore, based on the WoE model, almost 53% of the park is highly prone to fire, observably concentrated in the southern, south-western and eastern parts.

The study revealed that the geospatial techniques used are effective and can effectively model fire danger. The results also revealed that the study area’s fire regime was fuel-driven and caused by lightning and humans. This suggests that a fire management strategy should be based on a fuel management strategy that considers the local fuel conditions. Therefore, there is an urgent need to identify and determine the thresholds of fuel conditions (fuel load, fuel moisture content, and other fuel characteristics in relation to fire ignition and fire behaviour). This study assists in meeting sustainable development goal including SDG 15, i.e., life on land including protection of natural landscape and biodiversity, and conservation of mountainous ecosystems.