The threatened Gouldian finch preferentially forages in prescribed burnt savannas

Study area

The study was conducted during 2019 in the hills around Wyndham (15°28′S, 128°07′E; Fig. 1) in the east Kimberley region of Western Australia. Most of the average annual rainfall of ~910 mm falls between December and March (monsoon wet season; Bureau of Meteorology 2020). The annual mean day temperature for 2019 was 37.2°C, with daily maximums ranging from 25.0 (June) to 45.5°C (December). The main topographic features of this area are rocky sandstone ridges and gently sloping savanna hills with a tree canopy dominated by Corymbia dichromophloia and Eucalyptus miniata (Brazill-Boast et al. 2013), a ground layer in which the annual grass Sorghum stipoideum and the perennial grasses Triodia bitextura, Eriachne obtusa and Sorghum plumosum predominated. There is a pronounced fire season throughout the east Kimberley region. In the early dry season (before July), small, low-intensity hazard reduction burns are lit by land management agencies to create fire breaks. Few naturally or accidentally ignited fires occur during this season. In the late dry season (July–November), large high intensity wildfires are ignited annually, either by lightning, particularly during the build-up or storm season (September–November), or by anthropogenic ignition sources (roadworks, campfires, accidents, arson). An average of 60% of the savanna in the immediate Wyndham area was burnt annually from 2016 to 2019.

Fig. 1. 

Study region in the east Kimberley region of Western Australia showing the locations at which the finches were caught (star), the locations of automated telemetry receiver towers (black triangles), and the fire frequency in 2016–2019 (green to red shading).

Prescribed burning

Prescribed burning was undertaken both to increase the abundance and quality of grass seeds available for Gouldian finches (Weier et al. 2016, 2017, 2018, 2019), and also to reduce the negative impact of high intensity wildfires on tree nesting hollow density that finches rely on during breeding (Brazill-Boast et al. 2011; Radford et al. 2021). To achieve this, strips of inter-connected savanna vegetation were burnt early in the dry season. Finch habitat in the study area has a history of annual incursions by wildfires from the adjacent township, and one of the functions of networks of burnt habitat was to reduce the incursion of these high-intensity wildfires into finch habitat. Prescribed burning was applied by hand in a series of walked and driven lines using matches and drip torches. Fire lines were planned to be a sufficient distance apart that fire spread rates did not allow adjacent lines to meet prior to self-extinguishing (soon after ignition or overnight). To achieve small patches of prescribed burnt habitat (~<300 m wide or 10–100 ha), ignition was planned for the wet-dry season transition, immediately after annual Sorghum seed had dehisced, under humid, low wind conditions, and in grass fuels which were not fully cured (usually March–April in east Kimberley). The target for prescribed burning in Gouldian finch breeding habitat was ~20–30% of finch habitat annually. This burning regime was designed to provide sufficient newly burnt savanna to benefit finches while breeding and feeding during the dry season (Weier et al. 2016), and also provide habitat with infrequent enough fires (once every 3–4 years) to allow slow maturing perennial grasses such as Triodia spp. to produce seeds for wet season feeding (Dostine et al. 2001; Legge et al. 2015).

Bird sampling

In 2019, we conducted three finch sampling sessions in the early (April/May), mid (July) and late (October) dry season. Mist nets (12 m length, 2.7 m height, 25 mm mesh) were used to catch finches across 25 days at two waterholes approximately 11 km apart. Nets were positioned to capture birds leaving the water source following their morning drink and were continually monitored. After the birds were extracted from the nets, a uniquely numbered size-two aluminium band was placed on the left tarsus of any bird not captured previously. Each bird was weighed using a Pesola scale (±0.5 g), and the wing, head-bill, tail, and tarsus length were measured (±1 mm). The sex and age of the birds was determined from their plumage – adult male Gouldian finches are more brightly coloured than females, particularly the purple colouration of the breast. Juvenile Gouldian finches are a dull grey and cannot be sexed using plumage.

Monitoring movement

Coded VHF transmitters (Avian NanoTag model NTQB2-2, Lotek Wireless Inc.) were attached to Gouldian finches using a figure-8 leg loop harness made from an elastic thread (Rappole and Tipton 1991). The elastic thread degraded and fell off from 30 to 130 days after application, as confirmed by recapturing birds that had previously carried transmitters. Each transmitter and harness weighed 0.31 g, which was 2–3% of the average weight of the individual birds (13.8 g). Several sizes of harness (29, 30, 31, and 32 mm loop sizes) were used because size varied among individuals. Each transmitter was coded with a uniquely identifiable sequence of pulses emitted every 11–13 s at a radio frequency of 151.500 MHz. Tags were active for a 12 h on/off cycle to capture movement during the daylight hours (05:00 hours–17:00 hours) but save the battery at night.

Movements of tagged Gouldian finches were monitored using an array of six receiver stations. Each receiver station listened continually and recorded signal strength, receiver location, and the time and date of each detection of a coded transmission. Receivers were, on average, 1.7 ± 0.78 km apart and provided continual coverage over approximately 21.2 km² and encompassed all tagging locations. Each receiver contained a power source, a receiver, and an antenna tuned to the 151.500 MHz frequency. Each receiver was connected to a monopole antenna (5.1 dBi), which had a 360° detection radius and a range of up to 4 km (mean = 3 ± 0.963 km). Each antenna was vertically mounted to a 5 m length × 5 cm diameter galvanised steel post. The Sensorgnome device was mounted to the pole approximately 1 m above the ground. Receivers ran the open-source Sensorgnome software on a raspberry-pi 3 device (Compudata Systems, London, Ont., Canada) with Adafruit Ultimate GPS Hat maintaining time synchronisation for each device, and Nooelec NESDR SMArTee RTL-SDR radio dongles receiving the signals from the tags. Each receiver station was powered by solar panels and a 12 V deep-cycle battery. This arrangement provided 5 days of power redundancy and lasted throughout the dry season and well into the wet season, when cloud cover was more prevalent.

When multiple receivers detect the same bird through the unique sequence of beeps and pauses, the position of the bird can be estimated by the variation in signal strength of the detections. Using multilateration, the signal strength from several receiver stations (3+ simultaneous) was used to estimate the location of a tagged individual. Distance to position estimates were ground-truthed using a drone, as described in Collett (2023).

Fire mapping

Sentinel 2 imagery (bands 11 and 12; resolution 20 ×20 m) was used to map fire history and frequency from when it first became available from February 2016 to November 2019. Whilst other imagery was available from before February 2016, it lacked the fine-scale resolution of the Sentinel imagery required for the analysis. At least one image per month with minimal cloud cover (<5%) was chosen from February through to November each year. No fires were recorded in January and December. Images were downloaded as level 1C products from the Sentinel-hub EO browser (https://app.sentinel-hub.com/eo-browser), and the 11 and 12 bands were used as they reflect short-wave infrared reflection (SWIR).

Object-based image analysis (OBIA) in SAGA GIS (ver. 2.3.2; Conrad et al. 2015) was used to classify burnt areas. The two steps involved in OBIA are segmentation and classification. Firstly, the image was segmented into unclassified objects based on spectral properties, as defined and set by the user. Secondly, images were classified as either ‘burnt’ or ‘unburnt’ using a set of predetermined features and criteria. Manual on-screen editing also complemented the classification process and resulted in four burnt-area maps from 2016 to 2019. From these maps, we determined four variables attributed to fire history: (1) area burnt (m²); (2) year of burn; (3) month of burn; and (4) whether the fire was an early dry season (February–June) or late dry season burn (July–November).

The map of each variable was converted to a raster, and the raster calculator in QGIS was used to create fire frequency and time-since-last-fire maps. These maps showed: (1) overall frequency of fire through the area (FF); (2) frequency of early dry season burns (number of years burnt between 1 March and 30 June in the previous 4 years (EDSFF)); (3) frequency of late dry season wildfire (number of years burnt between 1 July and 1 December in the previous 4 years (LDSFF)); and (4) time-since-last-burn (months, TSLF); Table 1, see Supplementary Fig. S1).

Data preparation and analysis

Four levels of filtering were applied to the data. First, tracking data from the first 24 h following tag deployment were excluded from analysis to allow for a post-tagging acclimation. Second, detections with a run length of <3 were excluded to minimise the chances of false positives (Crewe et al. 2018), a run length being defined as the number of consecutive tag transmissions received by any one antenna. Third, detections were filtered to the six receivers that were active over the entire sampling period. Finally, birds were dropped from analysis if they were detected for fewer than 5 days per month in order to ensure that the sample of an individual’s behaviour was robust and representative.

Individual centres of activity (CoA) were estimated using non-linear least squares (NLS), whereby a position during each 7-min time interval was estimated by minimising the difference between estimated and ‘known’ distances from a tower for a minimum of three simultaneous detections. The ‘known’ distance from a tower was estimated using the relationship between signal strength and distance from each tower, collected using drones with nanotags attached (Fig. S2; Crewe, Udyawer, Collett, Radford, Micheli-Campbell, Campbell, unpubl. data). The 7-min interval was used, as this minimised the error in position estimates when comparing the known drone positioning (via onboard GPS) with the estimated position of the drone (with the maximum likelihood estimation method used here; Fig. S3).

We used the fixed kernel method to estimate utilisation densities (KUD) to assess core use areas, where positions were predicted via the non-linear least squares method, using the raw detections. For each individual, the 50% volume contour (in m²) of the KUD was created for each month of the tracking period using the ‘adehabitatHR’ package (Calenge 2006) in the software R. All KUD calculations used the default reference bandwidth as the smoothing parameter.

In order to determine if there was a relationship between core use areas and fire history, we clipped fire variable maps to the extent of the receiver station detection range and calculated the area for each level of the fire variable within the detection range. Each monthly core area was then overlaid with the clipped fire variables. Finally, the proportion of the available habitat that the birds used was calculated by determining the area of each fire level that intersected with the core use area using the ‘sp’ package in R software.

To determine the preferences in site use in response to the different fire variables, we used a permutational multivariate analysis of variance (PERMANOVA) in PRIMER ver. 7 (Anderson 2001; Clarke and Gorley 2015). Fire variables were normalised, and a Euclidean distance matrix calculated that was used as an outcome in the PERMANOVA models. Main and pairwise testing was conducted using 997–999 permutations to assess changes in habitat use according to each level of fire history. The available habitat was randomly sampled to test whether finches actively favoured a particular element of the fire history. To sample the survey area randomly, circular polygons were created that had an area equal to the average core use area of all birds. These polygons were located within the available habitat, where multilateration would be possible (maximum detection range for two or more towers). For each month, the number of these polygons sub-sampled was twice the number of core use areas estimated for that month. Up to half of the random samples were the same between 1 month and the next, with the rest randomly sub-sampled each month. The same random samples were used to assess the four fire variables in four different PERMANOVA models. A binary fixed factor was included in the tests to distinguish between site use (i.e. Kernel utilisation distributions (KUD) of birds and the random samples of available habitat; hereafter called ‘Bird vs Avail.’). Tag ID marking individuals was included as a random effect, nested within the ‘bird vs available’ factor, and both the ‘bird vs available’ variable and month were included as fixed factors.

Variables that showed a significant difference between the birds’ KUD and random samples of the available habitat or across months (P < 0.05) were further used for Similarity Percentages Procedure (SIMPER) analysis. SIMPER was used to identify the most influential fire interval or fire frequency for each fire history variable model, which accounted for most of the observed dissimilarity in fire patterns between months and/or between used habitat versus available habitat. A Non-metric Multi-Dimensional Scaling (nMDS) ordination was performed to visualise the multivariate data (i.e. the fire history patterns of each bird’s use area compared to a random sample of the available habitat (Fig. S4)). The nMDS for FF, EDFF and LDSFF had stress levels <0.2, allowing 2D interpretation of birds’ space use. However, the TSLF variable had an nMDS stress of 0.27, which did not allow for a confident interpretation of the results.

The most influential levels of each variable identified in the SIMPER analysis were analysed further to identify which variable had greatest influence on the birds’ space use. A generalised linear mixed model (GLMM) was used to assess the influence of fire history variables and the KUD months on individuals’ space use in response to selected fire variables. The level of each fire variable that contributed the most to the observed dissimilarity between the habitat in core use areas and available habitat (2-years EDFF, 1-year LDFF, 3-years FF and 16-months TSLF) was used as an explanatory variable in the model. As this model examined the proportion of the total area occupied within a bird’s core use area, a beta regression was used. The Tag ID was included as a random effect in the model. Models were run using the ‘glmmTMB’ package (Brooks et al. 2017) in R Software (). Model residuals were checked to ensure a lack of pattern across predictors and fitted factors, as well as to identify influential outliers.

Finch capture

Across three tagging sessions between 30 April and 10 October 2019 (tagging sessions – April, July, October), 68 Gouldian finches were fitted with nano VHF transmitters and their movements tracked. Of these tagged birds, 30 had sufficient data to conduct Kernel Density estimates (KUD; a minimum of five re-locations). These individuals were detected from between 7 and 130 days, with detections occurring at an average of 29.4 ± 2.8 tracking days per month.

Activity space

After the pre-processing, a total of 3 203 979 detections were collected from the 30 birds in this study (Table 2, Table S1). The mean core use area (50% KUD) in July was almost twice the size of the core use areas in other months (433 838 ± 68 842 m² s.e.). The smallest core area used by tracked birds, in October (176 462 ± 44 169 m² s.e.), was approximately half that used by birds tracked in July. There was little variation in the location of the core use areas throughout the dry season between individuals or months (Fig. 2, Figs S5, S6, Table S2).

Fig. 2. 

The number of 30 VHF-radio tagged Gouldian finches (Chloebia gouldiae) in the 50% KUD that overlap with one or more other tagged birds, in the east Kimberley region of Western Australia. KUD, kernel method to estimate utilisation densities.

Fire

Fire history variables had a strong influence on 50% of the birds’ core use area, as shown by a significant difference between the fire-related characteristics of an individual’s core use area and those of the surrounding area (Table 3). The fire patterns differed significantly between the birds’ core use area and available area for all fire pattern models (Permanova P < 0.01, for all).

There was a large difference in fire frequency patterns for early dry season fires between the habitat in core use areas and the available habitat (PERMANOVA: P-value = 0.001, the square root of the estimate of component of variation (ECV) = 0.46; Table 3). The SIMPER testing found that 48% of the observed dissimilarity between habitat in core use areas and the available habitat could be attributed to the 2-year (out of 4-year) prescribed burning frequency (Fig. 3). Between 75% and 99% of an individual’s core use area consisted of areas where there had been 2-year early dry season prescribed burning, compared to an average of 52% for the randomly chosen areas (Fig. 4a).

Fig. 3. 

Non-metric MDS ordination plot comparing the prescribed burning fire histories (fire frequency 0–4) chosen by birds (red ‘down’ triangles), with those histories that were available in the habitat (blue ‘up’ triangles). The vectors (FF 0–4) indicate the direction and strength of the Pearson correlation of the fire variables with the nMDS axes (with the centre point of the vectors placed at a random location on the nMDS). FF, fire frequency; MDS, multi-dimensional scaling.

Fig. 4. 

The proportion of an individual bird’s core use area, made up of (a) the varying frequencies by any fire, prescribed burns, wildfires for the bird’s core use areas (dark grey), and that which is available in the habitat (light grey); (b) the varying time-since-last-fire histories, between bird’s core use areas (dark grey), and that which is available elsewhere in the habitat (light grey).

The overall fire frequency also influenced the core use area of birds, albeit slightly less than the prescribed burning patterns (PERMANOVA: P-value = 0.001, ECV = 0.35 (Table 3); beta regression: −1.63; P < 0.001 (Table 4, Fig. 5)). On average, 65% of the birds’ core use areas had experienced fire three times over the past 4 years (Figs 4a, 5) – thereby contributing 74% to the observed fire pattern similarity for birds – compared to 28% for available random samples.

Fig. 5. 

Based on the beta regression model, predicted mean ± 95% CI proportion of the bird’s core use area occupied by the most influential level of the four fire variables (1-year wildfire frequency, 16-months-time-since-last-fire, 2-years prescribed burning frequency, and 3-years overall fire frequency).

The frequency of late dry season wildfires was the least influential fire frequency type, particularly compared to the prescribed burning patterns (PERMANOVA: P-value = 0.001, ECV = 0.21 (Table 3); beta regression: −1.69, P < 0.001 (Table 4, Fig. 5)). Gouldian finches avoided areas that were burnt more than twice in the past 4 years by late season wildfires with on average, 64% of their core use areas burnt once by late season wildfires (Fig. 4a) – thereby contributing 72% to the observed fire pattern similarity for core use areas, as compared to 53% of the randomly chosen areas.

Recently burnt areas (0–2 months) made up some of a bird’s core use area (contributing 11.5% [0 months], 9.2% [1 month] and 10.6% [2 months] to the observed dissimilarity of fire pattern between the birds’ space use and randomly chosen areas), while 16-months post-fire contributed slightly more (15% to the observed similarity in core use areas). However, time-since-last-fire was the least influential fire variable to explain a bird’s use of the core area (PERMANOVA: P-value = 0.001, ECV = 0.18 (Table 3); beta regression: −2.37; <0.001) (Fig. 4b).

The dissimilarity between the fire history of a bird’s core use area and the fire history of the landscape changed to during the year (Table 3) and between fire history variables. The average dissimilarity between the fire frequency patterns of a bird’s core use area and the surrounding habitat was greatest during November and least in August (Table 5). However, the greatest dissimilarity between the birds’ core use area and the surrounding landscape was in May for EDSFF and November for LDSFF (Table 5). This suggests that birds more strongly preferred high EDSFF habitat early in the dry season and high LDSFF habitat late in the dry season.

Management Implications

Current land and fire management varies across the Australian savanna. While fire management is largely determined by the proximity of human populations and local production industries, it should not overlook ecological impacts, particularly where savannas are managed for conservation. Ecological benefits of prescribed early dry season burnt savanna has been demonstrated here in terms of preferential finch foraging patterns but have been previously highlighted regarding finch nest box occupancy (Weier et al. 2016), grass seed availability (Weier et al. 2018, 2019) and retention of high-density nesting habitat (Radford et al. 2021). Areas burnt in patchy, low-intensity prescribed fires have greater temporal and spatial availability of annual and perennial grass seed throughout the year than extensively burnt habitats. This is because the patchy application of prescribed burning allows retention of both recently burnt and longer unburnt vegetation (Dostine et al. 2001; Dostine and Franklin 2002; Weier et al. 2018, 2019), which in turn enhances both annual and perennial grass seed availability and is thus associated with increased breeding success and body condition/stress levels for Gouldian finches throughout the year (Legge et al. 2015; Weier et al. 2016). We recommend that land managers use a mosaic of patchy, low-intensity burns in 20–30% of known breeding habitat annually to achieve both protection against extensive wildfires. This will improve grass seed resource availability and heterogeneity (both dry season annuals and wet season perennial grass seeds) across the landscape. Implementation of 20–30% annual prescribed burning will allow individual patches to be burnt no more frequently than once every 3years in line with greater Gouldian breeding success being achieved in recently but infrequently burnt savannas (Weier et al. 2016). We recommend limited strategic use of wet season burning to avoid burning large areas of annual Sorghum before it has set seed in March/April, thereby reducing Sorghum seed availability in the following dry season. A network of early dry season burns would reduce the overall within-site fire frequency whilst still maintaining recently burnt areas and patches of long unburnt vegetation to retain flowering and seeding perennial grasses, including the slow-maturing Triodia spp. (Dostine et al. 2001). Such a fire regime will help ensure the landscape is not dominated by high-frequency, extensive wildfires with all their negative consequences for biodiversity in northern Australia (Franklin et al. 1999; Andersen et al. 2005; Woinarski and Legge 2013; Lawes et al. 2015; Radford et al. 2020).