Introduction

The Amazon Basin comprises the largest contiguous tropical rainforest on Earth. This region harbours a large number of species1,2 and provides countless ecosystem services such as climate regulation3,4. The Amazon lowland forest is also an important long-term carbon sink5,6,7 that can help mitigate global climate-change. However, climate and land use changes are threatening the biodiversity, ecosystem functions and, particularly, climate regulation abilities of this rainforest4,5. The Amazon is still suffering from very high deforestation rates8 and even more from the adverse effects of forest degradation by selective logging9. At the same time, climate change is leading to warming temperatures and larger-scale droughts in the Amazon basin10. Severe droughts occurred in 2005 and 2010; these droughts were related to positive sea surface temperature (SST) anomalies in the northern tropical Atlantic. Negative rainfall anomalies were mainly observed in south-western Amazonia in 2005 and in south-western Amazonia, Mato Grosso and Bolivia in 2010 while the north-western (NW) area experienced positive rainfall anomalies11,12. During the severe 2015/16 drought caused by the El Niño phenomenon, negative rainfall anomalies occurred all over Amazonia, and the drought strength increased towards the eastern Amazon basin. The warming rates exceeded those of previous major El Niño events (e.g. 1982/83 and 1997/98)13,14,15. In addition to this general tendency towards longer and more intense drought conditions16, future projections conducted for the Amazon Basin have shown either an intensification of drought conditions, with negative rainfall anomalies in the central and eastern Amazon17, or drought hotspots in the south-central Amazon in June-July-August (JJA)18. Positive feedback loops between large-scale deforestation and rainfall reductions19 are predicted to even exacerbate the future drought situation13. As a result, scientists have expressed major concern that the Amazon region is close to a severe tipping point20 in which the westward transport of moisture from the Atlantic will be inhibited by the increasing reduction in evapotranspiration-driven water recycling20,21. This inhibition would endanger not only the carbon sink function2,10 of the Amazon but also its biodiversity22.

The most threatened species are those that directly depend on atmospheric moisture and rainwater such as epiphytic communities in the forest canopy. In general, global forest canopies host ~40% of all existing species and are among the most species-rich habitats23. Canopy epiphytes serve important ecosystem functions such as intercepting rain-water, sequestering carbon and regulating greenhouse gases. These epiphytes further provide habitats and food resources for organisms at higher trophic levels such as canopy ant communities24,25,26,27. Tropical montane cloud forests (TMCF) as well as tropical lowland cloud forests (TLCF) are characterised by both high species richness and high abundances of epiphytes. Both of the forest types receive water supply through cloud immersion. The difference between the two forest types is that advection of clouds leads to water inputs in TMCF, while in areas covered by TLCF longwave radiation and generated katabatic flows result in the formation of fog and low-stratus clouds (FLS)28,29,30. The difference between TLCF and tropical lowland rainforests (TLRF) is that the latter do not experience occult precipitation from FLS. This reduces water availability within the canopy and leads to a reduced radiation shelter both contributing to a reduced epiphyte biodiversity31. A good example for TLCF is an area previously studied in French Guiana28,31,32,33. However, the spatial distribution of this hygric habitat type is unknown to date, though these habitats may provide hygric climate change refugia (HCCR) in the Amazon basin if the favourable microclimate conditions can persist under the projected intensified drought conditions. In general, HCCR provide stable microclimatic conditions under macroclimatic changes such as warming or variable atmospheric moisture supply. These refugia warrant the survival of species that are poorly adapted to the increasingly adverse atmospheric conditions on the macroscale34,35,36,37.

Buffering warming has thus far been thought to be the most desired property of climate-change refugia. However, for moisture-dependent organisms, HCCR are needed to provide an additional buffer against increasing water scarcity38. For many canopy species, regular water inputs in the form of rain or occult precipitation from FLS and dew provide optimal conditions. For non-vascular epiphytes (mosses, liverworts and lichens), it is additionally important how these inputs are timed relative to the timing of light availability, as moisture is quickly lost from these organisms as air dries. FLS can increase the amount of water input to epiphytes and reduce evapotranspiration losses38 by providing a low-radiation, relatively cool and humid environment. Such conditions can be found in TLCF canopies, where relatively high air humidity, reduced morning radiation stress and wetted plant surfaces by FLS water have been observed28. Here, we hypothesise that mainly concave landforms may provide conditions similar to TLCFs over the entire Amazon lowland rainforest and may serve as potential HCCR under drought conditions (Fig. 1a). FLS is expected to preferentially emerge in these landform regions due to the relatively weak atmospheric mixing, increased humidity and low evapotranspiration losses28,38. Furthermore, the nocturnal katabatic flows not only facilitate the formation of FLS but also contribute to the reduction of temperature in valleys, thereby locally mitigating effects of global warming39. However, to date, it is largely unknown to what extent TLCF conditions persist under regional droughts and how these potential HCCR are spatially distributed. This knowledge is urgently needed to prioritise protection measures under climate change.

Fig. 1: FLS formation mechanism, frequencies and TLCF conditions measured in different landform types and sectors of the Amazon Basin.
figure 1

a Idealised cross section illustrating at which topographic position the threshold between TLCF and non-TLCF conditions is expected to be reached. This figure is intended to familiarise the reader with the mechanism of FLS formation in concave terrain of TLCF (for a process-based modelling analysis of this assumption, see S3). b Subset of the FLS frequencies of the three sites used to determine typical FLS thresholds corresponding to TLCF conditions (≥40% FLS). c FLS frequency map (2003–2020) indicating the location of three typical sites with convex (Amazon Tall Tower Observatory ATTO; 13% long-term FLS-frequency) and Station 1; 32% long-term FLS-frequency) and concave (Station 2; 50% long-term FLS-frequency) landforms, shown in 1b, that have undergone canopy epiphyte analyses and thus, TLCF condition analyses28,32,33,39,42. Boxplots depict the seasonal FLS frequency per sector and landform (median and upper/lower quartiles). d Total area of FLS frequencies meeting potential TLCF conditions depending on landform type. AL is the total area of concave and convex landforms in the study domain (grey areas), ATLCF is the area with suitable conditions for TLCF, defined by different FLS frequency thresholds (x-axis).

Full size image

Therefore, the main objective of this study is to utilise novel area-wide FLS data40 to examine the occurrence and resistance of HCCR to past droughts which could mimic future conditions in the Amazon due to climate change. In detail, we analysed 18 years of data representing FLS occurrence over the Amazon basin40 in four sectors (NW, NE, SW and SE). The sectors have been delineated from the drought anomaly patterns of the major Amazon droughts of the past decades15. We specifically investigated FLS frequencies (Fig. 1c) in these sectors, paying particular attention to the terrain and regional drought conditions, in order to understand the persistence of TLCF under varying climatic boundary conditions and identify potential HCCR.

Results

FLS occurrence and TLCF conditions in the Amazon lowland forest

We found that FLS is a common feature over the entire Amazonian lowland forest under normal conditions (Fig. 1c). Approximately 30% of the entire study area exhibits long-term FLS frequencies that are comparable to those found in the locations of French Guiana where TLCF is known to occur. Furthermore, about 15% of the study area has an even higher long-term average FLS frequency of ~50%. The annual FLS frequencies are generally higher in concave landforms than in other landforms and are higher during the rainy season than in the dry season (Fig. 1a, Supplementary Figs. 1, 2) over all sectors (Supplementary Fig. 2). There is an E-W FLS gradient, with higher frequencies in the east, which we suggest to result from the Atlantic moisture advection being reduced towards the western parts (Supplementary Fig. 2). This is supported by the relatively high specific air humidity particularly in the NE sector, which may foster FLS formation (Supplementary Fig. 3). Interestingly, the relatively strong rainfall seasonality in the west (Supplementary Fig. 2) is only weakly reflected in the seasonal patterns of FLS frequencies, which is quite low over the entire Amazon region (Fig. 1c). The NW sector exhibits a great seasonal stability in FLS occurrence; this can be related to the annually stable low saturation deficit in this region (Supplementary Fig. 3). In contrast to the NW sector, where close-to-saturation conditions presumably lead to only small differences in FLS frequencies among landform types (Supplementary Fig. 3), we show for all other sectors that FLS conditions are relatively stable throughout the year in concave landforms. At the same time, we observed a clear decrease in the FLS frequency towards the relatively dry months in convex landforms; this finding is particularly prominent in the SE sector (Fig. 1c).

To determine the FLS range corresponding to potential TLCF conditions, we extracted the FLS frequencies at sites with and without proven TLCF conditions. We revealed that TLCF conditions persist in the Pararé River valleys (concave landforms) where an annual FLS frequency of 40% was found (Fig. 1b, Station 2)31,33,41. We recognised the transition to non-TLCF conditions41 at an FLS threshold below 40% (Fig. 1b, Station 1). Another recent study revealed clear non-TLCF conditions with scarce epiphyte occurrences at very low FLS frequencies in a convex landform situation (Fig. 1b, Amazon Tall Tower Observatory, ATTO)42. Because no exact FLS threshold corresponding to TLCF conditions is known, we analysed the FLS frequency range of 40–60% as potential HCCR. We can show that the conditions required for such HCCR are more frequently found in concave landscapes than in convex landscapes, but convex landforms compose the dominant portion of the Amazon Basin (see Fig. 1d, AL). Increasing the FLS threshold beyond 40% reduces the potential refugial area nonlinearly in both landform types but results in a clearly higher remaining area in concave terrain than in convex terrain (Fig. 1d).

Potential HCCR area per sector and landform type

We found a higher proportion of concave landforms than convex landforms that could be considered HCCR in all four sectors (Fig. 2). However, some convex landforms are also found to be suitable as refugial areas in all sectors. We detect the lowest proportions of these landforms in the SE sector (Fig. 2c). We further showed that the available refugial area decreases nonlinearly with an increasing FLS threshold in all sectors. The NW sector is characterised by the strongest decrease in concave landforms. Here, the concave HCCR area is ~9500 km² smaller than the corresponding TLCF area. Thus, this sector shows the lowest stability (Fig. 2a). The southern sectors show the smallest decreases for concave landforms. Here, the concave HCCR area is ~3800 km² smaller than the corresponding TLCF area, but theTLCF areas begin at a lower amount at the 40% FLS frequency level (Fig. 2c, d). In the NE sector, the concave HCCR area is ~8000 km² smaller than the corresponding TLCF area. The convex HCCR area is ~4000 km² smaller than the convex TLCF area. The sector is characterised by a lower decrease in refugial area corresponding to both landforms compared to the NW sector. Here, the concave HCCR area is ~11,000 km² smaller than the corresponding TLCF area. The convex HCCR is ~9000 km² smaller than the convex TLCF area. The NE sector is more influenced by strong moisture advection from the Atlantic resulting in higher specific humidity conditions (Supplementary Fig. 3).

Fig. 2: Total area providing HCCR conditions within each landform type depending on the established FLS frequency threshold.
figure 2

a NW b NE c SW d SE sectors of the Amazon Basin under normal conditions (ATLCF convex,N, ATLCF concave,N; dark-red and dark-blue areas, all grid-cells with SPEI ≥ −0.5) and drought conditions with SPEI ≤ −0.5 (AHCCR convex,D, AHCCR concave,D; light-red and light-blue areas). Please note that the scale of the y-axis changes at 60*10³ km². From the value 100*10³ km² the axis is compressed by the factor 10. Density distributions represent Monte Carlo Simulations (Methods Section) derived samples of TLCF (dark-red and dark-blue) and HCCR (light-red and light-blue) areas per landform at FLS-thresholds of 40 and 60%. The HCCR area is the area within each sector where the FLS frequency under drought conditions remains higher than or equal to the respective FLS frequency threshold value shown on the x-axis. Potential refugial areas are significantly (p < 0.001) larger in concave landform regions than in convex-landform regions in all sectors and for all FLS frequency thresholds under both normal and drought conditions. The differences in TLCF/ HCCR areas between normal and drought conditions decrease with increasing FLS threshold.

Full size image

Losses of HCCR under drought conditions

Due to the intensification of drought conditions caused by climate change, moisture-dependent canopy species in the Amazon lowland are highly threatened, as noted by43 and22. We used all grid cells that exhibit past drought conditions during the study period as a proxy for future droughts. The definition of droughts is based on the Standardised Precipitation Evapotranspiration Index (SPEI)44,45 where widely used thresholds define drought intensity46. According to the calculation of the SPEI product, SPEI drought conditions represent larger-scale meteorologically driven spatial patterns characterised by negative rainfall anomalies combined with high evapotranspiration losses44,45. We first showed that the duration of drought conditions decreases with increasing drought intensity by analysing four drought severity classes based on the SPEI (S2, Supplementary Fig. 4). While the findings are consistent over all sectors of the Amazon lowland forest, the SW sector shows the lowest number of drought months occurring during droughts of light and moderate intensities. At the same time, we observed the lowest numbers of extreme droughts in the NW and SW sectors. This finding reflects that the western sector of the Amazon lowland is less affected by strong El Niño-type droughts than the other sectors14,15. We then compared the effect of droughts on the potential HCCR conditions per sector for both landform types (see the Methods section) and found a general loss of refugial area over all sectors and landforms under drought conditions (Fig. 2). We showed that concave landforms can conserve more refugial areas under drought conditions, while hardly any suitable area remains in convex landforms at very high FLS thresholds (e.g. 60%, Fig. 1). We observe a significantly stronger effect of drought in the northern sectors for both landform types (Fig. 2, Table 1). We found the largest effect, and thus the least stability, in the NW sector (Fig. 2a). The southern sectors do not show any large effect of additional drought conditions. We found that the overall drought influence is significantly stronger when the FLS threshold is low (Table 1). We assume that the few refugial areas remaining under high FLS thresholds are so extraordinarily moist that even droughts do not considerably limit the suitable local FLS condensation conditions. We suggest that the relatively poor moisture supply available in the southern sectors (Supplementary Fig. 3g, h) under the current low drought frequency (Supplementary Fig. 4c, d) results in nearly stable refugial areas at all FLS thresholds.

Table 1 Statistical results derived from virtual samples calculated by a Monte Carlo Simulation (n = 10% grid-cells per landform per sector, iterations = 500).
Full size table

Because it is expected that droughts in the Amazon Basin will increase in both frequency and intensity for the future13,16,17, we additionally question which potential HCCR areas have remained stable under past strong-intensity droughts (Fig. 3). We showed that under the lowest FLS threshold (40%) and intensive drought conditions (SPEI < −1.5), the refugial areas associated with concave landforms remain larger than those associated with convex landforms. Between the two most severe drought classes, we find that the greatest areal losses occur in the eastern sectors, particularly among concave landforms (Fig. 3b, d). In these sectors, El Niño-type droughts are the most intense11,15. The refugia in the western sectors remain almost stable regardless of the local landform class (Fig. 3a, c).

Fig. 3: Potential HCCR areas and examples of TLCF and HCCR in convex and concave landforms under normal and intensive drought conditions.
figure 3

a–d Potential HCCR areas in different sectors derived using an FLS threshold of 40% under intensive drought conditions per landform class and sector). e TLCF and HCCR areas in western Surinam (4.62° N, 56.78° W, convex) and Brazil, east to Santa Rosa (2.93° S, 53.35° W, concave). Only the upper part of the local mound remains stable as HCCR under normal and intensive drought conditions in the convex landform shape. The mean elevation of the convex HCCR differs comparatively strongly from that of the convex TLCF (difference is +22). The HCCR are located on average 22 metres higher than the TLCF in the same area. The TLCF is spreading throughout the concave valleys. The HCCR are located slightly further towards the centre of the valley compared to the TLCF whereas steep valley slopes and valley centres remain stable. The mean elevation of the concave HCCR differs very slightly from that of the concave TLCF (difference is −1). On average the HCCR are located at almost the same elevation and therefore topographic position in the concave valley as the TLCF.

Full size image

We further analysed the development of the topographic position index (TPI) under intensive drought conditions and different FLS thresholds to understand the role of the terrain configuration in affecting the stability of HCCR. We found that the remaining areas show an increasing steepness associated with both landform types (compare Fig. 3e, f). In convex terrain, only high elevation areas remain HCCR under severe droughts (Fig. 3e). Under intensive droughts, FLS is concentrating in the valley centre while the higher slopes become FLS-free and thus, do not belong anymore to the HCCR (Fig. 3f).

Loss of potential TLCF areas according to modelled deforestation scenarios in the Amazon basin

The ongoing deforestation threatens the TLCF areas and thus the HCCR. This also directly threatens the habitats of tree canopy epiphytes. To test the spatial impact of deforestation, we compared the potential TLCF areas (FLS frequency ≥40%) to two different deforestation scenario models47 in 2050 (method section, S4, Supplementary Fig. 6). Deforestation rates according to the Governance-Scenario would result in loss rates of about 9% of the potential TLCF area on average across all sectors. Deforestation rates as in the Business-As-Usual-Scenario would result in substantially higher loss rates that vary across the sectors. The lowest exposure to modelled deforestation and thus, loss of potential TLCF area is expected for the NW sector. The SW sector is much more vulnerable and could lose up to one-third of its potential TLCF area. The eastern sectors would be at much higher risk and could lose about half of the potential TLCF area (Table 2).

Table 2 Deforestation models47 and derived impact on potential TLCF/HCCR area.
Full size table

Discussion

In this work, we used past drought conditions in the Amazon basin as a surrogate to examine the potential future climate. We found that FLS and, thus, HCCR conditions exist all over the Amazon lowland forest. However, we demonstrated a clear decrease in the FLS frequency during the dry season in the different sectors of the Amazon basin. Relatively low FLS frequencies occur in the southern sector due to its longer dry season, resulting in lower relative HCCR proportions, especially those associated with convex-shaped topography. We found a general decrease in high-FLS areas among all sectors and landform types under past drought conditions. However, concave landforms showed by far the highest persistence of FLS, thus providing favourable microclimatic conditions under regional droughts. We identified concave landforms in the Amazon lowland as stable HCCR given a drought-prone future climate in the region. However, not all concave terrain areas exhibit high resistance under dry conditions. This is because the shape of the terrain serves only as a general indicator of local hydrological HCCR conditions and thus, its use is limited. Local water availability and atmospheric humidity are important for FLS formation. In addition, valley slopes and convex surroundings must be of sufficient steepness and size to allow for cold air production and the development of nocturnal katabatic flows as in the area of the Inselbergs in French Guiana28. Normally, concave terrain can receive soil water and overland flow from adjacent higher terrain, which provides a moisture source necessary for FLS formation, particularly under dry conditions. On the one hand, concave landforms without river courses are less effective atmospheric moisture sources. On the other hand, factors including soil type, depth, and surface roughness can adversely modify cold air, overland, and soil water fluxes and thus hinder FLS formation. All these factors should be considered in future higher-resolution studies on HCCR. We reveal that particularly steep valleys, but also the adjacent convex slopes and slopes at the base of high terrain, can withstand intensive droughts. The reason for this is that nocturnal outgoing radiation losses increase during droughts, thus fostering FLS formation also in the tropics when cold air drainage flows occur in steep terrain28,48,49,50 (Supplementary Fig. 5). To support the role of katabatic flows for FLS formations in tropical lowland valleys we compare the FLS occurrence with spatially derived cold air drainage flows using a process-based model (S3). This process is particularly important in the southern and the NE sector of the Amazon basin, where the adverse effects of strong land-use changes on the regional water cycles are most prominent and exacerbate the effects of global climate change51,52,53. However, we also show that past extreme drought conditions were most prominent in the NE sector. This is related to droughts following strong El Niño events, which are projected to become even longer and more intense under future climate change43,54.

It is of interest how the shelter effect of concave topography against local drying during meteorological droughts is linked to the current debate on Amazon forest dieback mechanisms under combined climate and land use changes. The starting point for forest dieback occurs at a certain level of drying, which corresponds to a reduction of ~3 mm of rainfall per day54. The importance of a closed canopy in the western Amazon due to the increasing fraction of rainwater recycling by the forest was emphasised by5. Only trees can access water from deeper soil layers during drought conditions, preventing drought stress and fire risks in the understory. The importance of maintaining the structure of natural forests to reduce self-amplified forest dieback under increasing meteorological droughts was underscored by55.

Moisture-dependent ecosystems are highly threatened by droughts, enhanced fires55 and forest degradation56 reducing atmospheric moisture. High resistance of HCCR during intensive droughts under sheltering concave terrain conditions will not directly buffer negative rainfall anomalies because the amount of occult precipitation from FLS at canopy level is likely too low to directly supply rainforest trees. However, there are indirect effects to consider. For instance, the reduction of radiation stress by FLS can buffer desiccation and even the risk for fire5. Therefore, persistent HCCR conditions in concave terrain could delay forest dieback, especially in the case of open canopies at the deforestation front, which are particularly vulnerable to increasing meteorological drought conditions5,55. Buffering desiccation in the understory is particularly crucial when the tipping point for self-amplified forest dieback in combination with grass invasion is reached under future climate conditions5. At the same time, the tipping point for a critical reduction of rainfall with regard to forest dieback is at around 30–50% deforestation of the Amazon forests57. Nocturnal and early morning canopy wetting by FLS in HCCR under intensive droughts might help to shift this tipping point towards enlarged deforestation areas by providing higher canopy ET during the day which can foster local rainfall formation. Rainfall formation in the western part of the Amazon is more and more depending on the recycled vapour through ET instead of rainfall directly formed from the Atlantic moisture advection as in the easternmost parts of the Amazon. Thus, the highest dependence of rainfall on ET is found in the far western sectors near the Andes58. Generally, FLS-moistened canopies can enhance moisture support and sustain aerial rivers and lakes58,59. Under arid conditions, the persistence of FLS in concave terrain is relatively high in the western sectors, indicating that these areas likely foster the regional resilience of the forest. In this regard, higher evapotranspiration in FLS-driven HCCR especially close to the deforestation front would contribute to the resistance of the Amazon forest against environmental changes. Based on our results we highly recommend protecting the HCCR, particularly in these vulnerable sectors. We finally suggest that areas with HCCR conditions should be priority areas of conservation, especially in the eastern part of the Amazon basin. In contrast to this urgent suggestion, these areas in the Amazon lowland forest have, to date, been primarily subject to multiple land use conflicts.

Since World War II, development policies for the Amazon have focused on fostering private colonisation programs instead of agrarian reforms in already settled areas60. This led to high deforestation and forest degradation rates, only interrupted by a period of deforestation control and sustainable development (~2004–201261), followed by dismantling environment laws and a lack of environmental governance62. Particularly riverine, and thus, major HCCR areas, were and are heavily affected by dam constructions, illegal mining and often are the starting point for deforestation61,63,64.

Among the focal areas of recent infrastructure programmes, further intensification is planned, including new dam65 projects. Thus, the effects of conserving concave-shaped HCCR areas might go far beyond safeguarding the diversity of moisture-dependent organisms in lowland forest canopies under climate change. First, protecting these areas would help other specialised and endemic organisms cope with climate change66. Second, comprehensive action could also help to prevent adverse effects on river-sediment transport, aquatic biodiversity, greenhouse mitigation and river connectivity67. Third, conservation would help to protect the riparian systems of the Amazon lowland, as these systems are fundamental for providing key ecosystem-regulating services68. Finally, it has been shown that reducing deforestation in areas close to human settlements, such as river valleys, through protection can have far-ranging influences on the resilience of the Amazon lowland forests over quite large spatial scales beyond the areas of refugial valleys10.

It must also be stressed, that riverine areas are heritage sites of great historical value, harbouring most of the archaeological sites and historical settlements of Amazon indigenous people69. While indigenous people are mainly threatened by activities such as land grabbing and illegal mining70 they at the same time have the highest knowledge on locally adapted and sustainable land use practices including agroforestry. Empowering their abilities would therefore be the most effective approaches to forest conservation in riverine areas71,72, thus also safeguarding HCCR. As a consequence, protecting HCCR might have value beyond conservation of biodiversity and stabilising the local hydrological cycle under droughts. Thus, we suggest embedding HCCR protection in a broader governance concept, following a landscape approach considering HCCR and adjacent areas61. Protecting HCCR areas should be aligned with the support of indigenous people and their sustainable agroforestry management, including forest restoration measures in deforested or degraded areas with special reference to maintenance of HCCR functionality. This, however, can only be achieved through a close cooperation of several actor groups: indigenous people, smallholders, the agrobusiness, and local to national governments61.

Methods

Study area

We focus our study on the region of the Amazon lowland forest with elevations ≤500 m above sea level (asl) (Supplementary Fig. 7). This area has been shown to harbour tropical lowland cloud forest (TLCF) ecosystems which are candidates for HCCR, particularly for drought-sensitive canopy organisms28,39,73. The prerequisites that were needed for a region to be included in this study were that (1) the area must be below 500 m a.s.l. according to the Advanced Spaceborne Thermal Emission and Reflection Radiometer global digital elevation map (ASTER GDEM)74 and (2) must harbour long-term pristine lowland forest areas. For the latter, the Food and Agriculture Organization (FAO) Land Cover Classification System (LCCS1)75 was used to define forested areas, and the Moderate-resolution Imaging Spectroradiometer (MODIS) land cover product (MCD12Q1)76,77 was incorporated. For the analysis, only grid cells classified as evergreen broadleaf forest in 2019 and for at least 10 contiguous years previously were considered.

We delineated long-term undisturbed evergreen tropical lowland forest areas based on the FAO LCCS1 and the MCD12Q1 (2001–2019)76,77. First, we aggregated the forest cover dataset at a 30 m spatial resolution to the 1 km spatial resolution of MODIS. This was conducted using a pixel-weighted resolution reduction process. Then, we classified pixels with tree crown coverages ≥60% as tropical evergreen broadleaf forests. Finally, we tested for long-term persistence by using the MCD12Q1 time series (2001–2019)76,77 and retained only pixels in our study area with at least 10 years of forest cover (a). Areas thus identified as long-term forests were filtered to obtain regions with elevations ≤500 m asl using the 1 km resampled ASTER GDEM74 (b). The combination of (a) and (b) yielded the object of investigation: regions in which the conditions implicate potential TLCF, serving as a surrogate of HCCR, where supported by atmospheric conditions (c). This process reduced the evaluated area to 3,443,845 km². We also subdivided the Amazon area into four sectors (NE, NW, SE, and SW) based on the contrasting spatial patterns of recent Amazon droughts (Fig. 1)17.

Derivation of landforms

Landforms are expected to be an important driver of FLS development in the Amazon basin28. In this article, we therefore considered two main types of landform classes: (1) convex-shaped and (2) concave-shaped landforms. Information on these landforms was derived from a digital elevation map (DEM) by calculating the topographic position index (TPI)78. For this purpose, we calculated the relative height of each grid cell within its 7-by-7 neighbourhood. Negative values describe cells that are lower than the mean elevation of the neighbouring pixels (indicating a concave landform shape), while positive values consequently describe pixels that are higher than their neighbouring grid cells (indicating a convex landform shape) (Supplementary Fig. 8). At a constant pixel resolution, a relatively high TPI indicates a greater height difference and thus a higher landform steepness (Fig. 3f). We assumed that pixels with TPI values equal to or larger than zero represented plain or elevated areas and thus classified them as convex landforms (1). Grid cells with negative TPI values were classified as concave landforms.

$${{{{{\rm{convex}}}}}}={{{{{\rm{TPI}}}}}}\ge 0$$
(1)
$${{{{{\rm{concave}}}}}}={{{{{\rm{TPI}}}}}} < 0$$
(2)

The amplitude of the TPI values is particularly important, as strongly negative TPI values account for steep valleys and valley centres (local minima), while highly positive TPI values account for steep terrain rises and high elevations (local maxima).

FLS and meteorological datasets

The study was based on a new comprehensive FLS occurrence dataset constructed over the Amazon basin from night-time satellite images acquired by the Aqua MODIS platform39,40. Aqua MODIS overflights conducted over the South American continent provide thermal infrared imagery between 2 and 6 a.m. local time; these imagery were used to spatially derive daily composites of nocturnal radiation associated with FLS39,40 for the entire study area. Based on reference cloud-free MODIS scenes and a spatially dynamic threshold derived while considering the potential subpixel FLS coverage, FLS pixels were identified using brightness temperature differences ΔBT10.8 µm–3.9 µm40. The derived daily FLS detection composites were subsequently aggregated to quarterly and total FLS frequencies (Fig. 1). We further used the European Centre for Medium-range Weather Forecasts (ECMWF) Reanalysis version 5-Land (ERA5-Land)79 data to analyse how relatively large-scale atmospheric conditions may influence FLS conditions in different sectors. FLS mainly occurs during the night and early morning hours (hours 23–8) as a result of radiation processes and katabatic winds. We thus used hourly ERA5-Land data characterising specific and relative humidity for this timespan. To assess rainfall, wind speed and wind direction daily totals were additionally considered. We further aggregated our results to four sectors in the Amazon (Fig. 1) that might adversely impact FLS occurrence and thus hygric refugia resistance against climate change. We aggregated the precipitation data from ERA5-land to mean-monthly quarterly precipitation sums (December-January-February (DJF), March-April-May (MAM), JJA, and September-October-November (SON)) and divided the results according to the NW, NE, SW, and SE sectors (Fig. 1) to identify the relationship between precipitation, wind direction or wind speed and the FLS frequency (Fig. 1). We also calculated the mean relative and specific humidity on a quarterly basis in each sector to identify the varying condensation conditions that most likely affect the FLS frequency.

Seasonal FLS occurrence

We analysed the seasonal FLS frequencies to understand the spatiotemporal dynamics in the Amazon (Fig. 1). This was necessary because the study area spans two hemispheres and experiences various seasonal courses comprising rainy and dry periods with an east‒west moisture gradient. We compared the sectoral (NW, NE, SW, and SE) FLS frequencies in different topographic positions with the seasonal wind field and atmospheric humidity conditions to separate climatological and terrain-induced spatial patterns throughout the year. We superimposed the average wind field onto the FLS frequency maps to understand the role of advection processes originating from meso- to larger-scale circulations and their interactions with topography with regards to the spatial FLS pattern (Supplementary Fig. 3).

Detection and classification of droughts

We analysed FLS frequencies under normal conditions and drought conditions. For this purpose, we used the Standardised Precipitation Evapotranspiration Index (SPEI) and considered only pixels where normal and divergent drought conditions occurred during the study period (2003–2018). To detect long-term drought effects, drought and non-drought conditions were separated based on the monthly SPEI grids. We use two different classification schemes. In the first classification scheme, droughts were divided from non-droughts. In the second classification scheme, only moderate and extreme drought conditions were considered80 (Supplementary Table 1).

The drought classification schemes provided us with a monthly time series of spatially explicit drought occurrences between 2003 and 2018.

Because no single FLS threshold was supported by the FLS conditions at the sites with known TLCF/HCCR occurrences, we applied a further analysis by considering the continuum between FLS frequencies of 40 and 60%. This analysis consisted of 3 steps:

  1. 1.

    Calculation of a landform-dependent TLCF/HCCR occurrence;

  2. 2.

    Sectoral calculation of a landform-dependent HCCR occurrence;

    and

  3. 3.

    Identification of a drought-dependent HCCR occurrence.

The three steps are described in the following text.

Calculation of a landform-dependent TLCF/HCCR occurrence

To account for the landform-dependent occurrence of TLCF/HCCR conditions, we partitioned the grid cells according to the landform classification. This allowed us to derive the absolute HCCR area in relation to the FLS frequency within each landform and sector as ATLCF/HCCR (FLS).

Drought-dependent HCCR occurrence

In the first step, all pixels were identified for each landform, thus providing HCCR conditions under non-drought conditions corresponding to the respective FLS threshold (AHHCR,N (FLS)). In the second step, drought classes were additionally considered. For each drought class, the corresponding pixels identified in step 1 were extracted, thus still providing HCCR conditions under drought AHCCR,D (FLS).

Calculation of the landform-dependent HCCR occurrence

To account for landform-dependent occurrence HCCR conditions, we partition the grid cells according to the landform classification. This allows us to derive the relative proportion of HCCR in relation to FLS frequencies within each landform considering the following equation:

$${{{{{{\rm{P}}}}}}^{{{{{{\rm{HCCR}}}}}}}}_{{{{{{\rm{L}}}}}}}={{{{{{\rm{A}}}}}}^{{{{{{\rm{HCCR}}}}}}}}_{{{{{{\rm{L}}}}}}}\left({{{{{\rm{FLS}}}}}}\right)/{{{{{{\rm{A}}}}}}}_{{{{{{\rm{L}}}}}}}$$
(3)

with:

PHCCRL,S is the relative proportion of HCCR in the landform area per sector S, AHCCR is the area with HCCR conditions in landform class L[km²] in relation to the FLS threshold, AL is the area of landform class L [km²].

Drought-dependent HCCR occurrence

In the first step, all grid-cells are identified for each landform, which provides TLCF conditions under non-drought conditions for the respective FLS threshold (ATNL(FLS)). In a second step, drought classes are considered. For each drought class, those pixels of step 1 are extracted, which still provide TLCF conditions under drought conditions ATDL,S(FLS). Mathematically, the approach can be denoted:

$${{{{{{\rm{PR}}}}}}^{{{{{{\rm{HCCR}}}}}}}}_{{{{{{\rm{L}}}}}}}\left({{{{{\rm{FLS}}}}}}\right)={{{{{{\rm{A}}}}}}^{{{{{{\rm{HCCRD}}}}}}}}_{{{{{{\rm{L}}}}}}},\left({{{{{\rm{FLS}}}}}}\right)/{{{{{{\rm{A}}}}}}^{{{{{{\rm{HCCRN}}}}}}}}_{{{{{{\rm{L}}}}}}}\left({{{{{\rm{FLS}}}}}}\right)$$
(4)

with:

PRHCCRL(FLS) is the area of drought resistant HCCR in the landform area according to the FLS threshold, AHCCRDL (FLS) is the hygric refugial area under drought conditions in the landform class L [km²], AHCCRNL (FLS) is the HCCR area under non-drought conditions in the landform class L [km²].

Monte Carlo simulation

To test the robustness of observed differences between regions, FLS thresholds and drought conditions, we applied Monte Carlo Simulations for convex and concave landforms (Supplementary Fig. 9). For the FLS-thresholds at 40 and 60%, 10% of grid cells are randomly selected. Then, FLS frequencies are calculated under normal and drought conditions. By repeating this 500 times, distributions of FLS frequencies are derived for FLS thresholds, drought conditions and landforms. These were normally distributed (p value of Shapiro–Wilk-Test = 0.9). The number of iterations was set to 500 to obtain an approximate normal distribution. The sample size was set at 10% to reflect a representative sample of the total data set. Afterward, we applied two-sample t-tests to test for significant differences between (i) landforms and (ii) normal vs. drought conditions.

Modelled deforestation

For the Amazon Basin, predictions of different deforestation rates until the year 2050 from the SimAmazonia model are available47. These modelled deforestation rates are divided into a Business-As-Usual model and a governance model. In the Business-As-Usual scenario, historical deforestation rates are used and combined with major roadway paving. Based on this, a future projection of the deforested area is calculated. Historical deforestation rates are also considered in the governance scenario. However, this scenario sets a maximum deforestation capacity of 50% per Amazon Basin subregion. In addition, existing and planned conservation regions are considered in this scenario. To derive the impact of deforestation trends derived from the different scenarios on the TLCF/ HCCR areas, we first calculated the area of the TLCF (FLS frequency ≥ 40%) in the forested areas according to the governance scenario for the year 2020 in the four described sectors. We then compared the spatial extent of the TLCF area in 2020 to the modelled remaining TLCF area in 2050 using the governance scenario and the Business-As-Usual scenario. Classification as a potential TLCF results from FLS frequency above 40% and identification as a forested grid cell.