Geophysics reveals forest vulnerability to drought

Drought and heat‐induced forest die‐off are being increasingly reported across the planet. As vulnerable areas tend to have thin soils and poor water holding capacities, quantification of soil depth thresholds, relative to drought intensity, has global implications for identifying forest areas at risk. Measuring soil depth at forest stand or regional scales is, however, difficult. Our aim was to quantify soil thickness across drought impacted forest stands using geophysics. We asked whether impacted sites had shallow soils and whether soil thickness was associated with drought effects and forest structure. Electrical resistivity measurements were conducted at three sites in the Northern Jarrah Forest, Western Australia, which experienced die‐off during a 2010 drought and subsequent 2011 heatwave. Multispectral imaging quantified stand structure and vegetation cover. Geophysics identified shallow bedrock in the centre of all drought sites. Soil thicknesses correlated well with stand structure and cover, consistent with increasing water limitation in thinner soils. Smaller cover by trees and shrubs, more ground cover and shorter canopies were observed in soils <20 m thick and were more likely in soils <12 m thick, while treeless areas had thin soils, <2 m thick. The apparent resilience of this forest to long‐term drying looks to be due to the region's deep soils. With predicted future drying it is expected die‐off patches will expand outwards and new die‐off patches will emerge over thin soils. Geophysics can identify areas of forest vulnerable to future drought events, suggesting the potential for landscape‐scale mapping of drought vulnerability via airborne methods.

relation to location, timing and species in forested landscapes will be crucial to adaptive management and conservation (Restaino et al., 2019).
First principles and broadly established evidence establish how soil depth, properties and related belowground processes regulate vegetation structure and, by extension, forest sensitivity to drought (Phillips et al., 2016).Soil depth is a significant driver of tree and shrub species distributions via plant traits and resource availability (Bernard-Verdier et al., 2012;Goebes et al., 2019).Soil depth differences also shape structural attributes within a forest such as tree height, canopy and vegetation cover and stem density (e.g., Carrière et al., 2021;Horst-Heinen et al., 2021).Drought effects are expected to be more severe on shallow soil sites due to resource constraints; however, some degree of adaptation and niche segregation is expected to have already occurred in the shallowest soils, which most frequently experience drought stress (Zhang et al., 2021).Sites most susceptible to drought are more likely to experience significant and sudden structural adjustments during climatic extremes.For example, substantial changes in community structure and composition in holm oak (Quercus Ilex) forests occurred after drought, which were especially severe on sites with shallow soils (Galiano et al., 2012;Lloret et al., 2004;Peñuelas et al., 2000;Rivas-Ubach et al., 2016).Similarly, plantations of Eucalyptus globulus at shallow soil sites were found to be most susceptible to drought events (Harper et al., 2009).
Forest die-off over shallow soils is particularly evident in Mediterranean climate regions, especially during summer, when evapotranspiration demand continuously exceeds precipitation (Doblas-Miranda et al., 2017).The supply side of water is controlled largely by climate as well as physical properties that affect water storage and redistribution such as slope, soil structure, soil thickness and depth to groundwater (Zhou et al., 2013;Zolfaghar et al., 2014).
Global change modelling of carbon and water fluxes has been shown to be particularly sensitive to assumed soil depths, and correctly parameterizing soil depth is critical for determining both the magnitude and direction of predicted changes (Peterman et al., 2014).
Therefore, while soil thickness data are considered fundamental, it is often limiting and rarely sufficient to quantitatively examine forest loss over time or space (Phillips et al., 2016).
Soil thickness data are difficult to obtain and often limited to point measurements and shallow depths.Practical and financial constraints of directly probing the subsurface using drilling machinery often limits assessment to a small number of data or shallow soil sites where manual methods can be employed.Despite these difficulties, even shallow data have proven useful; for example, Harper et al. (2009) found drought mortality in a E. globulus plantation to be significantly higher on shallow soil (<2 m) sites, but they also noted limitations in evaluating effects at depths greater than 2 m and predicting soil depth at scale.While statistical approaches to predict soil thickness from landscape attributes are common (e.g., Horst-Heinen et al., 2021;Penížek & Boru ˚vka, 2011), these relationships may have limited ability for accurate predictions at regional scales (Pelletier et al., 2016).Scaling up to regional predictions and broader inference requires more direct and robust data.Emerging geophysics methods are, however, bridging these gaps and, in the right circumstances, can do so across a range of spatial scales enabling application to predicting forest drought susceptibility.
Geophysical methods are well suited to mapping subsurface heterogeneity and explaining variation in forest characteristics and drought responses.For example, Challis et al. (2016) applied electrical resistivity tomography (ERT, an imaging technique which involves injecting an electrical current into the ground and measuring resulting voltage) to image groundwater in deep sandy soils, confirming depth to groundwater as a major driver of spatial differences in the drought mortality of a shallow rooted tree species.Carrière et al. (2021) applied ERT and electromagnetic induction (EMI, which measures the apparent electrical conductivity of soil), and characterized the vegetation using hemispherical photographs, to demonstrate that higher forest biomass occurred where soils were deeper.Resistivity methods have also been applied to measure temporal variation in soil hydrology related to hydraulic redistribution and water availability (e.g., Robinson et al., 2012).While not a complete panacea for direct measurements, when applied and interpreted with knowledge of their limitations (Leopold et al., 2013), ERT and EMI, in conjunction with ecological data, can provide useful information on the spatial structure of the subsurface of relevance to forest die-off.
Southwestern Australia is an important place to study globalchange type drought, as the region has experienced chronic drying since the 1970s, acute droughts in 2000, 2006-2007 and 2010, fol-lowed by a series of heatwaves in early 2011 (Bates et al., 2008;Ruthrof et al., 2018), which led to a canopy die-off event (Matusick et al., 2013).This Northern Jarrah Forest (NJF) region also contains ancient, deeply weathered soil profiles and thus provides an extreme end member to evaluate the effects of soil depth and climate change on forest health.A previous landscape-scale analysis suggested dieoff sites were associated with rocky soils and low water holding capacity, located at higher elevations, near bedrock outcrops and on steeper slopes, suggestive of the presence of thin soils (Brouwers et al., 2013).These sites therefore offer the potential to quantify soil thickness using geophysics and thus establish whether the method can help inform drought and heatwave vulnerability of forests.Given this, here we asked the following questions: Q1.Are historical locations of drought-induced canopy die-off, and current canopy heights and vegetation cover, associated with shallow soils, as imaged by geophysics?Q2.Is there a critical threshold of soil thickness associated with observed drought effects?
Our expectations were that die-off sites would be characterized by shallow soils, and geophysics would clearly indicate that shallow soils would intersect areas of forest die-off.We further expected a gradient of declining soil thickness from unaffected trees to drought impacted trees, to treeless areas.

| Study region
The NJF is a dry sclerophyll forest located in southwestern Australia and covers an area of 1,127,600 ha (Havel, 1975 (Dell & Heddle, 1989).
The climate of the region is a Mediterranean type, with cool, wet winters and hot, dry summers, with 80% of rain falling between April and October, and drought periods lasting up to 7 months (Gentilli, 1989).Annual rainfall ranges from 1,100 mm year À1 in the southwest to 700 mm year À1 in the northeast (Gentilli, 1989).As discussed earlier, southwestern Australian has experienced a chronic decrease in rainfall of 15-20% and increase in mean annual temperatures by 0.15 C per decade, since the 1970s (Bates et al., 2008), and acute climatic events including droughts and heatwaves.
The NJF occurs on the northwest section of the Darling Plateau with the elevations of hill tops at 280-320 m and valley floors 50-100 m lower (Churchward & Dimmock, 1989).Due to more than one billion years of relative tectonic stability and periods of intense weathering, the plateau is deeply mantled by a lateritic regolith derived from granite and gneiss bedrock.The soil is several tens of metres deep in many places and comprises surficial lateritic gravels, sand and sandy clays over deep mottled and pallid clays and clay rich saprolite (Churchward & Dimmock, 1989).Soil depth, however, is modulated by colluvial redistribution, and alluvial incision in addition to heterogeneous weathering and the presence of numerous dolerite dykes.The clays provide a large soil water storage capacity and contain numerous preferential flow pathways, providing a means to infiltrate and store winter rains, with surface runoff losses typically less than 5% of annual rainfall (Ruprecht & Schofield, 1990).As E. marginata does not strongly control water use via stomatal regulation, these trees are reliant upon deep stores of soil moisture and groundwater, accessible via deep roots reported to as much as 50 m below the surface (Dell et al., 1983).

| Study sites
Study sites were established to explore the ecological effects of the severe drought heatwaves that occurred between 2010 and 2011, which led to $16,000 ha of NJF experiencing canopy die-off (Matusick et al., 2013).Drought affected and treeless areas were identified and mapped using a combination of aerial photography and ground surveys.The mapped classification of forest health by Matusick et al. (2013) was retained for this analysis.These study sites have led to an understanding of immediate effects (Matusick et al., 2013), differential responses of canopy species (Ruthrof et al., 2015), longer term stand recovery (Matusick et al., 2016), site associations (Brouwers et al., 2013), spatial clustering (Andrew et al., 2016), fuels and modelled fire potentials (Ruthrof et al., 2016), carbon dynamics (Walden et al., 2019) and legacy effects of historical drought stress (Matusick et al., 2018).
In a study by Steel et al. (2019), soil depth to an impeding layer was measured by driving a 1 cm diameter steel rod into the ground until refusal or to the 180 cm height of the rod.However, refusal may also have occurred on the lateritic duricrust, which is not uniform, is highly porous and is often penetrated by roots or former root channels and therefore is a poor indicator of the depth of tree root accessible soil.While both the Steel et al. (2019) and the Brouwers et al. (2013) studies provided indications soil depth may be a factor, highly accurate, quantitative relationships between die-off severity and soil depth have not been established.
From 20 longer term study sites, three were chosen for this study (Table 1).Criteria for choosing the sites included (1) covering a broad geographical area, (2) sites that were small enough such that geophysics transects could feasibly traverse patches of die-off and surrounding healthy areas and (3) were close to vehicle access tracks, given the weight of the equipment (Figure 1).The sites encompass a rainfall gradient from 823 to 1,042 mm a À1 but otherwise have similar elevations and slopes.Two sites (1 and 166) experienced logging in the 1980s, and the remaining site ( 27) was partly thinned in 2004, while other sections were last harvested in the 1920s.

| Geophysics
ERT is a minimum-invasive geophysical imaging technique used to investigate the subsurface properties of the Earth.Different materials and fluids conduct electricity to varying degrees, and this can be quantified by electrical resistivity, the inverse of electrical conductivity.
ERT involves injecting electrical current into the ground through a pair of electrodes and measuring the resulting voltage at other electrodes on the surface.The tomography in ERT aims to derive from the measurements the spatial distribution of subsurface electrical resistivity and involves two steps performed iteratively.The first is to model the measurement by simulating how the electrical field changes shape in response the spatial distribution of subsurface electrical resistivity.
In practice, a uniform resistivity is often used as a starting model of the subsurface electrical properties.Then, at the second step, a numerical process of regularization modifies the modelled distribution of electrical resistivity so that modelled and measured resistances match more closely (e.g., Loke & Barker, 1996).
To quantify soil thickness and resistivity characteristics, ERT transects were established at the three study sites, spanning a range of measured drought impacts.A Lipmann 4 Point Light resistivity metre (Lipmann Geophysics, Germany) was used to measure resistivity.Electrodes were spaced at 4 m intervals along transects of 236 m (60 electrodes at site 166) or 316 m (80 electrodes at Sites 27 and 1).
Currents applied ranged from 1 to 100 mA with the resulting mean voltage measured a minimum of 5 to a maximum of 10 times with an acceptance criterion of <10% difference between measurements.A Wenner array with a maximum separation of 26 electrodes was applied as the electrode configuration.The software res2dinv version 4.05.39 was used to edit noisy measurements and conduct inversions.
The software applies a smoothness-constrained least-squares method to invert the resistivity data (Loke et al., 2013).We specified a maximum of seven iterations or an error of less 5% as termination criteria for inversion.The horizontal and vertical resolution of the results depends in part upon the measurement method, the electrode spacing and the chosen discretization of the finite element grid of the numerical inversion.Due to the high resistivity contrast near the surface, the finite element grid adopted used a horizontal resolution of 1-2 m and a vertical resolution of 0.5-1 m near the surface increasing to 4-6 m near the maximum depths of investigation.The sensitivity of surface ERT measurements decreases with depth and is affected by the spatial distribution of subsurface resistivity.To assess this sensitivity, a depths of investigation analysis was conducted using the method of Oldenburg and Li (1999).The method is described in more detail in the supporting information (Text S1).
A resistivity threshold of 2,000 Ω m was identified to separate clay soils from bedrock consistent with previous work in similar geology (O'Brien et al., 2019).Visual inspection of the inverted resistivity sections identified a thin resistive layer near the soil surface, which was interpreted to be dry soil.Below that a sharp gradient of the resistivity value at values around 2,000 Ω m tended to occur at depth and this shallowed close to where bedrock outcropped at Site 1 or where granite fragments were seen in surface soils at 166 and 27.No other validation of soil depth was possible at the time of the study.A smooth spline interpolation of this 2,000 Ω m isoline was conducted to derive bedrock elevation and thus soil depth.

| Canopy heights and vegetation cover
Drought sites were previously identified from an aerial survey following the drought-heatwave event in 2011; each of the sites was visited and delineated from aerial photography and an on-ground survey using a differential GPS (Pathfinder Pro XRS receiver, Trimble Navigation Ltd., Sunnyvale, CA) by an ecologist (Matusick et al., 2013).To T A B L E 1 Environmental characteristics and management history of study sites, Northern Jarrah Forest, southwestern Australia.For training the image classifier, six cover classes were selected: (1) bare ground and litter; (2) photosynthetically active groundcover; (  (Tucker, 1979) was calculated using the red and near-infrared bands.
The pixel values from the imagery, NDVI image and NSM for each of the points were extracted in R (R Core Team, 2021).These data were then used to create a random forest model (Wright & Ziegler, 2017) for each individual site to predict cover class images at 8 cm by 1 cm resolution.To avoid imbalance in the training of random forest model, each class had 20 to 30 training observations and 20 validation points randomly selected during the cross-validation processes.Accuracy and Cohen's Kappa statistic were used to measure overall classification accuracy (Cohen, 1960).To relate above ground forest structure to the depth to bedrock, the cover and height images were sampled at 8 m intervals along the ERT transects from non-overlapping areas of 113 m 2 .

| Statistical analyses
Differences in means of cover class fractions and canopy heights were assessed via Welch t tests with non-equal variances as assessed by Bartlett's tests.Effect sizes were quantified by Hedges' g statistic, which is the ratio of the difference in the means standardized by the pooled and weighted standard deviations.Generalized additive models (GAMs) were conducted to evaluate changes to maximum and median canopy heights to assess nonlinear relationships with soil thickness.Linear regressions were also applied to quantify the effect size where linear relationships provided a satisfactory approximation of observed data.Logistic regression was applied to assess the relationships between soil thickness and the classification of the areas as drought-affected or treeless, assuming the data stemmed from a binomial distribution after a logit transform: where D denotes either drought-affected/healthy or treeless/not treeless classifications, S is soil thickness, j denotes the site and i the individual measurements at each site; β 0j and β 1j are fixed effects that may vary by site; the error, e, is assumed normally distributed (N) with a standard deviation, σ.In this instance, the response variables are assumed to follow a binomial distribution.Odds ratios and contingency tables were used to quantify effect sizes.Statistical significance is reported at the 95% confidence level.
The pre-processed dataset is included in Text S2 and Tables S1   and S2, and the R code to reproduce the statistical results is included as Text S3.

| Geophysics
Depth cross sections from ERT transects at the three sites show the ability of the method to clearly image subsurface features (Figure 2).
The inverted resistivity models converged rapidly and had errors of less than 5% (Table S3).The depth of investigation analyses showed the modelled resistivities were primarily the result of measured resistivities and not artefacts of edge effects or parameters used in the numerical model inversion at the depths shown in Figure 2 (see Figure S1).Resistivities greater than 2,000 Ω m were identified as defining the occurrence of granitic bedrock, coincident with outcropping at site Extra 1.A thin surface layer, 0-1 m, with large resistivities >3,000 Ω m, was interpreted to be warm, dry surface soil.This was typically underlain by material with resistivities <500 Ω m, and together, they were interpreted to be lateritic soil and pallid zone clay.
A vertically oriented feature with resistivities in the range of 700-1,500 Ω m at site 166 suggested the presence of a dolerite dyke, which are common in the region (Figure 2c).Resistivity cross sections showed similar patterns of depth to bedrock across all sites (shallow in the middle, deep at ends; cf. Figure 3).Soil generally thinned at the middle of each transect where drought effects were most pronounced and where treeless areas tended to occur (cf.Figures 3 and S2), and soil thickened as transects spanned into healthy forest.A small outcrop of granite was observed near the centre of the ERT transect at Site 1 just to the west of the larger treeless area (Figure 3b).

| Forest structure
The classification of the RPA imagery into six discrete classes achieved a mean accuracy of 91.9% (see Table S4).In addition to accuracy, Cohen's kappa statistic and confusion matrices show excellent classification of cover classes by the RF models of each site (Table S4; Figure S3).The classified images show forest structural differences within and between the three mapped sites (Figure 3).Forest health classes generally were significantly different in terms of cover fractions and canopy heights (Tables 2 and S5).Across the three sites, mean groundcover increased significantly from 39% in healthy forest to 56% in drought-affected areas then to 84% in treeless areas, while the fraction of area occupied by trees or shrubs decreased from 36% to 25% to 4%, respectively (Table 2).
The fraction of area classified as logs was greater in healthy and treeless areas than in drought affected forest.Numerous logs appeared to have been felled by past logging.Dead branches comprised a smaller area in treeless areas than drought affected or healthy forest.The maximum canopy heights decreased by 5 m (2.7-7.2 m 95% confidence) from healthy to drought-affected areas and a further 9.2 m (6.7-11.8m 95% confidence) from drought-affected to treeless areas.Mean canopy heights followed similar patterns of shortening from healthy through to treeless areas.There were significant differences in heights between healthy, drought-affected and treeless areas when grouped by site and also between sites when grouped by these health classes.Tree cover and canopy heights declined with an increase in site aridity (cf.Table 1).

| Soil thickness, forest structure and health
The canopy heights increase with soil thickness and the depth to the underlying bedrock at Sites 166 and 1 (Figures 2-4 and S2).The generalized additive models found significant smooth terms describing maximum and median canopy heights varying with soil thickness at these two sites (Table S6).Generally, the responses looked linear except at Site 1 the maximum canopy heights tended to plateau at 20 m for soil thicknesses greater than 20 m (Figure 4a).Linear regressions found significant increasing canopy heights with increasing soil thickness for maximum and mean canopy heights at Sites 1 and 166, while canopy heights at Site 27 showed no significant relationship to soil thickness (Table S7).
Using the soil thicknesses derived from geophysics, logistic regressions were applied to predict drought-affected and treeless areas (Figure 5; Tables S8 and S9).Soil thickness was found to be a significant predictor of drought-affected and treeless areas at Sites Ground cover % 38.6 ± 3.4 bc 55.9 ± 2.6 ac 83.5 ± 2.6 ab Note: Letters denote significant class differences.See Table S5 for effect sizes.
thick.Soils less than 2 m thick were more likely than not to be treeless.

| DISCUSSION
While chronic drying may, in part, be a component of a much longer natural variability (O'Donnell et al., 2021), only two other droughts of similar duration and intensity have occurred in southwest Australia over the last 2,000 years (Zheng et al., 2021).With the future projections for southwestern Australia describing further climate drying and warming, we expect die-off events will accelerate and expand to include more locations of forest in vulnerable, shallow soil areas (Hope et al., 2015).This could lead to sudden changes in forest structure and ecosystem dynamics (Allen et al., 2015;Doblas-Miranda et al., 2015, 2017).Our results point to a means to predict vulnerable areas using geophysics to map soil thickness as a key environmental variable and driver of forest health.
Returning to the first of the two research questions, Q1: Are historical locations of drought-induced canopy die-off, current canopy heights and vegetation cover, associated with shallow soils, as imaged by geophysics?, ERT geophysics showed very similar patterns of decreasing soil thicknesses towards the middle of all three drought sites (Figure 2).Soil thickness from ERT was also shown to delineate drought-affected and treeless areas from healthy forest at Sites 1 and 166 (Figure 5).Separately, drone imagery confirmed significant differences in cover and canopy heights between healthy, drought-affected and treeless areas.Therefore, geophysics has the potential to explain stand-scale variation in forest structure related to drought stress.
While the subsurface structures at all sites were similar, the forest structure at site 27 looks to have potentially been modified by additional factors, possibly including a higher susceptibility to fire, drought or their interaction, and a silvicultural thinning prior to the drought, which did not occur at the other two sites (Table 1).The effect of both these factors, primarily through reductions in leaf area and subsequent regrowth over several years, may be responsible for reducing the magnitude of differences between drought affected and healthy  F I G U R E 5 Logistic regressions to predict from soil thickness the drought-affected (a) and treeless (b) forest health class at the study sites, Northern Jarrah Forest, southwestern Australia.Solid line is the mean response and the ribbon its 95% confidence interval.
evident at the other two sites.Site 27 also has the largest minimum soil thickness of 2.8 m (0.6 m at Site 166 and 1.9 m at Site 1), which may have further mitigated drought effects immediately under the ERT transect.Finally, the Jarrah Forest in the northeast, where Site 27 is located, is also generally more open, reflecting a long-term moisture limitation and in combination with other factors may be complicating interpretation.
The ability of geophysics methods to clearly distinguish subsurface structure that impacts forest susceptibility to drought may vary from site to site.A distinct resistivity contrast between conductive clay and resistive granite bedrock facilitates soil thickness estimation by ERT in the NJF.Soil thickness estimation may be more challenging where there is not a strong resistivity contrast between soil and bedrock, which is more likely where bedrock is shallow, less massive and more fractured.Soil moisture and temperatures strongly influence resistivity, and this is clearly seen in the highly resistive, near-surface soils seen in this study.However, at depth soil moisture and temperature, variations dampen, meaning they become a less important influence on resistivity.In Challis et al. (2016), a groundwater table was clearly visible as a low resistivity feature in a survey with a background of higher resistivity, a feature of a sand/karst landform.In the NJF, however, ERT may not be able to distinguish a water table in the low resistivity clays, particularly as groundwater, like parts of the unsaturated zone, can contain a degree of salinity (Johnston et al., 1987).At shallow soil sites, high in the landscape, which is characteristic of the drought-affected sites in the NJF (Brouwers et al., 2013), where groundwater is unlikely to occur, ERT is therefore expected to be a good indicator of soil thickness and drought stress limits to productivity.Forest canopies tend not to influence ground based surveys like ERT but more careful data curation to correct for ground retrieval in airborne methods may need to be considered.
Additionally, electric fences or power lines can interfere with ERT and airborne methods.Careful consideration of the regional geology may therefore be needed in the selection of a geophysical method to map drought susceptibility.
Our second objective, (Q2) Is there a critical threshold of soil thickness associated with observed drought effects?, looks to have been answered in the affirmative.The analyses suggest critical soil thicknesses of 12 m distinguishes drought-affected from healthy forest and critical soil thicknesses of less than 2 m distinguished treeless areas (Table S8).Furthermore, there were taller and denser tree canopies in healthy areas than drought-affected areas (Table 2), particularly those downslopes of the areas with shallow bedrock (Figures 3,4 and S1).The critical soil depth may therefore be smaller where there is the potential for enhanced groundwater recharge in treeless areas supporting subsurface flows over the top of the bedrock.
Thicker soils were associated with taller canopies, greater cover by trees and shrubs, and lower bare ground and groundcover classes.
Soil thickness is known to control water availability at the plot level in forests but also the nature of the bedrock (Phillips et al., 2016).The bedrock of the Darling Plateau is generally massive and weakly fractured, impeding the growth of roots into deep layers (Abbott et al., 1989).Comparing drought effects between Spanish sites differing in bedrock structure, Lloret et al. (2004) found more severe drought effects in holm oak (Quercus ilex) forests where bedrock was more massive and less fractured, suggesting a role for both deeper root penetration and water storage in fractured bedrock.
In the NJF, previous drought studies have provided mainly qualitative measures suggesting shallow soil depth was important.For example, trees on sites closer to ridge tops were found more likely to be affected than trees on sites further away as were trees on sites close to rock outcrops, whereas distance to drainages and valleys were found to have no effect (Brouwers et al., 2013).Soil depth and groundwater depth may be correlated factors in some landscapes; however, little data exist with which to validate this in the NJF.While deep soils are common in the region, numerous bedrock outcrops can be observed, and thus, irregularly distributed soil thicknesses may be expected.On the other hand, landscape position, i.e., hill top, midslope or valley floor, may be a better predictor of typical depths to groundwater (Ruprecht & Schofield, 1990).Drought affected sites in the NJF, therefore, tend to be, higher in the landscape, on shallow soils (Brouwers et al., 2013), where groundwater is unlikely to occur.
The dominant canopy species in this ecosystem, E. marginata and C. calophylla, have extensive root systems that find their way through the laterite duricrust and into the pallid clay layer.The former is known to maintain relatively high transpiration rates during summer (Silberstein et al., 2001;Szota et al., 2011), owing to its deep root system accessing vadose and groundwater (Dell et al., 1983).However, high transpiration rates may make this species more vulnerable to acute droughts on thin soils that have limited connection to groundwater (Crombie, 1992), e.g., close to bedrock outcrops.Pallid zone clays, 8 m thick, have sufficient storage capacity to retain annual rainfall in the study areas (Ruprecht & Schofield, 1990); however, seasonal depletion of soil moisture to a third of its peak and to depths more than 25 m has been reported in the Jarrah forest (Silberstein et al., 2001).Total evapotranspiration from similar systems has been measured to be higher than annual rainfall in some years, though comparable to a 10-year mean (Macfarlane et al., 2010;Mitchell et al., 2009).These observations suggest thick soils allow for the inter-annual carryover of water storage and supplementation during dry years, providing some drought resilience.
Soil thickness is a significant driving factor in drought response, but other factors are also important, such as historical disturbances (wildfires, pests), past management (prescribed burns, logging, thinning), stand developmental stage and combinations of these.For example, the interaction between soil depth and basal area of a site is important because the level of competition may be more intense on shallow soils (Galiano et al., 2012) or the combination of high severity fire and drought may lead to abrupt shifts in stand structure and height (Walden et al., 2023).Historical harvesting and thinning may influence the level of competition on these shallow soils.The southern section of site 27 did not experience die-off like the other sections for example.Further work is needed to examine whether the effects of soil thickness is moderated by management of forest structure.With improved data on the interaction of stand density, soil depth and drought, then nuanced plans for forest density reduction treatment can be created.
Another key point is that the climate of the last 100 years may have been the wettest in the past 800 years (O'Donnell et al., 2021), thus potentially facilitating tree encroachment onto shallow soil areas.Now, during one of the driest periods in 2,000 years (Zheng et al., 2021), observed die-off likely reflects these trees reaching a critical threshold in shallow soils.Another way to think of this is that the critical soil thickness is increasing and expanding outwards from areas with shallow bedrock, due to the drying and warming.Regardless of the longer term pattern of tree encroachment and retreat, understanding the conditions experienced at the time of retreat, and locations where retreat occurs, will help us predict where and when the forest will reach thresholds in a future that is projected to become drier and warmer.As depth to bedrock can be mapped across large areas using airborne geophysics, such data offer the potential to address the above knowledge gap at practical scales for management.

| CONCLUSION
This study examined geophysical characteristics, vegetation cover and canopy height of forest sites that had experienced canopy die-off.We found that in the historical locations of drought-induced canopy dieoff, current canopy heights and vegetation cover were associated with shallow soils, as imaged by geophysics.Shallow depths to bedrock were coincident with shorter trees with more dead branches, very shallow bedrock areas were treeless and deeper depths to bedrock had taller forest.We also found a critical threshold of 12 m of soil thickness was associated with drought effects.Soils were 20 m thick before drought effects were absent, suggesting large parts of the NJF may be drying to critical levels.With the regional forecast describing future climate drying and warming, we expect outwardly expanding patches of forest in vulnerable, shallow soil areas, to experience dieoff during extreme climatic events such as hotter droughts.This work raises new key questions that can better inform an assessment of forest vulnerability to warming and drying conditions.For example, what fraction of the forest has soils less than the critical threshold thickness?How does the threshold vary regionally?Will the threshold shift to greater depths as conditions warm and dry?The combination of geophysical and ecological measurements has the potential to map current and predict future drought and heat-vulnerable forest sites.
3) fallen logs; (4) trees and shrubs; (5) dead or bare branches (stags); and (6) shadows.Bare ground, litter and photosynthetically active groundcovers were distinguished from other classes using a 0.25 m threshold in the surface height model.While initially separated due to obvious colour and NDVI differences, they were later merged to one class (ground cover) to simplify subsequent analyses, particularly as bare ground may have seasonal cover by winter-spring annual plants, missed by the imaging in early autumn.The classes selected represent the key structural elements of each site while also being related to stand productivity (large logs) and past stress events like drought and fire (dead stems, more bare ground and concentration of more photosynthetic tissue closer to the surface).Short canopies were expected to be indicative of a soil moisture limitation on productivity and similarly, a larger proportion of bare ground and photosynthetically active groundcover relative to trees and shrubs.The mosaicked imagery was displayed in geospatial software, and for each site, 40 to 50 geographic points were created per class for training the classifier.The high resolution of the imagery (8 cm Â 1 cm pixels) meant that the classes were easily discernible.The normalized difference vegetation index (NDVI) Soil depth was estimated from the 2,000 Ω m contour located below the pallid zone clay, where distinct, or otherwise below the resistive surface layer where bedrock looked to be shallow.Estimated soil thicknesses are not particularly sensitive to the choice of a 2,000 Ω m threshold.The sharp interface separating zones with >2,000 Ω m from zones with much lower resistivities, <500 Ω m indicates a narrow interface likely exists between the two materials.This is also consistent with what we know of the geology with low resistivity clays and a weathered saprolite overlying massive and weakly fractured granite/gneiss bedrock (O'Brien et al., 2019).

F
I G U R E 2 Inverted resistivity cross sections and maximum canopy elevations at the study sites: 27 (a), Extra 1 (b) and 166 (c), Northern Jarrah Forest, southwestern Australia.Maximum canopy elevation shown as green lines.Vertical axis denotes the elevation; the horizontal axis is the distance along each transect.Dashed lines demark the interpreted granite/gneiss interface from clay or the dolerite dyke.

1
and 166.Soil thickness was a poor predictor of health class at Site 27.Drought effects are seen where soils are less than 20 m thick and were more likely to be drought-affected in soils less than 12 m F I G U R E 3 Cover types (left) and heights (right) at study Sites 27 (a), 1 (b) and 166 (c), Northern Jarrah Forest, southwestern Australia.T A B L E 2 Mean cover and canopy height characteristics by health class, of the study sites, Northern Jarrah Forest, forest classes.The mechanical thinning at Site 27 prior to the 2010-2011 drought would have reduced below-ground water competition and thus may have ameliorated the magnitude of drought effects Effect of soil depth on maximum (a) and mean (b) canopy heights.Smooth responses from GAM models shown as lines with 95% confidence bands (shaded regions).
).The forest ranges from a tall, closed forest in the south to an open forest in the northeast.The dominant canopy species is Eucalyptus marginata Donn.
Study sites in the Northern Jarrah Forest (grey shaded area), southwestern Australia.quantify vegetation structure in relation to ERT-measured soil thickness, a remotely piloted aircraft (RPA; DJI Phantom 4) was deployed at Sites 27, 1 and 166 in March 2022.The RPA was fitted with a fiveband multispectral camera and was used in conjunction with a realtime kinematic (RTK) mobile base station.At each of the drought sites, RPA captured imagery covering the area within 50 m of the ERT tran- a Mean annual for the period 1981-2010 (Bureau of Meteorology, 2022).bAlongelectrical resistivity tomography transects.cSince1940 with the last three fires at Site27: 2015, 2007 and 1991; Site 1: 2018 Site 1:  , 1998 Site 1:   and 1991;; Site 166: 2017 Site 166:  , 2002 Site 166:   and 1989..F I G U R E 1