Our study area is located within the southern Black Forest, in the south-west of Germany. The Black Forest is characterised by large elevation gradients (120 – 1,493 m) and corresponding variation in mean annual temperature ranging from 4 °C to 10.4 °C (Gauer and Aldinger, 2005). It covers about 5,000 km², of which 75% are forested.

The study plots are part of the ConFoBi (Conservation of Forest Biodiversity in Multiple-Use Landscapes of Central Europe) research network, comprised of 135 one-hectare forest plots in a mix of even- and uneven-aged, selectively harvested continuous cover forests (Storch et al., 2020). These plots were selected to assess relationships between forest biodiversity and the retention of forest structural elements in a managed forest context. Among public state forests, including the study area, a retention program was launched (ForstBW, 2015) that obliges the retention of approximately 15 habitat trees per 3 ha in state forests. In addition, efforts to support forest biodiversity on private land include contractual nature conservation, advisory services, and subsidies (Storch et al., 2020).

All plots are in temperate mixed low montane forest dominated by Norway spruce (Picea abies (L.) H.Karst.), European beech (Fagus sylvatica L.) and silver fir (Abies alba MIll). The plots have mostly closed canopies and are comprised of mature stands > 70 yrs. In 2018, a subset of 39 plots were randomly selected for acoustic recording to identify the acoustic index that best represented bird diversity within the study area. In 2019, a larger random subset of plots was selected (n = 80) and the survey was repeated ( Figure 1 ). Twenty-nine percent of the surveyed plots overlapped between the recording years (27 out of 92 plots).

We selected forest structural variables of known importance to forest bird diversity, and that also were collected in a 2017 live and dead wood tree species inventory (see Storch et al., 2020 for details). Our aim was to test our hypothesis with structural variables that are important to birds and also commonly quantified in national forest inventories, both in Germany (Kändler, 2006; Reise et al., 2019) and more widely across Europe (Motta, 2010; Gschwantner et al., 2016), to maintain relevancy for forest managers. Table 1 describes the final twelve variables selected to collectively describe plot-level forest structural heterogeneity, through diameter at breast height (DBH) and its standard deviation (DBH SD), tree species richness, ratio of coniferous to broadleaved species, tree density via basal area (BA) and number of living trees as well as standing and lying deadwood amount, size, height, and volume. We acknowledge that including a tree height variable would have provided a more complete picture of structural heterogeneity within our plots, but we were unable to do so due to lack of data available for this variable).

Acoustic surveys took place in the spring of 2018 and 2019 using LUNILETTRONIK Soundscape Explorer Terrestrial prototypes (Lunilettronik Cooperativa, Fivizzano, Italy). The recorders were equipped with one external microphone (EMY-63 M/P), with a sensitivity of −38 ± 3 dB and a SNR > 60 dB (EMY-63 M/P, sensitivity (0 dB = 1 V Pa-1. 1 kHz), input voltage of the ADC: 0.75 Vrms (personal communication with Lunilettronik Coop.) Recorders were installed 1.5 m high on trees, with the microphone facing perpendicular to the slope, when slope was present (Shaw et al., 2021b). Recordings were taken one minute every ten minutes from 4:30 - 9:00, gain of 25 db and a 24 kHz sampling rate. Recordings were saved in 16-bit WAV format on SanDisk 32 GB SD cards.

Recorders were in the field from 18.05.2018 - 10.06.2018 and 12.04.2019 - 04.06.2019 for each campaign. Bad weather days (presence of high wind or rain) were identified and removed from the dataset (5 days in 2018 and 29 days in 2019; details in Appendix 1 ). Given that sunrise timing shifts over the course of spring, we calculated each recording’s time in relation to sunrise. We used the coordinates of each plot centre to generate the exact sunrise time (defined as when the top edge of the sun appears on the horizon) for each plot-day-year using the ‘suncalc’ package (Thieurmel and Elmarhraoui, 2019) in R statistical computing environment (R Core Team, 2022).

A sub-selection of the one-minute acoustic files were manually reviewed by one expert observer to identify all recorded bird vocalisations to species level. To increase the probability of capturing birds whose vocal activity peaks at different points in spring, we selected one day from early-, mid- and late-spring per plot for bioacoustic identification. The campaign in 2018 was shorter than in 2019, spanning 25 compared to 55 days, respectively. Therefore, two days were randomly selected for bioacoustic identification from the 2018 campaign (one day from the first and last week of recording period), corresponding with mid- and late-spring. Three days (one from the first, middle and last week of the recording period) were randomly selected from the 2019 campaign, corresponding to early-, mid- and late-spring.

For each selected day, we sub-selected eight files per day for bioacoustic identification. Using time in relation to sunrise, we selected files to cover pre-dawn (n = 4) and late morning (n = 4) periods. We selected one minute every 20 minutes for the pre-dawn and late morning hours (minutes since sunrise = -60, -40, -20, 0, 160, 180, 200, 220). This selection method increased our probability of detecting the widest range of species across one morning (Shaw et al., 2022), in addition to species most vocally active at different times throughout spring, due to their migratory (residents, short- and long-distance migrants) or breeding phenology. The total number of files (n = 2,544) resulted in 42.4 hours of identification in the lab. The expert reviewer identified all vocalisations to species level (n = 10,660 observations) with her corresponding certainty (0 - 100%). All identifications under 100% certainty were removed (n = 77), mostly due to vocalisations being too faint or masked by other sounds. The final variable of bird species richness was computed, defined as all the unique bird species identified per plot per spring year.

We calculated acoustic indices on for each 1-minute audio file and only on good weather days (n = 71,229 data points across both years; identification of good weather days provided in Appendix 1 ). We selected the full suite of acoustic indices commonly used in ecoacoustic studies and available in the ‘soundecology’ package (Villanueva-Rivera and Pijanowski, 2018); these included the Acoustic Complexity Index (Pieretti et al., 2011), Acoustic Diversity Index (Villanueva-Rivera et al., 2011), Bioacoustic Index (Boelman et al., 2007), Evenness Index (Villanueva-Rivera et al., 2011) and Normalized Difference Soundscape Index (Kasten et al., 2012). All default settings were used for index computation, except for the ACI, for which we set the minimum frequency at 2 kHz, to automatically filter low-frequency noise that was observed in our files. We then filtered the dataset by retaining index values from recordings within 60 minutes before sunset and 220 minutes after sunset, corresponding with the shift in sunrise and the timeframe from which bioacoustic identifications were made. Lastly, we calculated the mean acoustic index values per plot, to eliminate the temporal autocorrelation in the dataset, as suggested in best practice guidelines for the suite of acoustic indices (Bradfer-Lawrence et al., 2019; 2023). Our final set of ecoacoustic variables was the mean acoustic index value per plot, per spring year, for the five acoustic indices.

All analyses were made in R statistical computing environment. Because the strength of the relationship between acoustic indices and bird richness varies across studies (Alcocer et al., 2022), we first conducted a Pearson’s correlation analysis between each acoustic index and bird richness in the 2018 dataset to identify the acoustic index that was most strongly correlated with bird richness in our study area (Step 1), using the ‘stats’ and ‘corrplot’ packages (Wei and Simko, 2021; R Core Team, 2022). Once that index was identified, we repeated the process on the 2019 dataset, to ensure the index’s reliability over a second year of monitoring. Interannual reliability is important because bird assemblages and their activity can vary; in our study area these years were notably different in terms of rainfall and temperature (2018 was a drought year). Only indices that were significantly correlated with bird abundance in both years were selected for further analysis.

Once we identified the most suitable index to reflect bird species richness in our study environment (Step 1), we used the 2019 data for our main investigation of how structural complexity affects bird richness (Step 2-3) and acoustic index values (Steps 4-5). We first aimed to assess which individual structures, if any, were significantly linked to bird richness (Step 2). To test the relative effects of each structural variable separately, we built twelve linear regression models with the ‘stats’ package and evaluated their relative effects on bird richness using an information-theoretic approach (Burnham and Anderson, 1998) using the ‘AICcmodavg’ package (Mazerolle, 2020). This approach allows for the identification of the ‘best’ model in the set of candidate models, which has highest predictive value, with the lowest corrected Akaike information criterion (AICc) score (Akaike, 1998; Burnham and Anderson, 1998). The Akaike weight (wi) describes what percentage of the time, if this study were repeated many times, a given model would be the optimal model (lowest AICc), given the available data (Garamszegi, 2011). Model diagnostics were performed with the ‘DHARMa’ package and model residuals were checked for spatial autocorrelation using the ‘testSpatialAutocorrelation’ function (Hartig, 2022). Any models showing significant autocorrelation via the Moran’s I test were further investigated (see Appendix 2 ).

The values of each structural variable were scaled between 0 and 1, to facilitate the comparison of effect size between models. Nine out of twelve variables were log- or square root-transformed, to meet assumptions of normality for linear regressions. Complete structural data was available for 75 of the 80 research plots from 2019; only these data points were used in our models. We assessed spatial autocorrelation in our models using the ‘DHARMa’ package (Hartig, 2022), given that the plots were not evenly distributed throughout the study area ( Figure 1 ). We conducted a Pearson’s correlation using the ‘corrplot and ‘qgraph’ packages (Epskamp et al., 2012; Wei and Simko, 2021) to assess the correlation coefficients of our structural variables, to aid in our interpretation of model results.

Next, we evaluated the combined effect of structural variables on bird richness (Step 3). We conducted a non-metric multidimensional scaling (NMDS) analysis based on Euclidean distance, which can coerce the data into two ordination axes. We then assessed the relationship between the combined structural variables and bird richness by fitting and plotting a generalised additive model (GAM) using a 2D surface smooth of bird richness values onto the ordination plot. This analysis was done with the ‘vegan’ package (Oksanen et al., 2022), using the ‘ordisurf’ command and a Gaussian distribution.

Lastly, we repeated Steps 2 and 3 using the selected acoustic index as the independent variable, in response to individual forest structural elements (Step 4) and overall structural heterogeneity (Step 5), which enabled us to evaluate potential similarities or differences in both dependent variables’ respective responses to forest structural elements.

Across 39 plots in 2018, bird richness ranged from 10-19. The NDSI showed the highest and only significant correlation to bird richness, with a correlation coefficient of 0.316 and p-value of 0.049 ( Figure 2 ). We found non-significant correlation coefficients for ACI, ADI, AEI and BIO, being 0.032, -0.024, 0.060 and 0.267, respectively. The NDSI, therefore, was selected for further analysis with the second year of data (Step 1). Across 80 plots in 2019, bird richness ranged from 4-25, and NDSI values for the same year ranged from -0.02 to 0.82. The significant positive correlation between the NDSI and bird richness held for the 2019 dataset: the correlation coefficient was 0.510, p-value < 0.001, a stronger relationship than the previous year. The first step in this study was to confirm the assumption that our hypothesis is built upon, that bird richness is significantly positively related to acoustic diversity (quantified by a robust acoustic index), which we confirmed with the NDSI.

Next, we sought to identify which individual structures, if any, were significantly linked to bird richness (Step 2). Our information theoretic approach used mean bird richness values per plot as the dependent variable and one structural variable at a time as the independent variable, resulting in twelve models ( Table 2 ). Tree species richness was the best performing model and the only variable that significantly affected bird richness, although variance explained was relatively low (adjusted R² = 0.08; Figure 3 ). We found evidence of spatial autocorrelation (Moran’s I test < 0.05) in no models except in the tree richness model, which showed a mild effect (estimate versus expected: -0.07, -0.01, p-value = 0.02). We did not alter this model’s structure, however, to enable its comparison with other models in Table 2 (see Appendix 2 for details and justification).

The NMDS of the twelve structural variables (Step 3) successfully converged with a stress value of 0.177, which is considered a satisfactory representation of the plots relative to each other when reduced to two dimensions (Clarke, 1993). The gam smooth of bird richness values onto the NMDS ordination revealed a significant effect of aggregate structural heterogeneity on bird richness (p-value 0.001, adjusted R² 0.20. Highest bird richness values were associated with high tree richness, tree age variability (DBH SD) and snag height variability, all which describe stand-level structural heterogeneity. Moderately high bird richness was observed in plots with larger trees (mean DBH), taller standing dead trees (snag height) and plots that were generally denser (BA). Bird richness was lowest in plots with high conifer share, snag DBH mean and SD and lying deadwood volume ( Figure 4 ).

When we repeated the individual models with the NDSI as the dependent variable (Step 4), the best-performing models were the standard deviation of DBH, standard deviation of snag DBH, snag DBH and lying deadwood volume, in decreasing order ( Figure 5 ), which all showed significant negative effects on the NDSI ( Table 3 ). These models, however, also explained little variance (adjusted R² = 0.08, 0.08, 0.06 and 0.04, respectively). No other structural variables significantly affected the NDSI, and we found no evidence of spatial autocorrelation in these models.

Compared to the bird richness models (Step 2), we observed a clear divergence between the structures that significantly affected NDSI values ( Figure 5 ) versus those that affected bird richness ( Figure 3 ). Additionally, the valence of the variables differed by dependent variable: the best-performing bird richness model showed a positive effect on bird richness, while the best-performing NDSI models showed a negative effect on NDSI values.

The Pearson correlation analysis between all structural variables described a complex web of correlations ( Figure 6 ), which allowed us to interpret model results between bird richness and the NDSI. Tree species richness, the significant variable positively affecting bird richness, was negatively correlated with DBH SD, the variable with the strongest negative affect on NDSI values. All four variables that negatively affected NDSI values (snag DBH SD, DBH SD, snag DBH and lying deadwood volume) were positively correlated with one another.

Average tree size (mean DBH) and number of trees were strongly negatively correlated, which is intuitive, as younger forests tend to be denser. Other notable relationships were between deadwood variables: snag height SD was positively correlated with snag height, snag DBH and DBH SD. Plots with high deadwood volume were negatively correlated with conifer share, meaning that high deadwood volume occurred in plots with more broadleaved trees species, while taller standing deadwood (snag height) was more likely to occur on plots with more conifers.

When repeating the NMDS analysis with NDSI (Step 5), we observed a very similar pattern to the bird richness GAM overlay, with high NDSI values significantly associated with the same plots as high bird richness (p-value 0.01, adjusted R² 0.11; Figure 7 ). Comparing the bird richness model to the NDSI model ( Figures 4 vs 7 ), the smooths diverged in two minor ways: plots with higher lying deadwood volume and number of snags had relatively moderate bird richness, while in those plots NDSI values continued a decreasing trend, skewing the otherwise similar shapes of both NMDS plots ( Figure 7 ). Secondly, highest NDSI values were predicted to occur in similar plots as for bird richness, although they were less associated with snag height SD and shifted slightly to the right of the plot, centring instead around plots with high basal area, snag height and mean DBH.

Confirming our assumption, bird richness showed a significant positive response to tree species richness (linear regression, Figure 2 ) and structural heterogeneity generally (NMDS gam smooth, Figure 4 ). Tree species richness had the strongest and only significant effect on bird richness in our set of candidate models, consistent with previous findings from other temperate (Fuller, 2000; Poulsen, 2002; Hanle et al., 2020) and boreal European forests (Versluijs et al., 2019). Other studies have shown that stands with higher tree species richness generally show higher canopy packing due to larger crowns of trees growing in mixed stands than in pure stands (Jucker et al., 2015), which increases stand structural complexity (Pretzsch, 2014). Previous research in our study area confirmed that tree richness is significantly positively correlated with understory and shrub layer species richness (data made available by Helbach et al., 2022). A diversity of vegetation in these lower strata provide a higher diversity of resources and leaf architecture at the stand level. This structural complexity increases opportunities for concealment and nesting on the ground or in shrubs and increase food supply from invertebrates, seeds or fruits, which have been linked to higher bird richness (Robinson and Holmes, 1982; Holmes and Schultz, 1988; Whelan, 2001; Muiruri et al., 2016; Shutt et al., 2018; Adams and Matthews, 2019; Hanle et al., 2020). Species-specific preferences for tree and shrub species have been observed in birds (Holmes and Robinson, 1981; Gabbe et al., 2002), thus it is likely that the increase in diversity of tree species richness and its associated understory plants would result in a greater diversity of birds in our plots (“diversity-begets-diversity”; San Roman and Wagner, 2021).

According to the NMDS model, bird richness was predicted to decrease in plots with increasing share of conifers. This is consistent with other research in managed European forests (Czeszczewik et al., 2015; Kebrle et al., 2021; Basile et al., 2021b; Zeller et al., 2023). Forests with higher percentage of broadleaved species host greater abundances of lepidoptera larvae (Summerville and Crist, 2003), an important source of protein for birds and their chicks in the breeding season. Indeed, an increase of broadleaved species in an otherwise conifer-dominated landscape—as is the case in our study area—generally benefits the entire bird assemblage via its provisioning of diversified food sources, most notably defoliating larvae (Vélová et al., 2021). This does not apply to all species, however, such as the goldcrest (Regulus regulus), crested tit (Lophophanes cristatus) or seedeaters like the Eurasian bullfinch (Pyrrhula pyrrhula), whose nesting or food preferences are specialised to conifers (Mikusiński et al., 2018; Basile et al., 2021b).

Contrary to our expectations, bird richness did not increase with deadwood variables. Rather, variables representing deadwood amount and variability (lying deadwood volume, number of snags and snag DBH SD) were non-significant as single predictors of bird diversity. These features were also associated with lower bird richness in the NMDS model, likely because plots associated with higher deadwood were different from the plots that otherwise had higher structural heterogeneity (tree species richness, DBH mean and SD, snag height, etc. in Figure 4 ) Deadwood is not an essential component for all forest birds, but rather for the guild of primary and secondary cavity nesters, such as woodpeckers, tits, owls, nuthatches and treecreepers (Mikusiński et al., 2018). In natural forests, high amounts of deadwood are often associated with other structural features contributing to structural heterogeneity (Bobiec, 2002; Hedwall and Mikusiński, 2015), but in our study area that has a history of forest management, the plots with high amounts of standing and lying deadwood were not the same plots that have a higher diversity of tree species and age classes, each which provision different microhabitats (Spînu et al., 2022). In our area, plots with high amounts of deadwood would not necessarily provide a sufficient diversity of additional niches for species from other nesting and foraging guilds to affect the richness of the entire bird assemblage (i.e. bird richness values), but more likely increase abundances of species that are reliant on deadwood.

Basile et al. (2021b) found in our study area that local structures were insufficient for predicting bird diversity, and that landscape-scale metrics such as broadleaf cover in the surrounding 5 – 25 km² and distance between forest patches also partially predict stand-level bird richness. Indeed, patterns of spatial autocorrelation indicate slight landscape-scale patterning in bird richness values, albeit mostly non-statistically significant ( Appendix 2 ). Both of our analytic approaches support this conclusion, that in this region, bird richness cannot be strongly predicted by stand-level structures alone. Additional metrics such as landscape-level variables would better predict bird richness (above 8 or 20% variance explained). Further, the more intensively managed a stand, the more it undergoes habitat simplification, which deprives birds of important resources such as suitable nest sites or food supply. They are then forced to exploit a broader range of environmental conditions, or increase their niche breadth (Mahon et al., 2016), which has been observed in our study area (Basile et al., 2021a). The degree to which bird richness will respond to individual forest structures is therefore additionally dependent on the intensity of management that stand is subject to, and the shift in generalist/specialist behaviours of the birds using that habitat (e.g. shifting their foraging or nesting locations, Basile et al., 2021a). The relationship between bird richness and local forest structures will therefore be modified by the surrounding landscape-level context, including the percentage of broadleaved forested area in an otherwise coniferous landscape (Basile et al., 2021a), a dearth of unmanaged forest patches (Storch et al., 2020) and the general homogeneity of the landscape matrix, which in our study area is comprised primarily of even-aged stands of low tree diversity.

Our results suggest, however, that one variable (tree richness) is sufficient to predict bird richness ( Figures 2 , 4 ). This is a notable finding, as tree richness is a variable that can both be easily measured and altered in a managed forest context to promote bird diversity. Tree species admixtures can also positively influence forest multifunctionality and resilience, as well as other forest-dwelling taxa (Ampoorter et al., 2020; Messier et al., 2021). Managing forests for higher tree diversity, and the associated diversity of canopy (Pretzsch, 2014) and understory structures (Helbach et al., 2022) that it begets, can improve bird diversity at the local assemblage level. Thus, our results support the recommendation of increasing tree species diversity in managed forests (Messier et al., 2021) to deliver more biodiversity-oriented forest management outcomes. Indeed, our results also support using tree species richness as a proxy for promoting bird richness, however acoustic index variation also provides within-season and interannual variations in the bird assemblage, which tree species richness cannot.

As outlined in the introduction, previous soundscape studies have examined the link between acoustic indices and habitat structural heterogeneity, the goal being that from increasing acoustic index values one can infer increasing habitat quality as it pertains to provisioning of habitats for birds. Acoustic indices have successfully distinguished between structural differences in ecosystems, although this is mainly differences between coarse habitat type classifications such as forests, pastures, plantations, or urban to rural gradients (Joo et al., 2011; Bobryk et al., 2016; Buxton et al., 2018; Turner et al., 2018; Bradfer-Lawrence et al., 2020; Do Nascimento et al., 2020; Hayashi et al., 2020; Dröge et al., 2021). There are fewer studies that identified differences in acoustic diversity within the same habitat type, as finer differences are more difficult to capture (Dröge et al., 2021). However, Shaw et al. (2021a) did so in a winter boreal forest, Chen et al. (2021) in nature reserves in southern China, Atemasov and Atemasova (2019) in oak forests in Ukraine and Farina and Pieretti (2014) in a Mediterranean shrubland. Our results support the latter studies that it is possible to distinguish between forests of lower or higher structural complexity within the same habitat classification, which in our case were low montane temperate managed forest stands.

At the assemblage level, we hypothesised that higher structural heterogeneity would provide a greater niche volume for more bird species, and that this richness will be reflected in acoustic index values. However, this is not a strictly linear relationship. Studies have found that bird species that prefer dense vegetation (e.g. blackbirds (Turdus merula) (Linnaeus, 1758) or robins (Erithacus rubecula)) (Linnaeus, 1758) can disproportionately contribute to the soundscape (Farina et al., 2015). This may explain why the highest NDSI values on the NMDS plots occur around plots with high basal area ( Figure 7 ), compared to bird richness ( Figure 4 ). Denser plots (high basal area) may thus in some cases yield higher acoustic index values without necessarily supporting higher bird richness, a limitation of this monitoring method.

Moreover, our study was not able to estimate bird abundance, which is also known to affect acoustic index values (Alcocer et al., 2022) and the NDSI specifically (Bradfer-Lawrence et al., 2020). Although we could not estimate abundance from our recordings, this explanation is consistent with findings from other work conducted in our study area showing that bird abundance, more than richness, was positively associated with local habitat quality (Basile et al., 2020, 2021b). Further, point count data collected during the same two-month period confirms an extremely high correlation between bird richness and abundance in this study area (Pearson’s correlation = 0.92, p-value = 2.2e-16; data made available by Basile et al., 2021a).

The NMDS model explained similarly low variance as with single forest structures (adjusted R² = 0.11 versus 0.08, 0.08, 0.06 and 0.04), while the variance explained by the bird richness NMDS more than doubled (adjusted R² = 0.20 versus 0.08). This suggests that the NDSI is subject to additional influences outside of forest structural heterogeneity and that it is not as sensitive as bird richness to one individual structure, but rather the overall structural heterogeneity within the plots. This is supported by the similar Akaike weights assigned to significant NDSI regressions (wi = 0.36, 0.22, 0.17, 0.09), while the wi was not as evenly distributed across bird richness models. The NDSI may differ from bird richness because it reflects dynamic components of a bird assemblage that the bird richness metric does not quantify. Species abundance, stage of breeding cycle, territory establishment and defence, predator alarms and chick begging are all conveyed by type, frequency and amplitude of vocalisations (Gil and Llusia, 2020) and would thus be reflected in NDSI, but not bird richness, values. These unknown species-specific factors highlight the NDSI as not just a proxy for bird richness but also a representation of a wider range of breeding season dynamics that are reflected through bird vocalisations.

Further, the NDSI was designed to measures the ratio of biophony to anthrophony, with biophony defined as any sound above 2 kHz, the typical range of biological sounds, including bird vocalisations. Anthrophony is defined as sounds below 2 kHz, which is the range at which low-frequency sounds from human machinery, such as road and air traffic or agricultural activity often occur (Kasten et al., 2012). The presence of faint background noise from roads and forestry operations varied per plot, and insofar as birds continue to use those plots despite this background noise, the NDSI would reflect this additional plot feature, while bird richness would not. Limitations, therefore, exist to this method, specifically that there is not a consistent, linear relationship between the NDSI, or any acoustic index, to bird richness; the strength of this relationship between the acoustic index and bird richness can be subject to changing environmental conditions, such as ambient anthropogenic noise. Appendix 4 investigates the effect of anthropogenic noise on the dataset, and indeed the plots with high bird richness and NDSI values were the plots with the lowest amount of anthropogenic noise. Our recordings were comprised almost exclusively of biophony (bird vocalisations), so the index’s ability to reflect bird richness was not drastically impeded by a preponderance of geophonic ( Appendix 1 ) or anthropogenic ( Appendix 4 ) noise, but rather occurring where there was a lack of it. There was greater structural heterogeneity where there was also less anthropogenic noise in our study area, the relative effects of which we cannot disentangle, but nonetheless which the NDSI values accurately reflect. Other study areas containing more anthropogenic sound sources, or anthropogenic noise near higher structurally heterogenous forest stands, might observe different results with this particular index. Overall, in our study area, the NDSI, successfully reflected not just bird richness (through positive significant correlations over two years), but also the response of birds to forest structural heterogeneity ( Figure 7 ).