Land use change impacts on regional climate over Kilimanjaro



[1] Glacier recession on Kilimanjaro has been linked to reduction in precipitation and cloudiness largely because of large-scale changes in tropical climate. Prior studies show that local changes in land cover can also impact orographic cloudiness, precipitation, and terrain-generated circulation patterns. This study uses the Regional Atmospheric Modeling System to simulate dry season orographic cloudiness, rainfall, and orographic flow patterns over Kilimanjaro for current deforested and reforested land cover scenarios. The simulations for current land cover show satisfactory performance compared to surface meteorology and satellite-observed cloudiness. Clouds occur less frequently in response to deforestation, with the magnitude of decrease increasing with deforestation. On the windward side, cloud liquid water path (LWP) and precipitation both show decreases at lower elevations (∼1000–2000 m) and increases at higher elevations (2000–4000 m) in response to deforestation. This pattern is caused by decreased aerodynamic resistance, leading to enhanced wind speeds and convergence at higher elevations. On the lee regions, LWP deficits found in deforested simulations coincide with regions of reduced moisture while precipitation increased slightly at lower elevations (1000–1800 m) and decreased at higher elevations (1800–4000 m). Kilimanjaro offers less obstruction to background airflow, and reduced moisture transport to the lee side is found for deforested conditions, causing reduced LWP and rainfall. However, land use change has little effect on cloudiness and rainfall at elevations in excess of 4000 m and is not expected to impact glaciers in the summit zone of Kilimanjaro during the dry season. The effect in other seasons requires further investigation.

1. Introduction

[2] Orographic effects on regional weather and climate are significant [Barry, 2008; Pielke, 2002; Whiteman, 2000] with the modification of the background flow by the terrain being the dominant effect. Direct orographic uplift and formation of orographic cloud systems often leads to enhanced precipitation on the windward slopes and, at times, rain shadow regions on the lee slopes, depending on the nature of the terrain and geographic location. Thermal circulations forced by the terrain also enhance convection over orographic features. Relatively strong variations in temperature, moisture, and wind in the vertical combined with terrain-driven precipitation inhomogeneities and orographic cloud cover create significant climatic variations and provide the environment for high levels of biological endemism. Extreme examples of species that are unique to a single geographic zone include the golden toad, an amphibian endemic to the Monteverde region of Costa Rica. Recent studies show that mountain weather and climate is disproportionately impacted by climate variability and change that is affected by both natural and anthropogenic forcing [Pounds et al., 1999; Still et al., 1999; Nair et al., 2003, 2010; Ray et al., 2006]. Since modification of mountain weather and climate has significant implications for regional water resources and ecological sustainability, there is a need to understand the impact of different anthropogenic and natural climate forcings and also their interactions.

[3] Prior research [Lawton et al., 2001; Nair et al., 2003, 2010; Ray et al., 2006; van der Molen, 2002; Bruijnzeel, 2004; van der Molen et al., 2010; Chase et al., 1999] shows that land use change in the vicinity of montane regions alters topographically generated circulation patterns as well as temperature and moisture of lowland air masses that ultimately impact orographic cloud formation. However, atmospheric response of deforestation in montane areas is very dependent on the geographical setting. In Costa Rica, modeling studies [Lawton et al., 2001; Nair et al., 2003, 2010; Ray et al., 2006] as well as radiosonde observational analysis [Nair et al., 2008] show that deforestation results in increase in cloud base height of boundary layer clouds. However, van der Molen [2002] found that coastal deforestation in an island setting in Puerto Rico resulted in enhanced slope circulations leading to invigoration of montane cloud formation.

[4] Land use may also impact the regional climate of Kilimanjaro in East Africa, where retreating glaciers on the peak of Kibo are one apparent phenomenon of large-scale climate variability and change [Thompson et al., 2002; Kaser and Osmaston, 2002; Kaser et al., 2004; Cullen et al., 2006]. The regional manifestation of the large-scale changes is characterized by drying, which might have its origin in the late 19th century and clearly appears in modern observations over the past decades [Mölg et al., 2009b]. Aside from glacier shrinkage, this drying has affected regional life forms, agriculture, and vegetation, which is also summarized by Mölg et al. [2009b]. The manner in which the regional climate of Kilimanjaro responds to land use change, however, may be significantly different compared to other areas since Kilimanjaro, as an isolated terrain feature in an inland area, occupies a very different geographical setting. Initial, idealized mesoscale numerical modeling experiments conducted by Mölg et al. [2009a] suggested that on Kilimanjaro a “flow-around-regime” may occur frequently, as opposed to a “flow-over-regime.” Thus, topographically generated convection may be as or more relevant than cloud formation because of direct topographical lifting of air. In this context, land use change has the potential to impact the regional climate of Kilimanjaro. Mölg et al. [2008a] hypothesized that for Kilimanjaro, the local anthropogenic forcing caused by land use change is potentially being superimposed on larger-scale climate forcing. However, the magnitude and exact nature of this forcing is unknown.

[5] The overall goal of the present study is to examine the impact of land use change on the regional climate of Kilimanjaro through a month-long regional atmospheric model simulation using the Regional Atmospheric Modeling System (RAMS) [Cotton et al., 2003] driven by larger-scale observed atmospheric flow features. The study area and data sets used are described in section 2; methodology and results are detailed in sections 3 and 4, respectively. Section 5 includes discussion of the results, and section 6 concludes.

2. Study Area

[6] Mount Kilimanjaro is located at 3.05°S and 37.33°E, near the northern border of Tanzania with Kenya (Figure 1a). It represents an eroded ancient volcano occupying an area of approximately 75 km × 75 km, with three peaks (Shira, Mawenzi, and Kibo) rising from 800 m in the savanna plain to over 5800 m at the peak of Kibo. Soini [2005] gave a history of land use changes in the region dating back to the early 20th century, describing as well as showing a drastic increase in land used for farming as the population on the south side of the mountain increased over the time period from 1961 to 2000. Since 1976, there have been multiple changes to the vegetation [Hemp, 2005, 2009], with a 9% decrease in montane forest, an 83% decrease in subalpine cloud forest, and an increase in bush and cushion vegetation. One hundred fifty square kilometers of Kilimanjaro's alpine forest cover has been destroyed by fire since 1976, which, along with 450 km2 of loss caused by logging, contributes to a 33% loss of total forested area since 1929 [Hemp, 2009].

Figure 1.

(a) Topography of Mount Kilimanjaro. (b) Model grid domains. Blue, grid 1; green, grid 2; orange, grid 3; black, grid 4. Also noted are sites used for analysis of temperature, dew point, and wind speed (red dots within the green grid).

[7] The north–south migration of the Intertropical Convergence Zone (ITCZ) and the associated Indian Ocean circulation are the dominant influences on climate of the study area resulting in the bimodal seasonality of precipitation over the study area [Hastenrath, 2000; Camberlin and Philippon, 2002]. The highest precipitation totals in the study area occur when the ITCZ is passing over the region and the flow near the surface is predominantly northeasterly. These conditions occur during two periods: March–May and October–November when the region experiences “long rains” and “short rains,” respectively. During the rainy seasons, the majority of moisture originates from the Indian Ocean [Chan et al., 2008]. The study region receives the majority of the annual rainfall during these two periods. For example, the average annual rainfall at Moshi, Tanzania, near Kilimanjaro, is 103 cm, of which 69 cm, on average, falls during the long rain period from April to May.

3. Methodology

[8] This study uses RAMS numerical modeling experiments to examine the hypothesis that deforestation in the lower slopes and surroundings of Mount Kilimanjaro impacts topographically induced circulation patterns and convection, altering the environmental conditions of the upper slopes and peak. The following sections provide a brief description of RAMS, the model configuration used in this study, and the simulation experiments during the dry season.

3.1. The Regional Atmospheric Modeling System

[9] RAMS [Cotton et al., 2003] is a nonhydrostatic numerical modeling system that uses finite difference formulation to solve conservation equations for atmospheric mass, momentum, heat, and water (all three phases). RAMS utilizes a polar-stereographic grid with Arakawa-C grid stagger in the horizontal and a sigma terrain-following coordinate in the vertical. Explicit microphysical parameterization [Walko et al., 2000] and convective parameterization schemes [Kain and Fritsch, 1990; Tremback, 1990] are both available within RAMS. Multiple options for turbulence and radiative transfer parameterization of varying sophistication are available within RAMS. The temporally varying lateral boundary conditions in RAMS are usually specified using analysis and forecast fields from other global models. The Land Ecosystem Atmosphere Feedback 3 (LEAF-3) submodel [Walko et al., 2000] is used in RAMS for representing soil-vegetation-atmosphere transfer (SVAT).

[10] Satellite observations of the normalized difference vegetation index (NDVI) from the advanced very high resolution radiometer (AVHRR) are generally used to specify vegetation phenology, which, in combination with vegetation type, is used to parameterize vegetation characteristics such as leaf area index (LAI), aerodynamic roughness, and other parameters within LEAF-3. However, in this study, the mean monthly moderate resolution imaging spectroradiometer-derived NDVI [Justice et al., 1998] is used to specify current day vegetation cover (see section 3.2). This will allow the vegetation to be consistent with reality over the period of study. The global, 30 s resolution land use categorization data set from Olson [1994] is used to specify vegetation type. The U.S. Geological Survey 30 s global digital elevation data set ( is generally used by RAMS to specify terrain, but when appropriate, the Shuttle Radar Topography Mission (SRTM) 90 m resolution topography data set [Jarvis et al., 2004] is also utilized in this study (see section 3.2). The global 1° United Nations Food and Agricultural Organization [1971] data set (FAO 1971–1981) [Gerakis and Baer, 1999] is used to specify the soil type in RAMS.

3.2. Numerical Model Experiments

[11] We use three different numerical modeling experiments, assuming land use scenarios in which (1) the current vegetation coverage persists, (2) the area is completely deforested, and (3) the area is completely reforested. The RAMS simulation assuming current vegetation coverage (CTL) is used as the point of comparison for the other vegetation scenarios, as well as to evaluate the performance of RAMS over the study time by comparing it against surface observations and satellite observations of cloudiness. The model experiment is then repeated for deforested (DEF) and reforested (REF) land use scenarios.

[12] All of the experiments utilized in this study use a hierarchy of four nested grids (Table 1), with the outermost grid of 64 km grid spacing covering a domain that includes a substantial portion of East Africa and the Indian Ocean (Figure 1b). Three interior nested grids of progressively decreasing grid spacing are used to establish a central domain with grid spacing of 1 km that covers Kilimanjaro (Figure 1a). Details of the numerical model configuration used in the experiments are given in Table 1.

Table 1. RAMS Model Configuration for All of the Experiments
ConfigurationGrid 1Grid 2Grid 3Grid 4
NX × NY56 × 5662 × 6266 × 6674 × 74
ΔXY (km)641641
ΔZ (m)/stretching ratio60/1.0860/1.0860/1.0860/1.08
ΔT (s)9022.55.6251.41
Center latitude/longitude3.05°S, 37.33°E3.05°S, 37.33°E3.05°S, 37.33°E3.05°S, 37.33°E
Radiation parameterizationHarringtonHarringtonHarringtonHarrington
Lateral boundary conditionsKlemp/WilhelmsonKlemp/WilhelmsonKlemp/WilhelmsonKlemp/Wilhelmson
Convective parameterizationKain-FritschKain-FritschExplicitExplicit
Eddy diffusionMellor-YamadaMellor-YamadaMellor-YamadaMellor-Yamada
Soil levels4444
Vegetation patches1111

[13] This study focuses on the dry season month of July, with the atmospheric conditions for 1–31 July 2007 being utilized to initialize RAMS. In the dry season, there will be less synoptic cloud cover and precipitation, so the precipitation and cloud cover that does occur will be more influenced by orographic processes and local terrain. In this context, regional climate, especially orographic cloud formation, is expected to have higher sensitivity to land use change and associated alteration of surface energy fluxes during the dry season. Note that analysis conducted in this study is specific to flow regime found at Kilimanjaro during the dry season, characterized by the Froude number. The results from this study are not directly applicable to terrain flow in other geographic regions where the Froude number may differ. However, at Kilimanjaro, Mölg et al. [2009a] did not find substantial seasonal differences in Froude number.

[14] In all the experiments utilized in this study, RAMS is integrated for 24 h each day of July starting from initial atmospheric conditions specified using Global Forecast System (GFS-ANL) three-dimensional 1° × 1° grid analysis of wind, temperature, and moisture fields at 0 UTC. The lateral boundaries and the model top are constrained by nudging it toward conditions consistent with the temporally varying GFS analysis fields available every 6 h. This experimental design, which is consistent with the type I dynamic downscaling classification of Castro et al. [2005], is chosen to retain realism of the large-scale atmospheric conditions around the study area. However, the soil moisture and temperature conditions in the RAMS experiments are initialized from GFS analysis only once, at 0 UTC of 1 July 2007. Thus, unlike the atmospheric conditions that are reinitialized every 24 h, soil moisture and temperature is allowed to continuously evolve in all the experiments during the entirety of the study period. This follows a type II dynamical downscaling approach to retain the evolution of the finer-scale features of the land boundary conditions that have a longer memory compared to atmospheric conditions.

[15] Land cover in the CTL experiment (Figure 2a and Table 2) is specified using the Olson Global Ecosystem (OGE) data set, the default land cover characterization data set used by RAMS [Loveland et al., 2000]. Note that the RAMS LEAF-3 SVAT submodel reduces the 96 land cover classes in the OGE data set to 30 land cover classes. In the DEF experiment, all evergreen and deciduous forest as well as wooded grassland areas are converted to short grass (Figure 2b and Table 2) extending the extremes of deforestation found along the western and northern slopes shown in the land cover used in the CTL experiment. While the arbitrary conversion of 53.5% of the total land cover from wooded grassland is not very realistic, it provides an upper bound for the effects of land use change. In the REF experiment, all the current areas that are farmland or urbanized are replaced by evergreen forests (Figure 2c) to match the land cover found at the middle elevations in the CTL experiment.

Figure 2.

Land use change scenarios for the three model experiments for model grid 4. Labels denote mainland use classes. (a) Control land use classification as given by LEAF-3. The high-elevation areas are denoted by semidesert (brown) and desert (dark brown) in the center of the domain. (b) Deforested land use scenario. (c) Reforested land use scenario. Kelly green denotes evergreen broadleaf classification, whereas dark green is the deciduous broadleaf classification.

Table 2. The Area Covered by the Various Land Use Classes in the Model Domain for the Three Differing Simulationsa
LEAF-3 ClassificationCTL AreaCTL %REF AreaREF %DEF AreaDEF %
  • a

    The area is in square kilometers and percentage of the total grid area. The REF case has no cropland, whereas the DEF case has no forested areas or any standing trees (given by wooded grassland).

Deciduous broadleaf1843.4%1843.4%00%
Evergreen broadleaf5449.9%143626.2%00%
Short grass5089.2%5089.2%431678.8%
Wooded grassland293253.5%284451.9%00%

[16] As discussed briefly in section 3.1, in the CTL scenario, the spatial distribution of NDVI is specified using a more representative Moderate Resolution Imaging Spectroradiometer (MODIS)-Terra-derived NDVI data set from July 2007 rather than the default AVHRR-derived NDVI data set used by RAMS. The domain-averaged AVHRR- and MODIS-Terra-derived NDVI over grid 4 are 0.522–0.553, respectively. Since the OGE land use characterization derived from AVHRR satellite observations in 1992, it does not account for additional land use change that has occurred in the recent years. Utilization of the MODIS-derived NDVI field will partially account for these additional changes in land use since the fractional vegetation cover and leaf area indices and vegetation albedo are parameterized to be dependent on the annual cycle of NDVI. In the DEF and REF scenarios, average values of NDVI of the altered land use classifications were used. For example, an average NDVI for the evergreen forested class, determined from the MODIS NDVI data used in the CTL experiment, is used to specify the NDVI for evergreen forest that replaces urban or farmland regions in the REF experiment.

4. Results

[17] RAMS experiments for the land use scenarios described in section 3.2 are used to examine the impact of land use on surface energy budgets, circulation patterns, and cloud formation for the month of July 2007. Note that except for surface meteorology and rainfall comparisons (see sections 4.1.1 and 4.1.2), analysis discussed in the study are for grid 4.

4.1. Evaluation of the RAMS Simulations

[18] The RAMS simulation for current land use conditions is evaluated by (1) comparing the model-simulated surface meteorology and observations at meteorological stations, (2) comparing the model-simulated precipitation to surface observations and satellite observations, and (3) comparing the model-simulated cloud fields against satellite observations.

4.1.1. Surface Temperature, Dew Point, and Wind Comparisons

[19] This study utilizes a statistical approach similar to that of Zhong and Fast [2003] for evaluation of RAMS-simulated 2 m temperature, dew point, and winds against corresponding observations (Figures 3a3c). Since East Africa is a relatively observation-poor area [Christy et al., 2009], only a few stations with observational fidelity sufficient for statistical analysis exist within the domain covered by the model grids. Therefore, the statistical approach compares hourly surface meteorological observations (NCDC DS-3505) from grid G2 (16 km grid spacing) (Figure 1b) to maximize the number of observations available for the analysis. The statistical analysis (Figures 3a and 3b) shows the magnitudes of root-mean-square error (RMSE), and bias for surface temperature and dew point comparisons shows a pattern in which a decreasing trend is observed, with the lowest values of RMSE and bias occurring toward the end of the simulation. For example, the daytime temperature bias, which was often in excess of −1°C during the early part of the experiment, approaches zero toward the end of the simulation (Figures 3a and 3b). Table 3 shows the mean RMSE, bias, and error standard deviation for the entire month for 2 m temperature, surface dew point, and surface wind speed. The comparisons indicate that the computed values of RMSE, bias, and error standard deviation all fall within ranges reported by modeling studies over other regions [Zhong and Fast, 2003; Miao et al., 2008; Mölders, 2008; Ray et al., 2010], demonstrating adequate performance of RAMS over the study area.

Figure 3.

Root-mean-square errors for (a) 2 m temperature, (b) dew point, and (c) wind speed following the methods of Zhong and Fast [2003]. The 24 sites used in this analysis are shown in Table 2, and average values are given in Table 4.

Figure 3.


Figure 3.


Table 3. Monthly Mean Values of RMSE, Bias, and Error Standard Deviation for Temperature, Dew Point, and Wind Speed
 2 m TemperatureDew PointWind Speed
RMSE1.99 K0.92 K1.52 m/s
Bias−1.49 K−0.18 K0.27 m/s
Error SD2.51 K1.34 K2.17 m/s

[20] This trend observed in the simulation is potentially related to adjustment of the coarse grid spacing soil moisture field from the GFS-ANL analysis that is used to initialize the RAMS simulations. Comparison to observations over other geographical regions shows that the GFS-ANL soil moisture fields have a wet bias and the improvements in the daytime RAMS-simulated surface temperature is related to the soil moisture field adjusting to the atmospheric conditions at scales finer than those resolved in the National Centers for Environmental Prediction global model used to generate the GFS analysis. The substantial biases in nocturnal temperatures found in the numerical simulation are due to the enhanced sensitivity of nocturnal boundary layer to local differences in surface characteristics between the land use scenarios [Nair et al., 2011], which is not reflected in the coarse grid spacing experiment.

[21] In addition to the statistical approach, point comparisons are conducted against observations from Kilimanjaro International Airport and Nairobi for which simulated values from the grid G3 (4 km grid spacing) are used. There is good agreement between the RAMS-simulated 2 m temperature and surface observations from Kilimanjaro International Airport, Tanzania, and the Nairobi, Kenya, Air Force base (Figures 4a and 4b) for the current land use condition with correlations of 0.76 and 0.84, respectively. The agreement between the RAMS-simulated dew point temperature and observations is substantially less, with the correlations of 0.56 and 0.41 found for the Kilimanjaro Airport and Nairobi Air Force base stations, respectively (Figures 4c and 4d). However, this decrease in correlation coefficient may be due more to the lack of diurnal signal in the model compared to the observations, as the average root-mean-square error is a lot less.

Figure 4.

Month-long comparisons of (a and b) temperature and (c and d) dew point for Nairobi, Kenya, and Kilimanjaro International Airport between the model (dashed line) and observations (solid line). Given is the correlation coefficient for the specific site.

4.1.2. Surface Rainfall Comparison

[22] The number of stations in the study area that report rainfall observations (Table 4 and Figure 5c) are substantially less (14) compared to those that report temperature, dew point, and wind speed observations (24). Therefore, numerical model-simulated precipitation (Figure 5a) is also compared to the Tropical Rainfall Measuring Mission (TRMM) [Kummerow et al., 1998] measurements over the study area (Figure 5b). However, note that while the 0.25° × 0.25° gridded TRMM 3B43 monthly total precipitation product (Figure 5b) does not adequately resolve the fine-scale patterns of orographic precipitation over Kilimanjaro, it provides information on larger-scale spatial patterns and complements the coarser scale surface observations.

Figure 5.

(a) RAMS G2 (16 km spacing) total monthly precipitation. (b) Total monthly precipitation from the TRMM sensor (0.25° resolution) over Kenya and Tanzania for July 2007. (c) Monthly total precipitation at 14 observation sites within the G2 domain. Numbers listed are total monthly precipitation (mm), and the color of the dots is on the same scale as Figures 5a and 5b.

Table 4. The 24 Stations Used in the Analysis of RMSE, Bias, and Error Standard Deviationa
WMO IDStation NameCountryLatitudeLongitude
  • a

    Shown are World Meteorological Organization (WMO) identification number, station name, country, and latitude/longitude location. These 24 sites are indicated as red dots in Figure 1b.

  • b

    The site was also used for precipitation verification in Figure 5c.

637910Kilimanjaro AirportbTanzania−3.4137.06
638450Pemba/Karume AirportTanzania−5.2539.81
638940Dar Es Salaam AirportbTanzania−6.8639.20
Table 5. Model Accuracy, Overprediction, and Underprediction of Cloud Occurrence at MODIS-Terra and MODIS-Aqua Overpasses for the 1 km Spacing Domain

[23] Comparison of the RAMS-simulated accumulated precipitation to TRMM and surface observations (Figure 5) in the second grid (G2) show that the spatial extent of local maximum on precipitation in the vicinity of Lake Tanganyika is substantially less in the numerical simulation compared to satellite estimates and ground observations. Similar trends were also found in the numerical modeling study by Ge et al. [2007] over the same region using RAMS. However, in the vicinity of Kilimanjaro, numerical model simulation overestimates precipitation compared to surface observations and also satellite estimates. At Kilimanjaro International Airport, surface observations show 4 mm of accumulation for the study period, while the numerical model-simulated value is ∼8 mm. Satellite-derived estimates show a maximum of <50 mm at Kilimanjaro, while numerical model-simulated values are in excess of 100 mm. Note that because of the two-way interactive feature of nested grids, the vertical velocity field in the vicinity of Kilimanjaro in grid G2 is enhanced compared to the rest of the grid, resulting in the tendency of overestimating rainfall. While the two-way interactive nested grid leads to better representation of orographic effects on coarser grid G2, the grid spacing is still not adequate to properly resolve the spatial variability of such effects.

[24] Comparison against ground-based observations and TRMM-derived estimates both show that over the entire grid (G2), RAMS underestimates precipitation. Compared to TRMM-derived estimates, model simulation (averaged to 0.25° × 0.25° TRMM grid) exhibits a RMSE of ∼23.1 mm and a bias of −17.9 mm. The RMSE and bias for model-simulated rainfall compared to ground-based rainfall observations are 37.5 and −34.3 mm. Intercomparison of TRMM rainfall estimates and ground observations show an RMSE of 18.11 mm and a bias of −5.0 mm. Ground-based observations or satellite-derived rainfall estimates with adequate spatial resolution for evaluating the simulated orographic precipitation patterns in the innermost grid (G4) of 1 km grid spacing are not readily available. However, the simulated altitudinal structure of precipitation compares favorably with those reported in the literature and will be discussed further in section 4.6.

4.1.3. Cloud Frequency Comparison

[25] As a result of the low density of surface observations, this study also uses a spatially distributed strategy of comparing the simulated cloud fields in grid 4 against MODIS satellite observations of cloud cover. The 1 km spatial resolution cloud mask derived from daily MODIS observations [Ackerman et al., 1998], made approximately at 1100 local standard time (LST) (Terra platform) and 1400 LST (Aqua platform) and reprojected to the RAMS polar stereographic grid (G4), was used for this purpose. To compare the satellite-derived cloud mask to the numerical model-simulated cloud field, the following metrics were computed:

equation image

where n1, n2, and n3, respectively, are the number of numerical model grid points for which clouds occur (1) in both the simulation and satellite observation, (2) in the model simulation but not in the satellite observation, and (3) in the satellite observation but not in the model simulation. Thus, the metrics A, O, and U (Table 5) quantify the percentage of comparisons where the model-predicted cloudiness is in agreement (accuracy) with the observations and is false positive (overprediction) and false negative (underprediction). Note that the following considerations should be made when comparing model-simulated cloud fields to MODIS observations: (1) the comparisons are made not exactly at the time of satellite overpass but at the hour closest to overpass time for which model output is available, and (2) the MODIS cloud masking algorithm may not adequately identify optically thin clouds, while the model-simulated cloud fields contain the entire spectrum of cloud optical thickness.

[26] Spatial distribution of frequency of occurrence of clouds (FOC) computed from all the available MODIS overpasses for the study period shows two arcs of orographic cloudiness, one on the windward side (south side) and the other on the lee side (north side) of Kilimanjaro (Figure 6a). The FOC sharply decreases toward the peak of the mountain, with the highest values around 2500 m in elevation and tailing off to near-zero values at high elevations. The spatial distribution of percentage accuracy (A) shows 70%–80% accuracy over the core regions of arc of cloudiness on the windward side (Figure 6b), comparable to the study by Ray et al. [2006]. Similar percentage accuracy is found for the leeward cloud bank, except along the western and eastern flanks, where lower values ranging from 40% to 50% are found. The percentage accuracy in the upwind slopes shows a decrease near the upper boundary of the orographic cloud bank, but above this region, the accuracy increases to 90%–95%. The spatial distribution of percentage false-positive comparisons (O) shows higher values along the upper boundary region of the windward orographic cloud bank (Figure 6c), indicating that the model overpredicts the occurrence of clouds in this area. Thus, compared to observations, RAMS predicts the upper bounds of the windward orographic cloud banks to be located at slightly higher elevations compared to observations. The spatial distribution of percentage false-negative comparisons show that the smaller accuracy values along the eastern and western flanks of the leeward cloud bank is due to underprediction of clouds in this area (Figure 6d).

Figure 6.

Spatial plots of accuracy of prediction of FOC from collocated MODIS and RAMS data for both MODIS-Aqua and MODIS-Terra time steps. (a) Average FOC (%) for all observations. (b) Model accuracy (%; see expression (1)). (c) Model overprediction (%). (d) Model underprediction (%). Domain average statistics are found in Table 5.

4.2. Response of Cloud Formation to Land Use Change

[27] Prior investigations of land use change impact on montane environments [Lawton et al., 2001; Ray et al., 2006; Nair et al., 2010] show changes to orographic cloudiness. To examine changes to orographic cloudiness, spatially averaged FOC as a function of altitude is utilized. FOC initially increases with elevation, reaching a maximum at altitudes of ∼2200 m (Figure 7a), and then steadily decreases above this altitude. Comparison between the CTL and DEF simulations (Figure 7b) shows that the impact of deforestation is to decrease spatially averaged FOC at all elevations. The impact of reforestation is the opposite, an increase in spatially averaged FOC at all elevations (Figure 7c). The maximum decreases in FOC are found between the DEF and REF simulations (Figure 7d). Note that the differences in orographic cloudiness between the three experiments are statistically significant (Student's t test, 95% confidence) at most elevations (Figure 7).

Figure 7.

Cloud frequency differences binned to 100 m elevation increments for the 1 km spacing model grid. Dots and triangles indicate mean values, and bars indicate ±1 standard deviation. Triangles indicate the differences between the simulations are at >95% statistical significance via t-means tests.

[28] Variations in the integrated liquid water path (LWP) among the CTL, DEF, and REF simulations (Figure 8) are more substantial compared to differences in FOC. Average LWP shows a band of convection to the south of the mountain, which is consistent with the average wind patterns for this season, and a smaller leeward area of higher liquid water path on the north side of the mountain (Figure 9a). In general, deforestation results in a decrease in LWP in the lower half of the windward cloud bank while causing an increase in the upper half (Figures 9b9d). The leeward cloud bank experiences a decrease in LWP in response to deforestation. Note that the magnitude of changes in LWP increases with the severity of deforestation. The magnitude of differences exceeding 40, 30, and 10 g−2 are found between DEF and REF (Figure 9d), DEF and CTL (Figure 9c), and REF and CTL (Figure 9c) simulations, respectively. Thus, while the general decrease in frequency of occurrence in orographic cloudiness in response to deforestation is accompanied by reduction in LWP at lower elevations, at higher elevations on the windward side, it is associated with increased LWP. Averaged over the entire grid, there is a decrease in LWP of 4.13 g−2 in the DEF simulation compared to CTL and an increase in LWP of 2.6 g−2 in the REF simulation compared to the CTL simulation.

Figure 8.

Comparisons of cloud liquid water path for the model simulations. Figure 8a uses the top color bar, and all other plots use the bottom bar. (a) Average cloud liquid water path for July 2007. (b) Differences between the DEF and CTL simulations. (c) Differences between the REF and CTL simulations. (d) Differences between the DEF and REF simulations.

Figure 9.

(a) Average of layer mean (0–1 km) horizontal wind speed and wind vector for the CTL simulation. (b) Average difference of layer mean (0–1 km) horizontal wind speed between the DEF and CTL simulations. R1 and R2 indicate areas of increased flow toward the mountain, and R3 and R4 indicate areas of increased flow around the mountain. (c) Same as Figure 9b except for the REF and CTL simulations. R5–R7 indicate areas of maximum change. (d) Same as Figure 9b except for the DEF and REF simulations.

4.3. Horizontal and Vertical Wind Field Differences

[29] The CTL, DEF, and REF simulations also show differences in wind patterns, averaged over the lowest 1 km of the atmosphere (Figure 9). The average pattern of low-level flow (Figure 10a) is a predominantly southerly flow forced to diverge around Kilimanjaro, accelerating as it flows along the eastern and western slopes. The average wind speeds are highest in the DEF simulation, followed by the CTL and REF simulations. Increased average wind speeds in DEF and CTL (Figures 9b and 9c) simulations are primarily caused by smaller roughness length associated with deforestation. In the DEF simulation, a decrease in surface roughness, and thus frictional effects, leads to an increase in wind speeds along the windward and leeward slopes (Figure 9b, regions R1 and R2) approximately directed along the topographical gradient and also along the western and eastern slopes of Kilimanjaro (Figure 9b, regions R3 and R4) with a significant component of the difference vector also directed along the topographical gradient. Note that the areas of significant wind differences in the DEF simulation (Figure 9b) correlate well with areas of deforestation (Figures 2a and 2b). In the REF simulation, the most substantial decreases in wind speeds and direction (Figure 10c, regions R5–R7) occur along the regions of reforestation in the windward side of Kilimanjaro (Figure 2c), as well as at the interface of the cropland and forested areas in the CTL land use scenario (Figure 2a). The pattern of wind speed differences between the DEF and REF simulations (Figure 9d) is very similar to that between the CTL and DEF simulations, but with the magnitude of differences being much higher in the former.

Figure 10.

Analysis of vertical velocity overlaid with surface horizontal wind vectors. Figure 10a uses the top color bar, and all other plots use the bottom bar. (a) Average CTL surface vertical velocity with horizontal wind vectors. S1 and S2 indicate areas of subsidence. (b) Vertical velocity differences (contoured) and horizontal wind vector differences between the DEF and CTL simulations. (c) Differences between REF and CTL simulations. (d) Differences between DEF and REF simulations. Note the unbalanced color bars in Figures 10b–10d.

[30] The average vertical velocity field for the CTL experiment (Figure 10a) shows regions of positive vertical velocity on the windward and leeward slopes, coinciding with location of orographic cloud banks in these regions. The impact of deforestation is an altitudinal shift in regions of convergence, with the increase in wind speed in the direction of the topographical gradient within regions R1–R4 (Figure 9b) resulting in enhanced vertical velocities in these regions (Figure 10b). A similar trend is found when the differences in average vertical velocities between the REF and CTL simulations are examined (Figure 10c). The differences in average vertical velocities between the DEF and REF simulations (Figure 10d) show a pattern very similar to the differences found between the DEF and CTL simulations, except that the regions of enhanced vertical velocities are enlarged (compare to Figure 10b). Deforestation also enhances subsidence, in regions where it is already pronounced (Figure 10a, S1 and S2) in the CTL simulation, along the eastern and western slopes just south of regions R3 and R4. The difference between the REF and CTL simulations shows that subsidence in the S1 and S2 regions are further enhanced compared to differences between the DEF and CTL simulations.

[31] The localized increase in LWP and precipitation at higher elevations and decreases at lower elevations following simulated deforestation (Figure 7) are consistent with the above discussed patterns of differences in horizontal and vertical wind speeds between the DEF and CTL, REF and CTL, and DEF and REF simulations.

4.4. Impact of Deforestation on Flow Modification by Terrain

[32] To quantify the modification of orographic flow patterns caused by land cover changes, the flow diversion (FD) ratio was computed following the approach of Rienecke and Durran [2008]. The flow diversion ratio is computed based on a control volume that is parallel to the direction of the large-scale flow. Of the two faces perpendicular to the direction of the large-scale flow, one is located upwind of the mountain where the terrain gradient is minimal. The other face is formed by the cross section of the mountain along the central ridge line. For the idealized mountain geometry (triangular prism) considered by Rienecke and Durran [2008], such a control volume is a cuboid. Assuming that the large-scale flow direction is southerly, the cuboid control volume is oriented south to north, with the northern face being defined by the middle cross section of the triangular-prism-shaped terrain with west-east orientation. Of the faces parallel to the flow direction, the bottom face is coplanar to the surface of zero elevation (mean sea level). The flow obstruction caused by terrain is quantified as expression (2):

equation image

where FD is the flow diversion ratio given by the ratio of the difference in mass flux in through the walls on the west side of the cuboid (ϕW), east side of the cuboid (ϕE), and south side of the cuboid (ϕS), respectively (expression (2)). In the situation of no flow obstruction, the incoming flow from the south is deflected in the vertical direction, and thus the mass flux along the east and west faces will be zero, yielding an FD value of zero. However, if the terrain offers complete obstruction to the flow, there will be no mass flux through the top face, and the net mass flux in the north–south direction will be same as that in the east–west direction, resulting in the FD ratio to be unity.

[33] Computation of the FD ratio for Kilimanjaro, whose terrain is more complex (Figure 11b) compared to the idealized topography assumed by Rienecke and Durran [2008], requires the use an irregular control volume instead of a cuboid (Figure 11c). The surface of the irregular control volume is a southward extrusion of the east–west cross section of Kilimanjaro through the highest peak. The net mass flux in the east–west direction is computed using the vertical surfaces that are normal and the velocity component parallel to this direction. Similarly, the net mass flux in the north–south direction is computed by considering the control volume surfaces normal to and the velocity component parallel to this direction. The equation used for this computation is given in expression (3):

equation image

where FD is the flow diversion ratio given by the ratio of net mass flux in the east–west (ϕE-W) and north–south (ϕN-S) directions, respectively.

Figure 11.

(a) Plot of longitudinal mass flux for all three simulations with flow diversion numbers. (b) Surface plot of model elevation for model grid 4, used in the flow diversion calculations. (c) Volume used to compute flow diversion. The incoming latitudinal mass flux for the blue wall to the south of the mountain was used as the denominator, whereas integration of the mass flux through the red faces was the numerator in the flow divergence calculations.

[34] The average control volume mass flux in the east–west direction, as a function of the longitude, show differences between the three simulations, especially in the western half of the control volume (Figure 11a). The magnitude of mass flux on the western half of the mountain increases as deforested land cover increases. The average FD ratio for the CTL, DEF, and REF simulations are 0.898, 0.884, and 0.908, respectively. These values show that, for the flow conditions characteristic of the study period, Kilimanjaro does offer substantial obstruction to the large-scale flow, with values indicating a mountaintop inversion [Reinecke and Durran, 2008]. However, the land cover change does indeed impact the FD ratio, with the FD ratio increasing by 2.4% as forest cover increases. Thus, reforestation results in an increase in the ability of the mountain to obstruct the flow, and deforestation leads to a greater proportion of flow upslope and over the mountain. Although the flow rate around the mountain increases with deforestation, so does the flow up the windward slopes. The result is an increased vertical velocity in the deforested setting, and therefore a greater proportion of the air rising over the mountain, versus diverging around it, relative to the CTL and REF simulations. The regions of enhanced vertical velocity are correlated to regions of increased horizontal winds (Figure 10).

4.5. Surface Temperature, Moisture, and Latent Heat

[35] Although consistent differences in model-simulated cloud, precipitation, and wind patterns are found in response to deforestation, these changes make it difficult to discern the direct impact of deforestation on surface temperature (data not shown). The DEF simulation shows that the surface air is generally warmer (data not shown) and drier (Figure 12b) over deforested regions, while REF simulations show cooler (data not shown) and moister air (Figure 12c) over forested regions in the windward side. Drier surface air is found on the leeward side of Kilimanjaro when comparing the DEF simulation to the CTL and REF simulations (Figures 12b12d). Deforestation results in a decrease in the water vapor mixing ratio, on average, but increases in near-surface wind speed result in very small positive changes to the mixing ratio in localized areas, including the peak. The impact of reforestation is to enhance surface moisture at lower elevations along the windward slopes, which coupled with changes to flow patterns increases moisture also on the lee side of Kilimanjaro (Figures 12c and 12d). This is evident in the patterns of monthly averaged water vapor flux along the east–west cross section through the peak of Kilimanjaro (center of grid domain 4), which shows enhanced northerly water vapor flux in both the CTL and REF simulations compared to the DEF simulation (Figure 13) above the surface layer. These produce decreases in the water vapor mixing ratio on the lee slopes and the areas on the side of the mountain that roughly coincide with regions of enhanced wind speed. Changes in the land cover do not show significant changes in average moisture or temperature at high-elevation areas downwind of the land use changes, instead occurring on the leeward side, which is consistent with the lack of changes at high elevations.

Figure 12.

Comparison of the monthly mean surface mixing ratio for the three differing simulations. Figure 12a uses the top color bar, and all other plots use the bottom bar. (a) Average CTL liquid water mixing ratio for July 2007. (b) Difference between DEF and CTL mixing ratios. (c) Difference between REF and CTL mixing ratios. (d) Difference between DEF and REF mixing ratios. Note the unbalanced color bar in Figures 12b–12d.

Figure 13.

Cross section of latitudinal moisture flux for the three simulations and differences at the highest elevation. (a) CTL latitudinal moisture flux. (b) Differences between DEF and CTL simulations. (c) Differences between REF and CTL simulations. (d) Differences between DEF and REF simulations.

[36] Only small differences in domain-averaged latent heat fluxes were found between the three experiments, with grid 4 average values of 36.6, 34.7, and 35.4 W m−2, respectively for CTL, DEF, and REF experiments. The domain-averaged latent heat fluxes are least in the DEF experiment because of reduced evaporation and transpiration over deforested regions. In the REF experiment, latent heat fluxes are indeed higher over reforested regions (data not shown), but the domain-averaged value is lower compared to the CTL experiment. Higher domain-averaged latent heat fluxes in the CTL experiment could be potentially related to enhanced rainfall on the windward slopes. Relatively small differences in domain-averaged latent heat fluxes also suggest that changes in flow diversion (see section 4.4) are the main contributor to differences in orographic cloudiness found between the three simulations.

4.6. Precipitation Differences

[37] Averaged monthly accumulated precipitation, as a function of surface elevation (100 m bins), increases up to an altitude of ∼2400 m and decreases with altitudes up to ∼4000 m (Figure 14a). This pattern is consistent with elevational profiles of precipitation reported for Kilimanjaro [Røhr and Killingtveit, 2003; Barry, 2008]. Spatial average precipitation differences between DEF and CTL simulations on the windward areas (the southern half of the fourth grid) are negative in the elevation ranges of 1000–2200 m (Figure 15a) and are positive between 2200 and 4000 m. However, the pattern for leeward areas (the northern half of the fourth grid) shows negative differences at all elevations except between 1300 and 1800 m, where a small positive difference is found (Figure 15b). The windward differences between the REF and CTL simulations are slightly negative below elevations of 1300 m and positive between 1300 and 4000 m (Figure 15c). Other than slightly positive differences between 1400 and 1800 m, very little leeward differences are found between REF and CTL simulations (Figure 15d). The differences between DEF and REF simulations are negative at an elevation below 1800 m and positive between 1800 and 4000 m on the windward areas (Figure 15e). On leeward areas, the differences are positive below 1700 m while it is negative between 1700 and 4000 m (Figure 15f). Note that the differences in orographic precipitation between the three experiments are statistically significant (Student's t test, 95% confidence) at most elevations (Figure 15).

Figure 14.

Average accumulated precipitation for July 2007 as a function of elevation. Averages are for 100 m elevation bins. The bars indicate a 1 standard deviation difference from the mean.

Figure 15.

Comparison of July 2007 total surface precipitation with respect to elevation, binned at 100 m intervals for (a, c, e) windward and (b, d, f) leeward sides. The bars indicate a 1 standard deviation difference from the mean values. Statistical significance >95% is denoted by triangles.

[38] The simulations thus show that the impact of deforestation is to cause a decrease in precipitation at lower elevations and an increase at higher elevations on the windward areas. On the lee region, there are precipitation decreases at higher elevations while there are some increases at lower elevations. The spatial pattern of precipitation differences (data not shown) shows that areas of increased precipitation coincide with regions of enhanced LWP (Figure 8), which in turn coincides with patterns of vertical velocity enhancements (Figure 10). Areas of reduced precipitation on the lee regions also coincide with patterns of moisture deficits (Figure 12).

5. Discussion

[39] The total deforestation of Kilimanjaro as given in this experiment is the upper bound of possible land use change. While such deforestation is a hypothetical end-member scenario, the reforested simulation, where anthropogenic land cover is replaced by tropical montane forests, is reflective of past land use scenarios documented by Hemp [2009] and Soini [2005]. Note that the differences between the reforested and current land use scenarios do show a consistent signal that is amplified when the extent of deforestation increases.

[40] While land cover change does indeed impact cloudiness at Kilimanjaro, physical mechanisms through which the changes are effected are different from those found in prior studies [Lawton et al., 2001; Nair et al., 2003; Ray et al., 2006; Bruijnzeel, 2004]. At Monteverde, Costa Rica, lowland deforestation impacted orographic cloud formation by causing the air mass to be warmer and drier, elevating the lifting condensation level and the height of formation of the orographic cloud bank. In an island setting in Puerto Rico, both changes in air mass and sea-breeze circulation patterns caused by deforestation was found to alter orographic cloudiness [Bruijnzeel, 2004]. In the present study for an inland site, while the changes in moisture at the surface potentially contribute to decreased cloudiness in deforested scenarios, altered flow patterns caused by reduced aerodynamic roughness and associated increase in the strength of thermally driven winds appear to play a dominant role. The pattern of increased LWP in the upper half of the windward cloud bank (Figure 8) correlates well with regions of enhanced vertical velocity (Figure 10). The increase in vertical velocity, in turn, is higher with horizontal wind speeds in the direction of the topographical gradient. Note that the vertical velocity component (ws) relative to slopes is related to topographical gradient (▿h) and horizontal wind (equation images) at the surface in the following manner [Barry, 2008]:

equation image

During the time period considered in the study, flow characteristics are such that obstruction caused by Kilimanjaro to the background flow decreases with the amount of deforestation, increasing near-surface wind speeds along the windward slopes at higher elevations where the topographical gradient is also higher.

[41] The above discussed differences in flow patterns lead to an increase in cloud thickness and precipitation at higher elevations in response to deforestation. The decrease in cloud cover and precipitation, attributed to large-scale changes in tropical climate and associated increase in absorbed shortwave radiation, is one of the mechanisms that has been shown to significantly impact glacier mass loss on the peak of Kilimanjaro [Mölg et al., 2003, 2006, 2008b, 2009b; Mölg and Hardy, 2004; Hastenrath, 2006]. The present study did not find substantial changes in cloudiness and precipitation at the peak of Kilimanjaro in response to deforestation during the dry season month of July, indicating minimal local forcing in this zone. However, this does not rule out the possibility of such impacts occurring in other seasons. The decrease in precipitation in response to deforestation at lower elevations complements the drying trend in the region associated with larger-scale changes in tropical climate, while increases at midelevations potentially mitigates such effects.

[42] Note that while the comparison of RAMS simulations for the current land use conditions against observations show adequate performance, there are also systematic biases exhibited by RAMS, especially in the simulation of precipitation. Since the magnitude of the biases itself could be impacted by land use [Ge et al., 2007], caution should be exercised when interpreting quantification of changes in cloudiness, precipitation, and surface thermodynamics. Another factor to be considered in the interpretation of the results is the inability of the 1° × 1° FNL analysis used for soil moisture initialization to adequately reflect small-scale variation of soil moisture in the study area. To assess the impact of land use change on orographic cloud formation, the approach utilized in this study focuses on the differences between the three experiments, provided that the CTL simulation performs adequately compared to observations. However, the possibility of soil moisture heterogeneity substantially modulating the differences between the experiments cannot be ruled out, and the results from this study should be viewed in the context of the initial soil moisture field utilized in the experiments.

6. Conclusions

[43] This study utilized numerical model simulations to investigate the impact of land cover changes at lower elevations of Kilimanjaro on the regional climate of the area. RAMS was used to simulate atmospheric conditions for July 2007, assuming current, deforested, and forested land cover scenarios. The findings from the comparison of these simulations can be summarized as follows.

[44] 1. Comparison of RAMS simulations for current land use conditions against surface meteorological observations and satellite observations of cloudiness show satisfactory performance of RAMS over the study region.

[45] 2. The RAMS simulations show that deforestation at lower elevations of Kilimanjaro lead to a decrease in the frequency of cloud occurrence at all elevations. The cloud liquid water path decreases in response to deforestation except at higher elevations on the windward side where it increases. Reforestation has the opposite effect, increasing frequency of occurrence of clouds at all elevations, increases in cloud liquid water path except at higher elevation on the windward side where it decreases.

[46] 3. Precipitation decreases at low elevations and increases at midelevations on the windward side in response to deforestation. On the leeward side, precipitation decreases at midelevations, while there is a very small increase at lower elevations. The magnitude of differences increases with the extent of deforestation.

[47] 4. Flow diversion values computed for the different scenarios also show that obstruction caused by Kilimanjaro is enhanced when the lower elevations areas are reforested.

[48] 5. Surface moisture patterns are also altered because of changes in terrain flow, with reforestation increasing moisture transport to the lee side of the mountain compared to current vegetation and deforestation.

[49] 6. While differences in surface moisture contributes to decrease in frequency of occurrence in cloudiness, changes in flow pattern caused by reduced aerodynamic roughness play an important role. When the lower-elevation regions are deforested, Kilimanjaro offer less obstruction to background flow, and the resulting increase in flow around the mountain causes reduced moisture transport to the lee side, causing reduced cloud liquid water path and precipitation. On the windward side, the increase in wind speed directed parallel to the topographic gradient at higher elevations, caused by reduced aerodynamic roughness in upwind areas, leads to enhanced surface convergence, cloud liquid water path, and precipitation.

[50] 7. Lack of precipitation at the peak during the period of study prevents making conclusions about potential impacts on precipitation at that level. Further study is required to investigate the possibility of such effects occurring during other seasons.

[51] This study addresses only the impact of deforestation on one dry season month. There are no compelling reasons for expecting the physical processes that cause the changes in clouds and precipitation to be substantially different if the analysis is extended to include the dry season month of July from other years. However, further study that extends the analysis to other seasons is required to establish the overall impact of land use change on the higher-elevation climate of Kilimanjaro.


[52] Jonathan Fairman was supported by NASA Headquarters under the NASA Earth and Space Science Fellowship Program (grant NNX08AU74H). Udaysankar Nair was supported by NASA NIP grant NNX06AF09G. Sundar Christopher was supported by the NASA Radiation Sciences Program. Resources supporting this work were provided by the NASA High-End Computing Program through the NASA Advanced Supercomputing Division at Ames Research Center.