1. Introduction

The effects of deforesting the Amazon basin on the global climate have been studied with general circulation models (GCMs; Henderson-Sellers and Gornitz 1984; Dickinson and Henderson-Sellers 1988; Lean and Warrilow 1989; Shukla et al. 1990). While these various modeling studies provide somewhat different results, in general, they agree that deforestation causes a reduction in precipitation and evaporation and an increase in surface temperature in the Amazon basin. They also indicate that the basin deforestation has no detectable, significant impact on the global hydroclimate. However, it is well known that El Niño has a major impact on the hydroclimate of many regions very far away from the eastern Pacific Ocean (e.g., Shabbar et al. 1997). With a relatively warm ocean surface, atmospheric moisture and instability above it are relatively high, providing appropriate conditions for the enhancement of thunderstorm activity. Thunderstorms are the conduit to transfer heat, moisture, and wave energy to higher latitudes (Riehl and Malkus 1958; Riehl and Simpson 1979; Ting 1996), which alter the ridge and trough patterns associated with the polar jet stream (Hou 1998), a mechanism known as a “teleconnection” (Glantz et al. 1991; Namias 1978; Wallace and Gutzler 1981). Since thunderstorms only occur in a relatively small part of the Tropics, not surprisingly, a change in their spatial patterns and frequency has the potential for global hydroclimate consequences.

The “fishbone” land-cover pattern created by deforestation in the Amazon Basin also modifies the frequency and location of thunderstorms in that region (Baidya Roy and Avissar 2002; Avissar et al. 2002). Therefore, Avissar et al. (2002) speculated that tropical deforestation, through teleconnections, should also be able to modify the hydroclimate of remote locations outside of the Tropics. In their numerical simulation of Amazon deforestation, Gedney and Valdes (2000) indeed noted a geopotential wave train, with its attendant precipitation changes, that extended into the winter midlatitudes. And using the National Aeronautics and Space Administration (NASA) Goddard Institute for Space Studies (GISS) general circulation model (GCM; Hansen et al. 1983), Werth and Avissar (2002) identified a clear relation between deforestation of the Amazon basin and a reduction of precipitation in North America.

As deforestation of tropical regions continues at an alarming pace [15 000 km² annually from 1978 to 1988 in the Brazilian Amazon alone according to Skole and Tucker (1993)], it is important to reevaluate conclusions from earlier modeling studies and quantify the effects of such a reduction in rainforest on the local, regional, and global hydroclimate. The main objective of this paper, which is a continuation of the study of Werth and Avissar (2002), is to assess the existence and importance of teleconnections resulting from the deforestation of all tropical regions (hereafter referred to as “land-cover teleconnections”).

2. Numerical experiments

In simulating the climate system with a GCM, it is customary to produce several simulations with perturbed initial conditions to differentiate between the variability due to a “real” signal from the natural variability of the climate system. Each simulation represents an independent “realization” and the set of simulations is referred to as an “ensemble.” Five ensembles of six 12-yr realizations (Werth and Avissar 2004a, manuscript submitted to Geophys. Res. Lett.) were produced with the NASA GISS GCM II (Hansen et al. 1983), with each ensemble representing a different deforestation scenario. The first 4 yr of each realization were discarded as spinup, thus resulting in each month of the year being simulated 48 times (six realizations of 8 yr each) for each ensemble.

In the version used for this study, the NASA GISS GCM II has 12 vertical layers and a horizontal grid size of 4° by 5°. Heat and humidity are advected with a quadratic upstream scheme, and momentum is advected with a second-order scheme. The model has both shallow and deep convection, and a second-order closure planetary boundary layer scheme for moisture and heat transfer is applied at the surface. The model uses six soil layers and a hydrology scheme that accounts for soil moisture transfer and root extraction (Rosenzweig and Abramopoulos 1997), the latter of which depends on the vegetation specified within a grid element.

Observed monthly mean sea surface temperatures (SSTs) derived from multiyear climatological records from the Hadley Centre were used for all simulations. With this type of forcing, the interannual variability that would be introduced in the climate system as a result of year-to-year oceanic variability is intentionally eliminated. While this suppresses important interactions between deforestation and hydroclimatic processes such as the El Niño–La Niña cycle, it considerably simplifies the detection of land-cover teleconnections. Also, with the long-term memory in the simulated climate system limited to only a few months after the ocean feedback has been concealed (Liu and Avissar 1999a, b), any month from a given year is assumed to be statistically independent from that same month in the other years.

In the “control” ensemble, a vegetation map developed by Matthews (1983) for the period starting in 1960 and ending in 1979 (before heavy deforestation started) was adopted for the simulations. In three other separate ensembles, Amazonia, Central Africa, and Southeast Asia were individually deforested and replaced with a mixture of shrubs and grassland. In the fifth ensemble, all three regions were deforested simultaneously. This is referred to as the “total deforestation” case. A global land-cover map illustrating the three deforested regions is presented in Fig. 1.

This land-cover change is expressed in the model by a change of albedo (from a value of 0.06 for tropical forest to 0.1 for the deforested land), vegetation height that affects surface roughness (from 25 to 5 m), leaf area index (from 6 to 1), cumulative root distribution, and stomatal conductance. The cumulative root distribution is defined by root depth and two empirical parameters, which are vegetation specific. For tropical forest, these three parameters are 0.8, 1.1, and 0.4 m, respectively. Corresponding values for the deforested land are 1.5, 0.8, and 0.4 m. Stomatal conductance is controlled by three empirical constants, which are also vegetation specific. It significantly affects the redistribution of energy received at the ground surface into sensible and latent heat. Werth and Avissar (2004b), Costa et al. (2004), and Avissar and Werth (2004) discuss the impact of this parameter on the Amazonian hydroclimatology simulated with the NASA GISS GCM II and other models.

3. Model performance

The NASA GISS GCM II has been used extensively to produce climate scenarios under various forcing conditions, and its capability to reproduce the current climate has been described in various publications (e.g., Hansen et al. 1983; Lau et al. 1996). While an additional detailed evaluation of this model’s performance is beyond the scope of this study, its capability to simulate regional and global precipitation, which is the key parameter for the present study, is briefly discussed below.

Figure 2 provides a comparison between the global monthly merged precipitation analyses of the NASA Global Precipitation Climatology Project (GPCP; see http://precip.gsfc.nasa.gov/ for a description of the data product) and the ensemble-mean control simulation. The GPCP images are based on data collected during the 25-yr period, from January 1979 until January 2004. Figures 3–5 provide comparisons of the annual mean, zonal average, and global precipitation, respectively, between the control simulations and the GPCP data.

While the general spatial and temporal pattern of precipitation is quite well represented by the model (Figs. 2–4), in general, it overestimates precipitation by about 10%–20%. This is well illustrated by the global precipitation comparison presented in Fig. 5. This seems to be due mostly to the too high precipitation simulated by the model near the highest mountains (i.e., the Andes and the Himalayas) and, more generally, near significant topographical features in tropical regions (e.g., Indonesia). This overestimate is particularly noticeable during the wet season in these areas (Fig. 2). The zonal means depicted in Fig. 4 confirm this tropical/topographical bias. It should be mentioned that this deficiency is not specific to the NASA GISS GCM II, as it appears that other GCMs suffer from the same problem (e.g., McCabe and Dettinger 1995).

In evaluating the significance of this bias between the model and the GPCP data, one needs to keep in mind that the model is forced by multiyear, monthly mean SSTs and fixed land-cover types dated from the 1960s. These conditions eliminate some of the interannual variability, which might have an impact on the monthly, annual, and global mean, through nonlinear interactions in the climate system. Furthermore, at the resolution of the NASA GISS GCM II, topography is only schematically represented by the model, and the numerical techniques used to approximate atmospheric flow near topography tend to create erroneous motion in the mountain lee.

However, it is also noticeable that the three regions where the deforestation experiments are being performed, namely Amazonia, Central Africa, and Southeast Asia, are generally well simulated by the model.

4. Impacts of deforestation

Figure 6 shows those locations worldwide where precipitation has either significantly decreased or increased for at least 3 months of the year as a result of deforestation of Amazonia (see next section for statistical analysis). The annual cycle and change of monthly mean rainfall at a few continental locations mostly affected in this scenario are also provided. The selection of these locations was subjectively made based on our perception of where the most impressive impacts occur, both in terms of absolute magnitude and time duration. For brevity, it is not possible (and justified) to present such an annual cycle for all locations (which are indicated by the color-coded points on the world map), and the purpose of the few examples presented here is just to illustrate various intricate ways by which the annual cycle of precipitation is perturbed by the deforestation in Amazonia.

Figures 7–9 present the corresponding results for the deforestation of Central Africa, Southeast Asia, and the three tropical regions together, respectively. Locally, deforestation causes a significant reduction of precipitation in all three tropical regions during the entire year. However, the season mostly affected by the deforestation is not the same in the three cases. In Amazonia, the wet season (summer) registers a decrease of precipitation as high as 50%–60% in December and January, but the dry season (winter), which experiences only a small amount of rainfall, is barely affected by the deforestation. In Central Africa and Southeast Asia, the impact of deforestation on rainfall is spread over the entire year. However, it is the dry season (June–July in Central Africa and May–June in Southeast Asia) that mostly suffers from deforestation, with a decrease of precipitation on the order of 30% during that period, versus a decrease of 10%–20% during the wet season. When all three regions are deforested together, the local impact of deforestation in each one of the tropical regions remains dominant as demonstrated by the lack of significant change in the annual cycle of precipitation of the three regions depicted in Fig. 9, as compared to those obtained in Figs. 6 –8.

While the major impact of deforestation on precipitation is found in and near the deforested regions themselves, a strong impact is propagated by teleconnections along the equatorial regions and, to a lesser yet still statistically significant extent, to midlatitudes and even high latitudes. One can notice that, as a result of the deforestation of Amazonia, the largest decrease of precipitation in continental regions outside of the Tropics is seen in North America, where this deforestation causes a decrease of rainfall in the Gulf of Mexico region, with a particularly severe impact in Texas (about 25%) and northern Mexico, during the spring and summer seasons. Asia, mostly in a region spreading from Turkistan in the west to the Gobi Desert in the east, is also affected by the deforestation of Amazonia. However, the extent of the impact in that region is less significant from a water resource perspective given the small, absolute amount of rainfall received there.

Deforestation of Central Africa causes a decrease of precipitation of about 5%–15% in the Great Lakes region, mostly centered in Illinois, with a peak decrease of about 35% in February. It also affects Ukraine and Russia (north of the Black Sea), where precipitation there is reduced by as much as 25% in May. The impact of the deforestation of Southeast Asia is mostly felt in China and the Balkan Peninsula, with a decrease of 20%–25% in western Turkey during a large part of the year.

As illustrated in Fig. 5, the global amount of precipitation is unaffected by the deforestation in any of these numerical experiments (only the control and total deforestation scenarios are presented in Fig. 5, but the three individual deforestation cases did not produce any different curves than the two superposed ones in this figure). This implies that the regional decreases described above are counterbalanced by an enhancement of precipitation at other locations, as is indeed depicted in Figs. 6 –9. Most remarkable is the impact that each one of the three deforested regions has on the rainfall in the Arabian Peninsula and East Africa (near the Red Sea). Deforestation of Amazonia results in an increase of rainfall as high as 45% in August and September, and the deforestation of Central Africa and Southeast Asia enhances the rainfall in that region by 15%–30% during this period (note, however, the decrease of precipitation during the June–July period). Other regions, such as northern Europe and North Africa, also experience a precipitation enhancement most of the time as a result of these deforestations, yet it is far less significant than in the Arabian Peninsula (Figs. 6 and 8).

The total deforestation case reveals that the global impact of deforestation is not equivalent to the cumulative impacts of the three individual regions, emphasizing some synergy (i.e., nonlinear effects) between the three deforested regions. For example, the strongest impact in the United States has shifted to California during the winter as a result of total tropical deforestation. The Midwest now experiences a less severe decrease of precipitation during a shorter period of the year, which nevertheless remains quite significant. There is a noticeable decrease of precipitation in southeast Africa that was not identified in the individual deforestations, and the enhancement of precipitation seen in northern Europe and Siberia as a result of the deforestation of Amazonia alone has disappeared. But a few regions do experience a cumulative impact. Most notably, the southeastern part of the Arabian Peninsula and northeastern India, which register enhancements of 70% and 35% of their summer (i.e., August–September) precipitation, respectively.

We explain these findings as follows: As a result of tropical deforestation, the sensible and latent heat released into the atmosphere is considerably altered (Shukla and Mintz 1982). The associated change of pressure distribution modifies the zones of atmospheric convergence and divergence, which shift the typical pattern of the Polar Jet Stream and the precipitation that it engenders as far away from the Tropics as mid- and high latitudes. Such a mechanism is not unique to deforestation, as it is also the probable reason for the impacts of El Niño on the global weather and climate (e.g., Trenberth et al. 1998).

5. Statistical analysis

For each scenario described in the previous sections, we assessed the impact of deforestation by calculating the change of monthly mean precipitation (averaged over the 48 simulated months) relative to the control case at each grid point of the GCM. We formulated the hypothesis that the possible change detected for each month of the year at a particular location (i.e., grid point) was caused by the deforestation. To test that hypothesis, Student’s t value was calculated using the means and standard deviations of the control and deforested ensembles. With 94 degrees of freedom [(2 × 48)−2], a one-tailed t value (increases and decreases are considered separately) of 1.66 indicates significance at 95%. This is similar to the method outlined by Chervin and Schneider (1976), but the formula used to get the combined variance is from Edwards (1971). To strengthen our confidence that the observed impact was indeed meaningful, we subjectively ignored any decrease or increase of precipitation that lasted less than 3 months a year. Thus, all points colored in Figs. 6 –9 indicate a statistically significant impact of deforestation that lasted at least 3 months. It should be noted that a statistically significant change of precipitation does not necessarily mean that its absolute magnitude is important enough to have any practical implication from a hydrometeorological point of view.

We also adopted the “Randomized Intervention Analysis” proposed by Carpenter et al. (1989) to check the impact of deforestation on precipitation away from the deforested regions. Accordingly, the root-mean-square difference (rmsd) between the ensemble, monthly mean global precipitation of the control and deforested scenarios (hereafter referred to as “true” ensembles) was calculated and compared to the rmsd of “false” ensembles of control and deforestation scenarios. These false ensembles were obtained by randomly combining three control and three “deforested” realizations performed for each scenario. A total of 200 such ensembles was obtained for each scenario. Figure 10 summarizes the results of this comparison. It clearly shows that in all months for the total deforestation scenario and most months for the individual deforestation cases, the rmsd of the true ensembles are more significant than those of the false ensemble.

Figure 10 also presents histograms of the rmsd of annual-mean global precipitation calculated for each of these 200 false ensembles and the position on these histograms of the true ensembles. For the Amazonian deforestation and the total deforestation cases, the annual-mean rmsd of the true ensemble was higher than any of the 200 false ensembles, clearly emphasizing that the precipitation difference obtained between the control and deforested cases is undoubtedly due to the land-cover change. While this measure is less significant for the deforestation of Central Africa and Southeast Asia (above 85% and 75% of the false ensembles, respectively), it remains nevertheless a strong signal based on this analysis. Reducing the period considered for this analysis from the full year to the season mostly affected by deforestation (i.e., September–December for Central Africa and March–June for Southeast Asia), positioned the true ensemble in the upper 93% and 89% of the histogram for Central Africa and southeast Asia, respectively.

Finally, we performed a correlation analysis between meteorological variables in the deforested regions and at the main remote locations identified to have experienced a statistically significant change of rainfall to emphasize the teleconnections between the regions. In general, strong correlations are found in all cases.

6. Conclusions

While our numerical experiments were carefully designed to provide a high level of statistical confidence in our results, we caution that our discovery is based on numerical simulations performed with a single GCM, with no interannual SST variability. Thus, our results emphasize the potential importance of land-cover teleconnections in hydroclimate studies, but they were not designed to simulate a specific, realistic hydroclimate. In spite of this cautioning remark, we trust that some of the strongest events identified in our study will be found as well with other GCMs under a more natural interacting climate system. To gain more confidence on the realism of our findings, we suggest that reproducing our experiments with other GCMs is desirable. Using global coupled ocean–atmosphere models could provide additional insights on the spatial and temporal variability of the teleconnections identified here. We are currently in the process of conducting such additional numerical experiments with the GISS Atmosphere Ocean Model (AOM2; Lucarini and Russell 2002), the Community Atmosphere Model (CAM) developed and maintained by the National Center for Atmospheric Research (Kiehl et al. 1998), and with the Ocean–Land–Atmosphere Model (OLAM), a new generation of earth system model developed at Duke University (Walko and Avissar 2004, manuscript submitted to Environ. Fluid Mech.).

While this study is still preliminary, if additional experiments with other GCMs and more realistic SST confirm its findings, land-cover change in tropical regions may have quite devastating consequences on agriculture, water resources, and related activities at various remote locations. This emphasizes that not only regional emissions of pollutants have global hydroclimatological impacts, but that regional land-cover change is another parameter that needs to be considered in climate change policies.

Explaining the dynamics of the land–atmosphere interactions involved in these land-cover teleconnections is not trivial. While our preliminary analysis indicates interesting correlations between the hydrometeorology of the deforested regions and that of remote areas, a careful analysis of geopotential heights, radiative forcing, and moisture convergence and divergence is still required to elucidate the main mechanisms involved in these global-scale interactions. But the activation of Rossby waves is likely involved in such teleconnections. We are in the process of performing such an analysis and plan to publish its results in a future paper.