Hong Kong is located on the coast of South China and has a marginally tropical climate with an annual average rainfall of 2,325 mm and an annual mean temperature of 23°C (Abbas, 2017). The zonal vegetation is usually considered evergreen broad-leaved forest (Zhuang and Corlett, 1997). However, due to widespread anthropogenic activities over the past several centuries, no primary forest remains in the territory (Zhuang and Corlett, 1997; Corlett, 1999). Our study site is located in the Tai Po Kau Nature (henceforth, TPK) Reserve, which is situated in Hong Kong’s New Territories with an area of 460 ha and a steep, hilly topography. Despite the prevalence of plantations containing a mixture of native (Castanopsis fissa, Cinnamomum camphora, and Pinus massoniana) and exotic (Acacia confusa, Eucalyptus tereticornis, and Lophostemon confertus) species planted between 1956 and 1963, TPK’s numerous remote, narrow gullies have protected small fragments of old-growth forest that pre-date WWII. As such, the reserve is regarded as one of the most biologically important secondary lowland woodlands in the territory (Nicholson, 1996).

A 400 × 500 m 20-ha plot was established inside TPK (22.42°N, 114.17°E) in 2012 (Figure 1A). The elevation of the plot ranges from 214 to 344 m, and the incline of the slope varies from 0.9 to 33.6° (Figure 1B). A full tree census (Condit, 1998; Anderson-Teixeira et al., 2015) was undertaken in the plot in 2015. A total of 81,019 individual trees (117,203 stems, including branches) with DBH ≥ 1 cm were recorded, belonging to 172 species, 111 genera and 53 families. The most abundant species was the shrub Psychotria asiatica with 26,646 individuals, which accounted for 33% of all individuals (Supplementary Table 1). Data pertaining to dead trees (1,876 stems or branches) were removed, including one species that was represented by a single individual, leaving 171 species for further analysis.

We obtained three black-and-white aerial photographs of TPK from the Lands Department of the Hong Kong SAR Government, one from 1956 (Supplementary Figure 1A) representing the initial degraded state, one from 1963 (Supplementary Figure 1B) representing the early regeneration state, and one from 1983 (Supplementary Figure 1C) representing the closed canopy state. As the entire area was forested by 1983, the following assignment of vegetation to transition types was conducted for 1956 and 1963 only. The aerial photographs for these 2 years were firstly geo-referenced, co-registrated and resampled to the same 1 m resolution for comparison. Then, we used the “segment mean shift” algorithms (Comaniciu and Meer, 1999) in ArcGIS 10.7 (ESRI, 2019, Redlands, California, USA) to delineate vegetation parcels with similar texture and grey tone. Segmented vegetation parcels were then divided into three groups based on the K-means clustering of averaged grey tone values within each parcel (Yadav and Sharma, 2013). Finally, manual refinements were made by considering the spatial continuity of each vegetation type.

Four vegetation types were identified in the aerial photographs: grassland, shrubland, pine plantation, and mixed natural forest. Whereas in 1956 the site comprised grassland, shrubland, and mixed natural forest (Supplementary Figure 1D), by 1963 the grassland had been substituted by pine plantation and shrubland (Supplementary Figure 1E). We overlaid the classified vegetation types in both images to derive six vegetation transition types (VTTs; Figure 2), namely, forest to forest (FF), grassland to pine plantation (GP), grassland to shrubland (GS), shrubland to shrubland (SS), shrubland to forest (SF), and shrubland to pine plantation (SP). Each VTT was regarded as reflecting a distinct vegetation change history. Because the FF VTT represented the occurrence of the oldest canopy cover in the plot, this vegetation was taken as reference old-growth forest. Whereas the GS, SS, and SF VTTs represented natural transitions were thus considered passive restoration, the GP VTT entailed management intervention was thus regarded as active restoration. The resulting vegetation transition map was converted to 20 m resolution (500 20 × 20 m subplots, Figure 2) based on the dominant rule (in which each grid cell was classified according to the VTT that accounted for the largest area within it) for further analyses. Since the area of SP (including six 20 × 20 m subplots, Figure 2) was <1 ha, this VTT was excluded from subsequent analysis.

The study area was found to have gradually transformed from a grassland-dominated landscape to a forest-dominated one as a result of natural succession and active restoration over the study period (Figure 3). Whereas 54.4% (108,770 m²) of the plot was grassland in 1956, with scattered patches of mixed natural forest occurring only in ravines along streams and accounting for just 15.9% of the total area (Figure 3), forests occupied 26.1% of the plot in 1963 and 100% in 1983. This transformation was mainly achieved through expansion of remnant natural forest from the streams. However, nearly half (GP: 50,814 m²) of the area occupied by grassland in 1956 had been replaced by pine plantation in 1963, with the remaining half (GS: 57,956 m²) becoming shrubland (Figure 4). More than half of the shrubland present in 1956 was still shrubland in 1963 (SS: 34,783 m²). Only a very small fraction (SP: 4,317 m²) of shrubland had been replaced by pine plantation in 1963.

Stem density was highest in SF and lowest in FF, and whilst there was no significant difference among GS, SS, and FF, it was significantly higher in GP than in SS and FF (Figure 5A). A similar pattern was observed in AGB, although the differences among VTTs were smaller, with that associated with GP being almost the same as for GS (Figure 5B). Differences in stem density were mainly caused by smaller trees (DBH 1–5 cm), while differences in AGB were mainly contributed by the trees with intermediate or large diameters (DBH ≥ 10 cm) (Supplementary Figure 2).

Numbers of species and individuals associated with different VTTs are shown in Table 1. Psychotria asiatica was the dominant species in all vegetation types (Supplementary Table 2). Whereas FF was ascribed the highest values for all Hill’s numbers, diversity curves were the lowest for GP (Figures 6A–F). GS was significantly higher and closer to FF than it was to GP. For the abundance-based curves, Hill’s numbers for richness (q = 0) declined in the order FF > SS > GS > SF > GP. However, 95% CIs overlapped for GS, SS, and FF, indicating no significant difference. Hill’s numbers for the exponential of Shannon’s entropy index (q = 1) and for the inverse of Simpson’s concentration index (q = 2) declined in the order FF > SS > SF > GS > GP. In both cases, most curves were widely separated from one another, apart from GS and SF for q = 1 (Figures 6B,C). In the incidence-based curves, whereas the 95% CIs for GS, SS, and SF overlapped with one another and were close to those for FF, those for GP were much lower and clearly separated (Figures 6D–F). Coverage of sample size approximated to one for both abundance (Figure 6G) and incidence data (Figure 6H).

Species composition was closely correlated with topography and varied across VTTs (Figure 7). Meanwhile, compositional differences were captured along the elevational gradient in all VTTs except for the area of pine plantation represented by GP (Figure 7).

Variation partitioning showed that VTT and topography accounted for 17.1% of total variation in species composition. Whereas VTT alone explained 4.4% of total variance and topography alone explained 9.7%, both factors together explained 2.9% of variance (Figure 8).

Of the 108 species tested in the torus translation test, 70 (64.8%) were significantly associated with at least one VTT (Supplementary Table 2). Twenty-seven species (25.0%) were negatively associated with GP (Table 2), which was the highest number of species among all VTTs, and 17 species (15.7%) were negatively associated with SF. Equally, different VTTs were significantly associated with certain species. FF had the highest number of positively associated species (31) (Table 3), which was significantly greater than that for all other VTTs, whereas FF, GS, and SS had relatively few negatively associated species (Table 2).

Species colonising during early succession are usually regarded as generalists or “winner” species (McKinney and Lockwood, 1999; Zhao et al., 2015) owing to their ability to tolerate, and attain dominance under harsh conditions. This appears to be the case at our study site. The common understorey shrub Psychotria asiatica alone accounted for 33% of all stems, and the fast-growing trees Aporosa dioica, Diospyros morrisiana, and Machilus chekiangensis were among the most abundant species throughout the entire plot (Supplementary Table 1). These species exhibit several ecological traits characteristic of generalists, including fruits that are attractive to common wildlife and which are consequently widely dispersed, and high drought-tolerance during the early growth phase, allowing them to compete with grasses and so colonise grassland (Tabarelli et al., 2012; Zhao et al., 2015). The seed source of these winner species may have been nearby patches of old-growth forest or shrubland during the initial degraded phase, or they could have dispersed into the plot from further afield.

Classical models of forest succession predict that once the canopy has closed, mid- or late-seral species arrive and establish, thereby facilitating community enrichment (Ashton et al., 2001; Poorter et al., 2021). Our results show that this may be unlikely to happen rapidly, given that late-seral species are often rare and overwhelmingly restricted to small patches of old-growth forest (FF), and that seed dispersers are locally extinct, hence limiting opportunities for recovery.

Whilst initial vegetation type was found to have a significant impact on present-day abundance-based diversity indices, impact in terms of stem density, AGB and incidence-based diversity indices, which relate to spatial distribution, was limited. In terms of forest age, the different VVTs identified in the TPK plot decline in the order FF > (SF, SS) > (GP, GS). In old-growth forest, individual tree girth is usually greater and stem density is lower as compared with younger secondary forest (but see Augusto et al., 2000). Indeed, our results confirm that, at TPK, younger forest (i.e., SF, SS, GP, and GS) tends to have a higher stem density than does older forest (FF). This difference appears to be primarily caused by the presence of very large numbers of small saplings and shrubs (1 cm < DBH ≤ 5 cm) in shrubland and young forest.

Higher stem densities in these younger forest communities could indicate that competition among individuals is intense (Picard, 2019) and that natural stem-thinning is therefore likely to ensue (Field et al., 2020; Wu et al., 2020). However, stem density peaks in SF, despite this representing a comparatively advanced state of secondary succession in the sequence observed at TPK, suggesting that self-thinning has failed along this successional pathway. We note that an over-dominance of generalist shrubs (e.g., Psychotria asiatica) and faster growing, early seral trees (e.g., Machilus checkiangensis) in the SF VTT is linked to lower species diversity, higher stem density, and crucially, an absence of late-seral species, as compared with old-growth FF. This appears to limit tree girth increment and prevent species turnover, two features consistent with arrested succession (Kennard, 2002; Whitfeld et al., 2014). The markedly different species composition observed between the SF and FF VTTs reinforces this inference and suggests that the successional pathway taken by post-disturbance secondary vegetation at TPK does not lead to reinstatement of authentic late-seral communities.

Poorter et al. (2016) found that AGB can recover by up to 90% of old-growth values after more than 60 years of succession, and that the impact of land-use history on present-day forest composition and structure may be limited in the neotropics. In contrast, our findings demonstrate that land-use history as reflected by historical VTTs can have a significant impact on the recovery of AGB, even after six decades of post-disturbance recovery, which is more in line with predictions made by Jakovac et al. (2021). Previous studies have shown that AGB is mainly accounted for by large-diameter stems (Stephenson et al., 2014; Lutz et al., 2018) and that large trees are expected to occur more in old-growth forest than in younger associations (Brown et al., 1997). Our results show that in tropical Asia, where old-growth forest has been extensively converted into agricultural lands, this may not be entirely correct. We find that comparatively young secondary communities, represented by GS, GP, and SF in our study, have higher AGB than FF, the oldest forest community in our analysis. This appears to be because the number of small stems is significantly greater in the younger communities and the number of larger trees is low in FF. Indeed, tree size in South China’s remnant old-growth patches are limited owing to several factors. Firstly, although the remnant patches may contain late-seral species, they too have usually been subject to intensive disturbance histories, including fire (Chau and Corlett, 1994), firewood collection (Corlett, 1999), and selective logging (Zhuang and Corlett, 1997) over a period of centuries. Secondly, because these patches are generally restricted to steep ravines, tree size may have been limited by rocky conditions and a lack of top soil due to seasonal flooding, which can lead to stunted tree growth (Lopez and Kursar, 2003). Thirdly, once the canopy of immediately adjacent secondary forest has been able to close through regeneration, the environment along the streams could be too dark and therefore less suitable for remaining late-seral species to recruit and spread (Lebrija-Trejos et al., 2011).

In comparison, GS, GP, and SF appear to have higher AGB because winner, generalist species (McKinney and Lockwood, 1999; Tabarelli et al., 2012; Filgueiras et al., 2021) can successfully establish and grow to a larger size or attain a higher stem density on nearby slopes (Zhao et al., 2015). Indeed, significant increases in AGB are mainly generated by the presence of larger stems (DBH > 20 cm) throughout all VTTs, but because high stem density is associated with higher AGB, it does not necessarily reflect a recovery in species composition (Rozendaal et al., 2019). The exceptionally high number of stems in the young secondary forest communities studied here is primarily made up of generalist and winner species, indicative of arrested succession (Goldsmith et al., 2011).

In analysing forest succession, it is common to substitute space for time to represent a chronosequence (Abbas et al., 2016, 2019; Poorter et al., 2021). In the present study, however, we used historic aerial photographs to identify past vegetation types and reconstruct successional pathways, allowing us to estimate the impact of historic land use more accurately. Although species composition was not documented in each subplot at the time the photographs were taken, we were able to retrospectively determine vegetation type using a statistical approach. By overlaying vegetation type distributions as inferred from the resulting time sequence, the specific disturbance history of each subplot could be inferred. This approach reduces ambiguity and highlights subtle but important differences between diverging successional pathways (Norden et al., 2011; Arroyo-Rodríguez et al., 2017; Jakovac et al., 2021).

Vegetation transition types alone was found to explain 4.4% of variation in present-day species distribution, as compared to 9.7% contributed by topography. A further 2.9% of variation was explained by both factors combined. It is not surprising that the effects of VTT are confounded with those caused by topography, given that both factors tend to act synergistically in constraining ecological tolerance (Guo et al., 2017; Schmitt et al., 2021). For example, it is conceivable that in ravines that are too steep or too rocky for conversion to agriculture, pre-existing forest cover was deliberately retained to maintain water supply (Luke et al., 2017).

Previous studies have shown that, although species richness can recover within a few decades in the tropics (Abbas et al., 2019), species composition may only partially recover (Xu et al., 2015). Similarly, our data reveal that, whereas species richness in the various VTTs was able to recover within six decades, species composition has not converged with that observed in old-growth patches (FF) over the same period, indicating that passive forest succession in these VTTs is slow. Potential underlying reasons for this include dispersal limitation (Palma et al., 2021), reduced species pool (Zobel et al., 1998), priority effects (Fukami, 2015; Fukami et al., 2016), competition (Connell and Slatyer, 1977; Chen et al., 2019), and environmental filtering (Jakovac et al., 2016, 2021), all of which are known to affect both tree seedlings and saplings.

Multiple studies have found that differences in species composition can be explained by dispersal limitation (Condit et al., 2002; Shen et al., 2009; Palma et al., 2021), environmental gradients (Graco-Roza et al., 2022), and stochasticity (Asefa et al., 2020). However, the term “dispersal limitation” may be too general, encompassing numerous inter-related factors. In fact, there may be several inter-dependent reasons for the failure of establishment of late-successional species (Uhl, 1987), such as those which were found to be restricted to FF in the present study.

Firstly, the number of seeds that allow these species to establish in neighbouring areas may be far below the minimum threshold required for successful expansion (Palma et al., 2020) due, for example, to high rates of seed predation (Chung and Corlett, 2006) or because mother trees are confined to less than optimal conditions and are thus suppressed by over-dominant surrounding trees (Ding and Zang, 2021).

Secondly, a species’ ability to germinate and persist will be affected by seed quality, which is itself influenced by genetic factors (Leimu et al., 2010). Trees occurring in old-growth patches within secondary forest are often isolated from other conspecific populations, and hence subject to reduced gene flow (Zeng et al., 2012). Consequently, these species typically exhibit the effects of inbreeding (Leimu et al., 2010; Young et al., 1996) and so exhibit reduced seed viability and depressed seed-set (Kuussaari et al., 2009; Dullinger et al., 2012). Since this is often indicative of extinction debt (Kuussaari et al., 2009), conservation attention is important.

Thirdly, even if the species were able to produce a sufficient number of viable seeds, an absence of suitable dispersal agents may prevent them from recolonising surrounding vegetation types (Corlett, 2017; Silva et al., 2020). Brancalion et al. (2018) have shown that late-seral species tend to have larger fruits or seeds that are dispersed by animals or birds. In our study, a number of species within the FF VTT have larger fruits or seeds that are attractive to animals, for example, Lithocarpus elizabethae, Sloanea sinensis, Artocarpus styracifolius, and Elaeocarpus nitentifolius (Corlett, 2011). Successful dispersal of these species into adjacent areas is dependent on a range of dispersal agents, including hornbills, large squirrels, gibbons and other birds and mammals which are no longer present in Hong Kong (Corlett, 2011). One solution for overcoming this ecological barrier and associated extinction debt would be to reintroduce these locally extinct animal species back into Hong Kong’s secondary forests (Svenning and Faurby, 2017; Carver et al., 2021). However, it is questionable whether today’s secondary vegetation could support the survival of these large dispersal agents: Hong Kong’s secondary forests are mostly too homogenised to provide year-round food supply (van Schaik et al., 1993; Lichstein et al., 2004; Keith et al., 2009; Corlett, 2011).

Pine trees have been widely used for afforestation in Hong Kong (Corlett, 1999). However, the pine trees previously planted in TPK (P. massoniana) were mostly eradicated by wilt disease during the 1970s, resulting in reversion to grassland and shrubland (Zhuang, 1997). The impact of this can be seen in our results: species diversity was significantly greater in GS than in GP, and it approximated more closely to that found in FF, indicating that natural regeneration was more effective than low diversity active restoration in terms of overall ecological recovery. Indeed, species composition in areas previously under pine plantation differed markedly from that in other VTTs, even after decades of recovery and natural succession. The number of negatively associated species within the pine plantation area was much greater than that in other VTTs, particularly with respect to species indicative of old-growth primary forest, such as Turpinia arguta, Mallotus hookerianus, and Ficus pyriformis. Previous research suggests that Pinus species impose an intense priority effect on natural regeneration (Tomimura et al., 2012) through niche pre-emption and modification (Fukami, 2015). In addition, monocultural plantations are known to negatively affect soil properties, including cation exchange capacity and availability of phosphorous and potassium (Caravaca et al., 2002; Maestre and Cortina, 2004), plus they can adversely affect the composition and diversity of the seedling layer (Comita et al., 2010). It thus seems important to recognise that any short- to medium-term benefits associated with monocultural plantations in terms of soil stabilisation and reduced fire risk (Corlett, 1999) could be counteracted by these kinds of long-term detrimental impacts, which can in turn lead to arrested succession (Goldsmith et al., 2011). Indeed, planting low diversity species mixes can be risky (Holl et al., 2022), strongly influencing the successional pathway and available species pool in a way that can mean complete canopy coverage is not attained.

Although biotic homogenisation is often prevalent in secondary forests (McKinney and Lockwood, 1999; Holl et al., 2022), secondary forests themselves are not necessarily homogeneous. In particular, patches of old-growth forest that harbour a diverse set of late-seral species may be scattered within a secondary forest matrix in a non-random manner (Liu et al., 2019; Liu and Slik, 2014). Such patches are likely to retain higher species diversity and thus be of greater conservation value (Cardoso et al., 2022), and our results underscore the importance of ensuring that they serve as reservoirs of regenerative potential by identifying and removing key barriers to spread.

Indeed, we reveal that the effects of historical VTTs can persist for 60 years or longer, demonstrating that past disturbance regimes should be taken into account in studies of community assembly and forest succession in degraded tropical landscapes. Our data indicate that many species are unable to colonise monocultural plantations, reinforcing their artificially low diversity, and secondarily making them more vulnerable to pests and diseases, thereby retarding succession (Siemann and Rogers, 2006; Liu et al., 2018; Wu et al., 2021). Conversely, multispecies forest plantations are likely to offer preferable outcomes in terms of ecosystem services (Feng et al., 2022).

Conservation intervention is clearly necessary to secure the long-term persistence of rare, late-seral species, and overcome the bottlenecks discussed here, by, for instance, selectively thinning the understorey to improve fine-scale conditions that support dispersal, establishment and turnover (Deng et al., 2020; Ding and Zang, 2021). Enhancement planting should also be considered to increase overall biodiversity after thinning (Corlett, 2011; Abbas et al., 2020), and such interventions may not only be critical for trees and shrubs, but also herbaceous elements, too.