Loss of Amazon Rainforest Resilience: Conclusions

Table of Links

• Abstract and Introduction
• Materials and Methods
• Data
• Results
• Conclusions, Open Research Section and Conflict of Interest/Competing Interests
• Author Contributions, Supplementary Information, Acknowledgments and References
• Additional Insight from the Conceptual Model
• AMSR2 until 2022 Excluded Due to Missing Human Land Use Data
• Investigation of Resilience Loss in AMSR2’s X-Band
• Results Proving Robustness
• Potential Driving Forces and Sources of Bias

CONCLUSIONS

In spatially coupled systems, the spatial correlation is expected to increase prior to a critical transition, establishing an indicator of CSD. In this work we first used a conceptual model to show that for a system with spatial extension and coupling similar to the ARF, variance, AC1 and spatial correlation increase as it approaches a critical transition. This is the case even if a cascade of tipping is induced by a single cell (Bathiany et al., 2013b). The simulations revealed that for strongly coupled cells where a transition is caused by a reduction of the incoming recycled moisture, spatial correlation is an especially reliable and early mean of detecting the loss of resilience and an approaching transition.

Recent studies have shown that satellite data are appropriate only under certain conditions for investigating changes in the resilience of the ARF (Smith et al., 2022, 2023). In particular, it has been shown that time series which combine different data sources might inherit artifacts resulting from the merging procedure (Smith et al., 2023), and thus in our work we exclusively analyzed single sensor data.

While time series of several decades would be favorable to capture long-term vegetation dynamics, shorter time series are still capable of sensing physiological responses to droughts and other environmental conditions. The sensors AMSR-E and AMSR2 provide acceptably long VOD time series (2002-2011 and 2012-2020, respectively), which we analyzed on a monthly resolution, following Boulton et al. (2022). For the early 2000s we find an overall increase of the CSD indicators, with more striking signs of resilience loss in AMSR-E’s X-band. The spatial pattern is consistent across the two bands, with the largest losses of resilience occurring in the southwest and north. From 2012 to 2020, AMSR2 data reveals a less clear picture. Yet, the cells in the C1-band where all three indicators increase reside mostly in the northeast, coinciding with the resilience loss detected by AMSR-E in the preceding years. The cells that are likely undergoing destabilization according to AMSR2’s C2-band are concentrated in the southwest.

Overall, even though the results differ somewhat for the individual data sets, we can conclude that the ARF’s vegetation experienced a loss of resilience during the first two decades of the 21st century. More pronounced signals were found for the time period from 2002 to 2011, with the regions of destabilization comprising the western Amazon basin, the band along the northern boundary as well as the northeastern parts. Interestingly, the regions in the southwest where destabilization is detected in all data sets correspond to the regions downstream from the ‘atmospheric rivers’. Hence, they are highly dependent on moisture recycling, implying that they are considerably spatially coupled and found to be more vulnerable to tipping due to network effects (Wunderling et al., 2022). In combination with the results from the conceptual model, which show that spatial correlation gives an especially reliable indication of CSD in a highly coupled sub-system, the spatial correlation can be considered the most reliable indicator in the southwest. This is in line with the fact that all data sets show increasing spatial correlation in parts of the southwestern Amazon. These increases could hint at a destabilization due to changes in incoming recycled moisture, which in return could be an effect of the high deforestation rates in the ‘arc of deforestation’ further upstream of the ‘atmospheric rivers’.

The main forcing affecting the ARF’s vegetation and potential resilience changes is presumably the precipitation. Thus, it is essential to ensure that the detected changes in the vegetation’s indicators of CSD are not a direct representation of corresponding changes in precipitation statistics. We thus calculate the same CSD indicators for the precipitation at each grid cell, and analyse the regions in which the sign of their trends agree with the sign of the trend in VOD CSD indicators (see Fig. S16). This analysis shows that the signals found in Fig. 6 are not a consequence of statistical changes in precipitation. On the other hand, it is still possible that the vegetation resilience loss is related to a decrease in the annual precipitation. Yet Fig. S17 shows that the signs of CSD cannot always be explained by negative trends in precipitation. However, this could be caused by a lag in time between the changes in precipitation and the resilience loss in vegetation. Furthermore, the lack of a negative trend in mean annual precipitation could still coincide with overall drying: in case of an increasing evapo-transpirative demand or, as many studies have suggested, shifts in the dry season length and strength, and increasing frequency of droughts that could be evened out by increases in precipitation during the rainy season or flood years (Marengo et al., 2018). Thus, further work is needed to better understand the interplay of causes that can drive the ARF towards a dieback.

This work has focused on spatial correlation as an indicator of CSD, due to the spatially coupled structure of the ARF and the theoretical advantages this indicator shows in numerical experiments. Yet, multiple other potential (spatial) CSD indicators exist, such as spatial variance, spatial autocorrelation, spatial skewness (Dakos et al., 2011; Donangelo et al., 2010; Dakos et al., 2010; Guttal and Jayaprakash, 2009), or spatial permutation entropy (Tirabassi and Masoller, 2023) . Several of these would be applicable in this setting, but their thorough investigation and comparison was beyond the scope of this study. Still, a comparison of different spatial resilience indicators could improve our understanding of their applicability as well as their reliability to detect changes in resilience in the ARF, as well as in other spatially coupled ecosystems.

If we are to robustly capture resilience changes, our efforts must be focused in a few key directions. First, long single-sensor time series are preferred to reliably trace the dynamics and potential resilience changes of vegetation ecosystems. Second, sensors and their derived VIs must be adequate to address the question of interest. To that end, it is important to find measures to assess the suitability of data sets. In the context of residuals-based resilience analyses, a sufficiently high signal-to-noise ratio is crucial. Furthermore, with respect to dense vegetation such as the ARF, it would be helpful to better understand problems induced by saturation in the VIs on their higher order statistics (Smith et al., 2023). Third, the applicability of indicators of CSD in different settings must be better understood, such that the most suitable approaches can be chosen depending on the system analyzed.

Overall, the complex changes we find in the ARF suggest that combining multiple datasets and indicators can give a clearer picture on the applicability of CSD, and the statistical robustness of trends in different parts of the Amazon.

Our findings suggest that the previously found loss of resilience in the early 2000s (Boulton et al., 2022; Smith et al., 2022) can, in parts, be confirmed by our approach and data, with less distinct signals for the years 2012 to 2020. Nevertheless, we find a destabilization of the vegetation in the ARF since the beginning of the century, independent of the data source or the indicator of resilience change, that is especially pronounced in the southwestern Amazon basin.

OPEN RESEARCH SECTION

All data used in this study is publicly available.

For this study, only cells within the Amazon basin (//worldmap.maps.arcgis.com/home/ item.html?id=f2c5f8762d1847fdbcc321716fb79e5a, accessed on January 28, 2021) are considered. Human Land Use is extracted from the MODIS Land Cover dataset (Friedl and Sulla-Menashe, 2015) (MCD12C1, Version 6) available at //lpdaac.usgs.gov/products/mcd12c1v006/(accessed on November 11, 2021), based on Land Cover Type 1 in percent. The Hansen deforestation data (Hansen et al., 2013) was downloaded on May 31, 2022, from //storage. googleapis.com/earthenginepartners-hansen/GFC-2021-v1.9/download.html.

The VOD from AMSR-E (Vrije Universiteit Amsterdam (Richard de Jeu) and NASA GSFC (Manfred Owe), 2011) (LPRM-AMSR_E_L3_D_SOILM3_V002, C- and X-band) and AMSR2 (Vrije Universiteit Amsterdam (Richard de Jeu) and NASA GSFC (Manfred Owe), 2014) (LPRM-AMSR2_L3_D_SOILM3_VC1- and C2-band) can be found at //hydro1.gesdisc.eosdis.nasa.gov/data/WAOB/LPRM_ AMSRE_D_SOILM3.002 and //hydro1.gesdisc.eosdis.nasa.gov/data/WAOB/LPRM_AMSR2_D_SOILM3.001/ and were last accessed on November 8 and December 24, 2022, respectively. The precipitation data from CHIRPS (Funk et al., 2015) is available at //data.chc.ucsb.edu/products/CHIRPS 2.0/global_monthly/netcdf/ and was last accessed on November 16, 2022. It was downscaled to the same resolution as the VOD data by selecting only the grid cells matching VOD’s grid cell centers (center of 5 × 5 cells).

CONFLICT OF INTEREST/COMPETING INTERESTS

The authors declare no competing interests.