Regional dry-season climate changes due to three decades of Amazonian deforestation

Transcription

1 In the format provided by the authors and unedited. SUPPLEMENTARY INFORMATION DOI:./NCLIMATE Regional dry-season climate changes due to three decades of Amazonian deforestation Jaya problemkhanna by using 1 * three-decadal, David Medvigy satellite 1, observations, Stephan of Fueglistaler clouds 1, and Robert Walko scales of deforestation (see also Methods). Use of another cloud More than 0% of the Amazon rainforest has been cleared in and precipitation for trend detection, and numerical simulations to 1 Program in Atmospheric and Oceanic Sciences, Princeton University, Princeton, New Jersey 0, USA. Department of Geosciences, Princeton University, Princeton, New Jersey 0, USA. Rosenstiel School of Marine and Atmospheric Science, University of Miami, Miami 1, USA. Present addresses: Jackson School of Geosciences, University of Texas at Austin, Austin, Texas 1, USA (J.K.); Department of Biological Sciences, University of Notre Dame, Notre Dame, Indiana, USA (D.M.). * jkhanna@jsg.utexas.edu NATURE CLIMATE CHANGE ADVANCE ONLINE PUBLICATION 1 NATURE CLIMATE CHANGE Macmillan Publishers Limited, part of Springer Nature. All rights reserved.

2 SUPPLEMENTARY INFORMATION Cloud Detection Algorithm ISCCP GridSat visible (VIS, 0. µm) and infrared (IR, µm) channel data were used to infer cloud cover over Rondônia. This -hourly dataset is available from to present at a spatial resolution of km. We analyzed the data at 0 LT and 0 LT for the months of June, July, August and September between 1 and 00 (visible data being severely incomplete in other years). We used a standard cloud detection algorithm 1 to generate our cloud maps. The main test in this algorithm is to determine cloudy pixels by applying 1) a lapse rate threshold on the surface temperature and ) a threshold on the surface albedo of the corresponding pixel: IR #$%&' IR )*+%$ > IR -.'%/.0$1, VIS )*+%$ VIS #$%&' > VIS -.'%/.0$1. 1 Here, IR clear and VIS clear are the brightness temperature and albedo, respectively, of the pixel under clear conditions, IR pixel and VIS pixel are the brightness temperature and albedo of the pixel at the time of observation, and IR threshold and VIS threshold are thresholds based on the atmospheric lapse rate and observed statistical differences between clear and cloudy pixels. All the threshold values used in our implementation have been taken from Rossow and Garder The cloud detection algorithm is designed to generate the cloud cover field by making the best possible estimates of IR clear and VIS clear. In the ISCCP cloud detection algorithm 1, the surface temperature and albedo for each group of pixels is determined based on the statistical characteristics of cloud occurrence in that region, both over space and time, and then applying several spatial and temporal tests to find pixels which have a high chance of being clear in every

3 -day period. The algorithm then assigns a -day spatio-temporal averaged surface temperature and surface albedo to these pixels, which are then used, along with equation 1, to distinguish between cloudy and clear pixels. For optimal performance, IR statistics are collected over spatial regions that are sufficiently small to minimize the probability of false clear detection due to large spatial variability in land properties, but sufficiently large so that the contrast due to cloud and clear pixels is well-captured. In our algorithm this area is chosen to be km by km in size. Sensitivity tests done with somewhat different sizes produced similar results. The output of the algorithm is a binary image of cloudy pixels at each time snap (Supplementary Fig. g) The algorithm is applied on GridSat VIS and IR images of the type shown in Supplementary Fig. a-d. Supplementary Fig. e-g show a sample of the binary cloud image produced given a VIS and IR scene. The algorithm identifies cloudy and clear pixels in a scene and hence divides the VIS and IR histograms into cloudy and clear parts, the former occupying high albedo and low brightness temperature regions (Supplementary Fig. h-k). We also evaluate the algorithm through its ability to reproduce persistent regional natural features like orographic convective triggering over the hills (. W,. S and. W,.1 S) and thermal convective triggering over the natural savanna (1. W,. S) shown in Supplementary Fig. 1. These features are robustly reproduced throughout the -year analysis period as seen in Fig. 1a-c and Supplementary Fig. a-c. Our MATLAB implementation of the algorithm is provided in the data repository. 1 A second, simpler cloud detection algorithm was also used for comparison. This algorithm assumes that at a given time 0% of the pixels in a given scene can be cloudy. The scene in this

4 study is defined as the area between W to 0 W and 1 S to S. Of these top 0% brightest pixels, the ones that were more than K cooler than the monthly average surface temperature of the scene were defined to be cloudy. The monthly average surface temperature was defined to be the average of the top % warmest pixels in the whole month. This brightness temperature cutoff ensures that we do not flag non-cloudy bright pixels, like bare land, as cloudy. A binary image at each time snap is generated with the algorithm, which is then combined to generate the maps of percentage occurrence of cloudiness. Due to the design of this algorithm it will underestimate cloud cover in situations of widespread overcast; however, widespread cloud cover is uncommon in the dry season Model Evaluation We evaluated our simulations primarily against the in-situ LBA-ECO CD- Flux Tower Network Data Compilation. We utilize the eddy covariance, meteorology and radiation data collected at two eddy flux tower sites in Rondônia - Fazenda Nossa Senhora (pasture site at. W,. S) and reserve Jaru (forest site at 1. W,.0 S) (see Supplementary Fig. a,b). These sites have been a part of both the Anglo-Brazilian Amazonian Climate Observational Study (ABRACOS) and Large-Scale Biosphere-Atmosphere Experiment in Amazonia (LBA) providing valuable data for climate studies of the impacts of deforestation, model evaluation and parametrization. Our model surface parametrizations also come from the data collected from these and similar sites during ABRACOS. The ABRACOS campaign was active between and 1 and the LBA-ECO CD- data was collected between March, 1 and September, 00. As the current study is focused on analyzing climatic effects of present-day deforestation, we used the LBA data for model evaluation. Moreover, the data from

5 these campaigns cannot be used to detect signals of spatial redistribution of clouds or precipitation over the whole deforested region because such an analysis would require simultaneous measurements from more than one pasture site Owing to the fact that in-situ measurements provide data for a very limited area (and therefore are bound to be affected by local, site specific conditions), we also evaluate some model fields against satellite data. Simulations are evaluated against precipitation data from the monthly TRMM satellite data product B (resolution 0. x0. ) and surface radiation fluxes obtained from the CERES surface EBAF product (resolution 1 x1 ). Satellite data were averaged between 000 and 01 and over the deforested boundary of 00 due to their coarse resolution and because the dynamical mechanism can especially effect the spatial precipitation patterns. Hence the simulated precipitation reported in Supplementary Table is also averaged over the 00 deforested boundary. Because of this reason it is also not appropriate to compare simulated precipitation with rain gauge measurements from an earlier time period We compare data averaged over the month of August in accord with the time period simulated with OLAM. Model evaluation is performed for the numerical experiment DEF0SST00 because this simulation has land cover and SST boundary conditions closest to the LBA-ECO CD- field data. The comparison is presented in Supplementary Fig. and Supplementary Table. All variables (except precipitation) reported from the numerically simulated data are averages over a ~0. by 0. area around the LBA-ECO pasture site (. W,. S) or the LBA-ECO forest site (1. W,.0 S). The simulations are labeled respectively DEF0SST00 pas and DEF0SST00 for in Supplementary Fig. and Supplementary Table.

6 We evaluate boundary layer processes, specifically surface sensible and latent heat fluxes and boundary layer height, during the afternoon hours as the mesoscale circulations studied here are primarily afternoon phenomenon. The simulated diurnal cycle of surface latent and sensible heat fluxes are shown in Supplementary Fig.. Over the pasture these fluxes have a reasonable agreement with observations. The forest latent heat flux also agrees reasonably with observations although with a lag of ~1 hour. The forest sensible heat fluxes are in disagreement with observations during the morning hours. This disagreement can arise because in our simulations the LBA forest site, reserve Jaru, falls almost inside the deforested boundary of our simulations (Supplementary Fig. a,b). This is because the deforestation extent in our simulations is larger than that when the LBA-ECO CD- data were collected. However, the afternoon averaged fluxes (both latent and sensible) over both pasture and forest sites are in good agreement with observations (Supplementary Table ). Because the response time of the planetary boundary layer to surface processes is an hour or less, we compare model output with site observations in the afternoon only The near surface air temperature in the LBA-ECO CD- data is measured at heights (from ground) of 0 m and m over the forest and pasture canopies respectively. It should be noted that there is no analogue of this measurement in our simulations as the lowest prognosed level in our model is at 0 m altitude. Hence, in the table we have compared observed near surface temperature with the closest model analogue - canopy air temperature. Canopy air is defined by that part of the near surface air that is directly in contact with and is effected by soil, snow, surface water and vegetation. Turbulent fluxes of heat and moisture are transferred between these

7 components and the canopy air and between canopy air and the atmospheric boundary layer. OLAM is observed to have a near surface cold bias as compared to in situ observations. This is possibly because convective triggering simulated by the convective parametrization happens early in the day in the model as compared to observations, which is also a recognized discrepancy of convective parametrizations in many other GCMs. This early onset of convection reduces the incoming solar radiation due to increase in cloud cover during midday hours resulting in a cold bias in surface temperatures. But the simulated near surface temperature difference between pasture and forest vegetation is comparable to that observed We also compare the effect of simulated surface heat fluxes on the boundary layer characteristics during afternoon hours. For this purpose we calculate the simulated boundary layer height following prescriptions of previous studies. Averaged between 0 LT and 0 LT the boundary layer height for pasture is m ( m) and that over forest is m (1 m). Averaged between 0 LT and 0 LT the boundary layer height for pasture is 1 m (1 m) and that over forest is m ( m). The values represent ensemble mean and standard deviations. These values are comparable to boundary layer heights measured between 1 and August 1 at the LBA pasture and forest sites which are respectively: m () and 0 m (0 m) at 0LT; and m ( m) and m ( m) at 0 LT. It is however noted that the average difference between the simulated pasture and forest boundary layer height is smaller than the observed difference. 1

8 It is noted that the dipole in simulated relative humidity is observed at all model levels close to the boundary layer top (figures not shown) but is strongest slightly above the boundary layer top. Hence, the simulated data presented in this study is reported at ~m altitude.

9 Supplementary Figure 1. Increasing scales of deforestation in Rondônia over the last three decades. Deforested regions of Rondônia in a,1, b,, c, 1, g, 00 and h, 00. d, e, f, j, and k, are the corresponding zoomed in images over the red box. Highway BR- running from south-east to north-west of the deforested domain is represented by the black-dashed line. Land cover data is obtained from the m resolution LBA-ECO ND-01 land cover maps 1 derived from LANDSAT images. i, Topography around Rondônia. l, 0 LT JJAS ambient winds at 00 mb, mb and 0 mb averaged over - W, -0 W, -1 S and - S between 1 and 0. Wind data is from NCEP/NCAR reanalysis 1. Latitude and longitude are in degrees.

10 Supplementary Figure. Figure to evaluate the performance of the cloud detection algorithm in separating cloudy (high albedo and low brightness temperature) pixels from clear pixels of a GOES scene (see Methods for detailed discussion of evaluation). 0 LT JJAS average a, b, albedo and c, d, brightness temperature between 1 (a, c) and (b, d) respectively obtained from GridSat 1. Data is shown as anomalies from the x area average (reported in upper right in each panel). e, f, g sample of a cloud occurrence map obtained using the cloud detection algorithm 1. e, Albedo, f, brightness temperature at 0 LT on 1 st August 00 and g, corresponding binary cloud cover image. h, i, j, and k, - frequency distributions, derived from 0 LT JJAS albedo and brightness temperature, for the full scene ( W to 0 W and 1 S to S) and for cloudy and clear pixels identified on a daily basis in 1 (h, i) and 00 (j, k). 1

11 Supplementary Figure. Emergence of the dipole structure in cloud occurrence and precipitation with increasing deforestation. JJAS percentage occurrence of clouds between (a, d) 1 and, (b, e) and 1 and (c, f) 001 and 00 using GridSat 1 data. a, b, and c, are maps at 0 LT and d, e and f, are maps at 0 LT. g, JJAS daily precipitation from TRMM B, h, JJAS 0 LT precipitation from TRMM B and i, JJAS daily precipitation from PERSIANN 1 averaged between 00 and 01. Data is presented as a percentage deviation from deforested area mean (reported at the top of each panel). Stippling shows differences significant at the 1% significance level. Solid lines represent deforested boundaries in the corresponding decades (see Supplementary Fig. 1). Dashed lines represent deforested boundary in 00 and is provided as reference.

12 Supplementary Figure. Time evolution of spatial patterns of cloud occurrence and precipitation over the deforested area in June, July, August and September. 0 LT and 0 LT averaged percentage occurrence of clouds in June (a, e and i), July (b, f, and j), August (c, g and k) and September (d, h and l) in 1 to (a-d), to 1 (e-h) and 001 to 00 (i-l). Data is derived from GridSat 1 measurements. Data is presented as percentage difference from the deforested area average. 1

13 1 1 Supplementary Figure : Correlation between patterns of cloudiness, scale of deforestation and spatial location in the early and present periods (see Methods for more details). Bivariate probability distribution functions (PDF) (a, c) of 0 LT JJAS percentage occurrence of cloudiness and fraction of deforested area under the corresponding grid cell in a, 1- and c, 001 to 00. Corresponding univariate PDFs of % deviations of cloudiness from area mean (b and d). These figures show that there is a transition in the relationship between cloud occurrence and local deforestation scale between the early and present periods and that in the present period the dipole location is independent of the local scale of deforestation. e-h Bivariate PDFs between latitude (. S set as origin) and percentage deviations of cloudiness from area mean for 0 LT JJAS cloud cover maps averaged between e, 1 to 1, f, 1 to 00 and g, difference between the two maps. The data in (e, f, and g) is obtained using 000 bootstrapped samples from each period. These figures show that the difference between the cloud patterns in the early and present periods is robust. 1 1

14 Supplementary Figure. Evaluation of simulated diurnal cycle of sensible and latent heat fluxes (see Supplementary information for detailed model evaluation). a, and b, respectively show land covers in 1 and 00 used in simulations. The LBA forest and pasture sites are also displayed. Comparison of the in situ observed and simulated diurnal cycles of c, surface sensible and d, latent heat fluxes around the LBA pasture site Fazenda Nossa Senhora and forest site Reserve Jaru averaged over the month of August. Simulated data is obtained from the experiment DEF0SS00 and observed data is obtained from the in situ measurements from the LBA-ECO CD- dataset. Sensible Heat flux (W/m ) c LBA Pasture LBA Forest OLAM Pasture OLAM Forest Latent Heat flux (W/m ) d Local time of Day (Hours) Local time of Day (Hours) 1

15 1 Supplementary Figure. m altitude relative humidity averaged between 0 LT and 0 LT in a, DEFSST0-FORprSST0, b, DEF0SST00-FORprSST00 c, DEFSSTclFORprSSTcl and d, DEF0SSTcl-FORprSSTcl. e, DEFSST0, f, DEF0SST00, g, DEFSSTcl and h, DEF0SSTcl - precipitation totaled between 0 LT and 000 LT shown as a percentage difference from deforested area average. m altitude relative humidity averaged between 0 LT and 0 LT in i, DEF0SSTcl-dyn - FORprSSTcl, j, DEF0SSTclthrm - FORprSSTcl k, DEF0SSTcl-topo FORprSSTcl-topo, l, DEF0SSTcl-topo. l shows percentage difference of the field from deforested area average. All results are averaged over all days in August and over all ensemble members. Stippling shows differences significant at the 1% significance level. 1

16 Supplementary Figure. Simulated mesoscale circulations in the early and present-day time periods. Horizontal cross sections of vertical and horizontal wind averaged between 0 LT and 0LT at a, b, m, c, d, m and e, f, 1 m for a, c, e, DEFSSTcl FORprSSTcl and b, d, f, DEF0SSTcl FORprSSTcl. The data is averaged for all days in August and over all ensemble members. The horizontal wind vectors are not to scale between panels but individual scales are provided at the top of each panel. a - 0. ms ms -1 b c ms -1 e ms d ms -1 f ms Vertical Wind (m/s) (difference from baseline experiment)

17 Supplementary Table 1. Cloud and Precipitation Observations showing statistically significant increase in polarity in individual months calculated from years of data. Table listing linear trends (p value) in cloud and precipitation dipole strength in J, J, A, S and JJAS, and spatial correlations between monthly avg. and JJAS avg. cloud occurrence field (calculated within the deforested boundary). Annual Trend (p-value) % km year -1 GridSat Cloud Occurrence Spatial corr. with JJAS avg. (p-value) PERSIANN Precip. Occurrence Annual Trend (p-value) % km year JJAS (e-0) (.e-0) June 0 (0.00) 0.1 (0) 0. (0) 0.1 (0) (0.) July (0.00) 0. (0) 0. (0) 0.0 (0) 1 (0.0) August (0.00) 0.1 (0) 0.1 (0) 0. (0) (0.001) September (0.000) 0. (0) 0. (0) 0. (0) 1 (.e-0) 1

18 Supplementary Table. Numerical design. Table summarizing OLAM numerical experiments. EXPERIMENT Land Cover SST Purpose DEFSST0 1 1 average DEFSST0 versus FORprSST0 DEF0SST average FORprSST0 Forested Rondônia 1 average FORprSST00 Forested Rondônia average and DEF0SST00 versus FORprSST00 - capture the combined roles of SST variability and land cover change in the observed transition in cloud cover. DEF0SSTcl 00 0 average Capture the atmospheric response DEFSSTcl 1 0 average due just to a change of land cover FORprSSTcl Forested Rondônia 0 average between 1 and 00. DEF0SSTcl-dyn 00, pasture vegetation as high as evergreen forest 0 average DEF0SSTcl-dyn - FORprSSTcl to separate out the role of horizontal surface roughness variations DEF0SSTcl-thrm DEF0SSTcl-topo FOR0SSTcl-topo 00, all properties of pasture vegetation same as evergreen forest except vegetation height 00, average topography between W to W and 1 S to S 0 average DEF0SSTcl-thrm FORprSSTcl to separate out the role of horizontal sensible heat flux variations 0 average Separate the coupled effect of topographical variations from vegetation variations on the regional hydroclimate 1

19 Supplementary Table. Comparison of numerical experiments to field data and satellite data, (see Supplementary information for detailed model evaluation). Model evaluation is done by comparing data from in situ and satellite measurements with the numerical experiment DEF0SST00. The comparison is performed by calculating averages of simulated results over a ~0. by 0. area around the LBA-ECO pasture site (. W,. S) and over a ~0. by 0. area around the LBA-ECO forest site (1. W,.0 S). The two experiments are respectively named as DEF0SST00 pas and DEF0SST00 for. Only the simulated precipitation is averaged over the whole deforested area in 00 (see Supplementary information). The values reported from other numerical experiments are not provided for model evaluation but only to document the respective surface energy components. All values presented are averaged over the month of August. In situ data are averaged between 1 and 00. Satellite data (TRMM B and CERES) are averaged over the deforested area between 000 to 01. Sensible heat (Sens Heat), latent heat (Lat Heat) and near surface air temperature (Temp) are averaged between 0 LT and 0 LT. Net radiation (Net Rad), incoming short wave radiation (Incom SW) and precipitation (Precip) are averaged over the whole day in August. Sens Heat Wm - Lat Heat Wm - Net Rad Wm - Incom SW Wm - Temp K Precip mm/day Pasture site Forest site TRMM CERES DEF0SST00 pas DEF0SST00 for DEFSST DEFSSTcl DEF0SSTcl DEF0SSTcl-dyn DEF0SSTcl-thrm DEF0SSTcl-topo