How does local tropical deforestation affect rainfall?



[1] The aim of this study is to investigate the potential impacts of vegetation-breezes on locally-generated rainfall and its distribution on the mesoscale. Ensembles of simulations with a 2D large-eddy model were performed using various heterogeneous land surfaces. Rainfall was found to be 4–6 times higher over warmer surface anomalies, associated with cropland, compared to a homogeneous surface, but rainfall was reduced to half or less over the forest. While the suppression of rainfall tended to occur throughout the forest with an intensity comparable to the surface anomaly, the exact location of the maximum in rainfall was less predictable. The location of peak rainfall depended on an interplay between the size of the heat flux gradient (governing the vegetation-breeze strength), the size of the anomaly (as vegetation-breezes organize in certain preferential length-scales), and the distance to other anomalies (since convection in one location could suppress it elsewhere). The presence of surface heterogeneity also increased the total rainfall in the domain by 13% on average, with higher increases in the presence of more intense surface variabilities.

1. Introduction

[2] Land-use change, such as conversion of part of a forest to pasture, can lead to sharp gradients in surface fluxes, which in turn can generate thermally-induced circulations analogous to sea-breezes. Convergence of vegetation-breezes can lead to enhanced shallow convection over the warm temperature anomalies [Wang et al., 2000; Roy and Avissar, 2002; Kawase et al., 2008; Roy, 2009]. This enhancement of convection is due both to an enhancement of vertical motion at the convergence zones, aiding convective initiation, as well as an increase in the convective available potential energy (CAPE) at the convergence zones increasing the depth and organization of clouds [Garcia-Carreras et al., 2011] (hereafter GC11).

[3] Vegetation-breezes have been directly observed by aircraft measurements over Benin in West Africa, with convection occurring at the convergence zones over the cropland boundaries [Garcia-Carreras et al., 2010]. Satellite studies have also extensively corroborated the link between land cover and convection over deforested regions [Garcia-Carreras et al., 2010; Wang et al., 2000; Roy and Avissar, 2002; Wang et al., 2009], consistent with initiation of convection caused by land-surface induced flows.

[4] There has been relatively little work looking at the subsequent impact on rainfall amounts and distributions. Mesoscale modelling of the impact of fishbone deforestation in Amazonia has found that resolving mesoscale processes is important to correctly characterize the rainfall [Ramos da Silva and Avissar, 2006] and that mesoscale circulations lead to increased cloud cover and rainfall over the deforested patches [Roy, 2009]. This could attenuate the decrease in rainfall caused by deforestation predicted by coarser resolution models [Ramos da Silva et al., 2008]. Satellite rainfall measurements also show enhanced rainfall over the deforested side of the land-surface boundaries and reduced rainfall either side [Negri et al., 2004; Knox et al., 2011]. These results are consistent with the presence of peak CAPE at the convergence zone with subsidence either side, as found by GC11.

[5] The impact of deforestation on rainfall distributions is crucial to understand subsequent feedbacks back to the surface, and thus the long-term impact of deforestation. For example, enhanced rainfall over deforested regions may enhance the regeneration of the forest, thus acting as a negative feedback by reducing the initial perturbation to the surface. On the other hand, West African rainforests, together with those of the Congo, have lower mean precipitation rates than any other rainforest in the world [Malhi and Wright, 2004]. These regions may therefore, potentially, be more sensitive to reductions in precipitation, leading to vegetation degradation beyond the deforested region itself. These possibilities highlight a non-linearity of the feedback of deforestation on the water cycle which has not been investigated in depth. Given that deforestation rates in West Africa are particularly high [Food and Agriculture Organization, 2009], understanding these feedback mechanisms is of particular importance for the region.

[6] The aim of this study is to determine the impact of different vegetation heterogeneities on locally-generated rainfall distributions, in particular differences in rainfall over cropland and forest. Various realistic distributions of surface patterns are used to look at the impact of the size and locations of the surface anomalies on the resulting mean rainfall distributions.

2. Methodology

[7] Version 2.4 of the Met Office Large Eddy Model (LEM) [Gray et al., 2001], a nonhydrostatic model with a Boussinesq equation set, was run in 2D for the simulations, using a similar setup to GC11. The main difference in this study has been the inclusion of rainfall with a full 3-phase microphysics parameterisation. A 312.5 km horizontal and 13 km vertical domain was used, with a horizontal resolution of 250 m and a varying vertical resolution (14–240 m over 107 levels). The model was run from 06:00–24:00 LT.

[8] The use of a 2D model was imposed by computational constraints, particularly given the large number of runs that were performed. The impact of using 2D as opposed to 3D is partly discussed in GC11. The use of 2D may emphasize the presence of mesoscale circulations, as these are artificially confined in one dimension. Clouds in a 2D model experience reduced entrainment of environmental air, but also weaker updraughts due to a moister lower troposphere, which is a result of lower detrainment from shallow clouds. Overall cloud cover is generally similar for 2D and 3D simulations [Petch et al., 2008]. The focus of this study is on the distribution of rainfall and the impact of boundary layer dynamics on this distribution, as opposed to quantifying total rainfall amounts absolutely. Furthermore, a comparison of 2D and 3D in GC11 showed a similar cloud cover pattern, although details differed, suggesting that the use of a 2D model is justified.

[9] The same initial and prescribed surface boundary conditions as in GC11 were used, which in turn followed an observational test-case over Benin [Garcia-Carreras et al., 2010]. The initial conditions were taken from ECMWF analysis data on 16 August 2006, and the diurnal cycle of surface sensible and latent heat fluxes was estimated from ground station data in Nangantchori, Benin. A run with a homogeneous surface was used as a control (CTL). The surface heterogeneity was imposed by varying the Bowen ratio (lower over forest, higher over crop) but keeping the net radiation constant. The domain mean sensible and latent heat fluxes were the same as in the control run.

[10] One ensemble was run with the same surface flux heterogeneity as in GC11, derived from remote-sensing data (HET). In addition to this, another 4 random land surfaces (RN1–RN4) were generated by creating red-noise distributions with variability at similar length-scales to HET. This was done by using an autoregressive function, i.e. a distribution where each value corresponds to a weighted sum of its previous values with a white noise error, tuned to give variability of the desired wavelengths. The range of the distribution was then set to be the same as HET. The actual land-surfaces are shown in the top plots of Figure 1. The largest heat-flux gradient among all the simulated land surfaces is found in HET, as the full range is covered by a single boundary (50–60 km). To explore the impact of the magnitude of the gradient at land-surface boundaries, a final ensemble (RN5) was run with the same surface distribution as RN4 but doubling the Bowen ratio offset range. RN4 was chosen as it had the smallest land-surface gradients.

Figure 1.

(top) Ensemble mean daily rainfall and (bottom) Bowen ratio offset (black and red lines respectively) and their wavelets (coloured and line contours respectively) for (a) HET, (b) RN1, (c) RN2, (d) RN3, (e) RN4 and (f) RN5. In the top plots the grey shading represents the interquartile range. In the bottom plots the line contours correspond to powers of 0.1,1,2 and 5 W. Solid lines are for positive wavelets and dashed lines for negative wavelets.

[11] For each land surface, an ensemble of 20 runs was performed. The only difference between the ensemble runs was a random perturbation to the initial conditions of 0.2 K in temperature and 0.3 g/kg in specific humidity at each point. For a 20 run ensemble, the mean rainfall in CTL was relatively smooth, with few outliers. The domain-mean rainfall also converged to a single value between using 10 and 20 runs, suggesting that the average behaviour was well represented by 20 runs.

3. Results

[12] The dynamical features of the simulations are similar to GC11, with some differences due to the presence of rainfall-generated cold-pool outflows (not shown). As in GC11, the heterogeneity in surface fluxes leads to temperature differences of up to 1 K between cropland and forest. The temperature gradients drive the winds, producing convergence over the cropland boundaries and divergence over the forest. As in GC11, the wind patterns organize the cloud cover, with divergence suppressing cloud, and convergence leading to “primary” triggering of convective clouds. Once showers develop, low-level cooling arising from downdraughts enhances the temperature gradients at the land-surface boundaries, and displaces these gradients further, with “cold pools” of air [e.g., Simpson, 1999] propagating across the domain. The cold pools lead to convergence, causing “secondary” triggering of convective clouds at their leading edge (or gust front).

[13] The effect of the cold pools is that the resulting cloud cover distribution is less closely linked to the land-surface in these simulations compared to GC11, due to secondary triggering by the cold-pool outflows away from the land-surface boundaries, where they originate early in the afternoon. However, because rainfall is concentrated in the early afternoon hours (14:00–16:00) and where clouds are deepest, the rainfall distribution is still closely linked to the land-surface features, as can be seen in Figure 1. Thus, although rainfall has an impact on patterns of convergence in the PBL via the presence of cold pool outflows, the cold pools do not have a large impact on the observed land-atmosphere coupling.

[14] In order to better understand the link between the land-surface heterogeneity and rainfall amounts, both the amplitude and spatial extent of the surface anomalies need to be taken into account. To do this, wavelet analysis has been used. Wavelet analysis decomposes a distribution into wavenumber space as a function of domain position. In this study, the wavelet chosen was the derivative of a Gaussian function. The advantage of this wavelet is that it is well suited for describing isolated peaks, which are characteristic of rainfall distributions, and positive and negative signals can be separated, so that the influence of cropland (positive surface anomaly) and forest (negative surface anomaly) can be assessed separately [Torrence and Compo, 1998]. In Figure 1 the power of negative anomalies is multiplied by −1 to differentiate them from positive anomalies.

[15] Figure 1 shows, for all 6 land surfaces, the Bowen ratio multiplier, which describes the factor by which the Bowen ratio was varied from the mean at each point, the ensemble daily mean precipitation (Figure 1 (top plots) red and black lines respectively), and their respective wavelet powers (Figure 1 (bottom plots), line and coloured contours respectively). The wavelet power spectrum is given by the absolute value of the wavelet transform squared, similar to the power spectrum derived from a Fourier transform. The precipitation wavelet was normalised by the domain mean wavelet power at each wavelength (λ) for the CTL ensemble, to account for the background bias towards longer wavelengths. The grey shading in the top plots represents the ensemble interquartile range in precipitation amounts.

[16] In all cases the heterogeneous land surface exerts a strong control on locally-generated precipitation, leading to strong gradients in rainfall amounts. In CTL the average rainfall is 1.02 ± 0.02 mm, homogeneously spread across the domain (not shown). In the heterogeneous runs precipitation is concentrated on the cropland boundaries, with mean rainfall amounts of between 3.5–6 mm. The interquartile range is evenly distributed around the mean at these peaks, consistent with persistence over many runs as opposed to a single intense event dominating the mean. The precipitation over the forest is strongly suppressed, with rainfall amounts consistently below 1 mm (e.g., −160 – −100 km in HET), decreasing to negligible amounts close to certain boundaries (e.g., −15 and 80 km in HET). This contrast between cropland and forest can lead to a persistent difference in rainfall of an order of magnitude over a distance of just 10 km (c.f. 110–120 km in HET). This link between the land surface and precipitation is robust for all ensembles, even when the surface flux gradients are not particularly strong, noting that the full Bowen ratio offset range represents a maximum change in sensible heat fluxes of 60 W/m2.

[17] This rainfall pattern is consistent with the mechanisms identified in GC11. At the vegetation-breeze convergence zones CAPE is higher compared to both adjacent crop and forest. This is due to vertical redistribution at the convergence zone of high CAPE air originating from the forest. The enhancement of CAPE causes increased convective development, leading to enhanced rainfall at the convergence zones. This effect is particularly significant for the ensemble runs given that the location of the convergence zones is persistent over all the runs. Subsidence over the forest forms a warm capping layer strongly suppressing cloud cover, leading to the reduction in precipitation observed in Figure 1. It is also worth noting that the rainfall pattern cannot be attributed to the magnitude of fluxes alone, but is associated with the presence of the boundaries. This can be observed in the difference in rainfall amounts between −20 km and 40 km in RN2, or −50 km and −10 km in RN4 and RN5.

[18] Figure 1 (bottom plots) reiterate the strong link between the land cover and rainfall, with the peaks in wavelet power closely collocated, albeit with a positive y-direction shift of about 10 km in precipitation relative to the surface. Thermally-induced breezes are expected to be shallower and more coherent with a head wind [e.g., Garcia-Carreras et al., 2010] and the background winds in all runs are directed from positive to negative y, with an approximate strength of 1 m/s in the afternoon. Given that the land-surface configuration is different for all the runs, this suggests that it is the background winds that cause the shift in precipitation.

[19] For wavelengths of 10–100 km the peak power of negative surface and rainfall anomalies occurs at the same wavelengths. This means that suppression of rainfall occurs over all the forest in the presence of mesoscale variability. This is consistent with subsidence due to developing convection at the convergence zones, as its impact will extend over large areas, as shown in GC11. For example, in HET rainfall is suppressed (compared to the average rainfall in CTL) for 80 km between 140 km and −100 km. The magnitude of the anomalies also track closely, with stronger anomalies at the surface associated with stronger suppression of rainfall. For example there is stronger suppression at 70 km compared to −50 km in RN2, and at 130 km compared to 0 km or −70 km in RN3, in both cases associated with more pronounced anomalies in the land-cover wavelet power. This implies that stronger heterogeneity increases the extent of the rainfall suppression, at least relative to the enhanced rain over the crop. Overall, the magnitude and spatial extent of rainfall suppression closely matches the land-surface heterogeneity in all the simulations.

[20] Although the presence of increased rainfall over cropland is robust, the peak values of the positive wavelet power anomalies (cropland and enhanced rainfall) do not match in the same way as for the negative anomalies. There is, in general, a shift of the power in the precipitation wavelet towards smaller wavelengths relative to the land surface, particularly for the larger surface flux anomalies. This is explained by the presence of considerable small-scale variability in the precipitation patterns. For example in HET, for λ = 2–10 km, there are distinct peaks in the precipitation wavelet power at 55, 120 and 130 km. These correspond to sharp peaks in the rainfall at these locations. These peaks are also associated with land-surface features, albeit weak ones, at 50, 115 and 125 km (λ = 10 km). Similarly, in RN2 a large cropland region leads to two separate rainfall peaks at 25 and 40 km, coinciding with small changes in surface fluxes (on the order of 10 W/m2) embedded within the larger cropland region. This is consistent with Roy et al. [2003], who argue that for large-scale heterogeneities, smaller perturbations within them will break up the land-surface induced flows, leading to a preferred scale for such flows of 10–20 km.

[21] The timings of the rainfall maxima sometimes differ. For example, in the above example from RN2, the rainfall at 40 km occurs between 14:00 and 16:00, whereas at 25 km, which lies over a much weaker land-surface feature, it is between 16:00 and 18:00. This later convection is a result of triggering from cold-pool outflows which originate at 40 km. Thus, even precipitation arising from secondary initiation is consistent over a number of runs. Given that rainfall often occurs over weak surface features, it is possible that although the surface flux gradients might not be sufficient to trigger convection, the combination of the small surface flux gradient and the arrival of the cold pool leads to enhanced precipitation over these regions.

[22] The total domain-mean rainfall is affected by the surface heterogeneity. The mean rainfall for all ensembles (excluding CTL) is of 1.15 ± 0.01 mm as opposed to 1.02 ± 0.02 mm in CTL, a statistically significant increase (at the 0.01 level) of 13%. Furthermore, there is a positive correlation between the domain-mean rainfall and the mean wavelet power of the surface heterogeneities (averaged for wavelengths between 20 and 200 km, the scales at which thermally induced breezes are expected to be significant [Baldi et al., 2008]). The relationship is linear expect for the two extremes in the land-cover wavelet power (Figure 2), with a maximum increase in rainfall compared to CTL of 22% (in RN5). This suggests that there is a cutoff heterogeneity intensity above which the impact on rainfall amounts does not vary. This is to be expected, as total rainfall amounts will ultimately be limited by the total moisture fluxes in the model. It is also likely that the relationship has low sensitivity of rainfall to surface heterogeneity for low wavelet powers, where the heterogeneity may not be large enough to initiate breezes.

Figure 2.

Scatter plot of domain-mean rainfall and mean surface wavelet power averaged for wavelengths of 20–200 km for all ensembles. The error bars represent the standard error in calculating the mean rainfall.

4. Discussion

[23] The above results demonstrate that there is a significant coupling between the land-surface heterogeneity and locally-generated rainfall, with a consistent (over a 20 run ensemble) 4–6 fold increase in rainfall over cropland boundaries compared to the homogeneous case, and a strong suppression of rainfall over the forest. These results were remarkably consistent for a range of randomly generated land-surfaces of the same spectral distribution. These results are also consistent with satellite observations of rainfall [Knox et al., 2011], and with the processes described in GC11. Subsidence over the forest acts to suppress cloud initiation, thus considerably limiting rainfall amounts. The cropland boundaries, on the other hand, are regions where CAPE is maximized (GC11) leading to increased cloud depth and organization. This mechanism leads to a persistent increase in rainfall over the cropland boundaries.

[24] The magnitude and spatial coverage of the suppression in rainfall follows the size and intensity of the surface anomalies closely. In contrast the location and magnitude of peak rainfall, although always occurring over a cropland boundary, appears to be hard to predict. There are three main factors which appear to determine where peak rainfall occurs (although this list is not necessarily exhaustive):

[25] 1. Heat flux gradient, manifest in the wavelet power for heat flux. The amplitude of the local gradients in surface heat flux controls the strength of the vegetation-breeze, which in turn promotes convective triggering.

[26] 2. Size of patch related to the wavelet wavelength. Land-surface breezes tend to organise in certain preferential length-scales [Roy et al., 2003], and therefore patches close to these scales will tend to cause stronger convective triggering.

[27] 3. Distance to other patches, related to the distance between peaks in the wavelet analysis. As convection in one location may suppress convection nearby, we anticipate that proximity to other surface features will influence the statistics of convective triggering at each location.

[28] There are a number of factors which are not covered by the modelling framework used in this study, a topic for further research. The background wind will in reality vary in space on the mesoscale, and in a 3D environment can reorient, as well as advect, mesoscale rolls [Weaver and Avissar, 2001; Weaver, 2004a]. Extending this study to include the interaction between the mesoscale flows and the background winds, particularly in a realistic 3D environment, would therefore be of interest.

[29] This study focuses on locally-generated rainfall. There is evidence that large-scale organised systems, relatively insensitive to convective inhibition (CIN), rain more over humid surface anomalies [Clark et al., 2003]. The change in total rainfall at a location could therefore depend on the proportion of local versus large-scale propagating convection it receives.

[30] The final factor involves land-surface feedbacks. Cloud shading and increased precipitation over the cropland will decrease the Bowen ratio, and potentially lead to accelerated vegetation regrowth. Complex interactions with the larger scale can occur, as the surface soil moisture patches will interact with future storms, thus producing new patterns in surface heterogeneity [Nykanen et al., 2001; Weaver, 2004a, 2004b; Ramos da Silva and Avissar, 2006]. Observations show that land-surface induced breezes are climatologically significant in various regions despite these feedbacks. Their impact must, however, be better understood to correctly quantify the climatological impact of deforestation on rainfall.


[31] LGC was funded by a NERC studentship NE/F007477/1. The work was also supported by NERC grants NE/B550538/1 and NE/G018499/1. The authors would like to thank John Marsham and Chris Taylor, as well as the two anonymous reviewers, for constructive comments on the results and manuscript. Based on a French initiative, AMMA was developed by an international scientific group and funded by a large number of agencies, especially from Africa, European Community, France, UK and USA. More information on the scientific coordination and funding is available on the AMMA International web site:

[32] The Editor thanks two anonymous reviewers for their assistance evaluating this manuscript.