Leaf area index (LAI) is defined as one half of the total leaf surface area per unit ground surface area (projected on the local horizontal datum). It is a key biophysical vegetation property describing biome-specific canopy structure [1
], and an essential variable in models of ecosystem processes and productivity [2
], crop productivity [4
] and hydrology [5
]. Prescribing these models with accurate LAI parameters is, however, challenging due to the scarcity of landscape scale LAI measurements for most part of world’s vegetated biomes. Data deficiencies are especially acute in tropical regions and across degraded woodlands [6
Earth Observation (EO) data can be exploited to fill these data gaps [8
]. LAI products derived from EO data describe variation in regional and global vegetation leaf area at 250 m to 6 km spatial resolutions and at 8-day to monthly temporal resolutions. Some are generated via inversion of physically-based models against observations of surface-leaving radiation [10
]. Others are produced using empirical and semi-empirical relationships between LAI and spectral vegetation indices [12
The Global Inventory Modeling and Mapping Studies (GIMMS) LAI3g product was derived from the Advanced Very High Resolution Radiometers (AVHRR) sensors. It provides one of the longest LAI time-series currently available for canopy structure trend assessments [8
], spanning the period from July 1981 to December 2011 at 15-day intervals and at 1/12 degree spatial resolution [14
]. Applications to date include the assessment of latitudinal vegetation growth trends [15
] and trends in vegetation greenness for pastures [16
]. The GIMMS LAI3g product was generated using a feed-forward neural network, using AVHRR third generation normalized difference vegetation index (NDVI3g) between 1981 and 2011. The neural network itself was trained through back-propagation using MODIS LAI between 2001 and 2009, and GIMMS data over the same period [14
]. Validation of the GIMMS LAI3g product was carried out using field data from just 29 sites, predominantly in northern latitude cropland, grassland and boreal forest biomes. There have been no validation studies of GIMMS LAI3g in tropical biomes.
Here, we use a new dataset of ground-measured LAI (hereafter, LAITrue
] to validate the GIMMS LAI3g product (hereafter, LAIavhrr
) in East Africa. We compare biome-specific LAI values, and test whether both LAITrue
capture similar trends in environment-LAI response. In so doing, we characterize LAI within and between East African biomes, and quantify spatial variation in response to climate, topography and disturbance.
The ultimate objective of linking EO data to biophysical land surface attributes (e.g., vegetation leaf area, biomass and productivity) is to characterize those attributes on large spatial scales and over time with minimal need for further fieldwork [42
], so that the logistical, financial and subjective sampling constraints of fieldwork can be overcome. This requires that EO products, such as GIMMS LAI3g, accurately reflect conditions on the ground, both in absolute values and in their variation across space and time [43
]. In this study, the increase in LAI with vegetation height for natural biomes (from shrubland/bushland to woodland and forest) is captured by both LAIavhrr
. However, we find that LAIavhrr
significantly overestimates LAI relative to LAITrue
for all biomes except for shrubland and cropland. The coarse-spatial resolution of LAIavhrr
may cause a bias towards the global mean of the region. Yet in East Africa, one would expect such a bias to result in lower LAI compared with estimates derived at high spatial resolution such as field-based LAITrue
, because these landscapes are highly heterogeneous, and increasingly modified towards low LAI biomes such as degraded woodlands and crops. Moreover, LAIavhrr
extracted from large tracts of homogeneous forest in Eastern Congo was found to be similar to LAIavhrr
extracted for forests in our study area, again suggesting a consistent overestimation in the GIMMS LAI3g product.
In tropical Africa, vegetation phenology is particularly tuned to precipitation and its seasonal shifts [45
], which may be impacted by rising global temperatures [47
]. Rainfall triggers the emergence of green leaves in deciduous biomes, and controls vegetative growth and growth-duration in semi-arid and arid environments [48
]. Vegetation indices such as NDVI, as a surrogate of vegetation productivity and phenology, have been linked to rainfall and rainfall and rainfall variability [49
] in East Africa via a log-linear relationship. Nonlinear responses of LAI to PPT (and derived MI) in our study area are captured by both LAIavhrr
by the same functional relationship. LAI increases with water availability up to a maximum, declining at higher values of annual rainfall and moisture. However, the response curves differ significantly, in that field-derived LAI reaches its maximum at lower values of the respective environmental driver, and exhibits a more rapid rate of change.
Further to climatic constraints, our analyses show that socio-economic factors play an important role in observed vegetation structure. LAI was found to be higher with increasing terrain steepness and within protected areas. These findings add evidence, based on biophysical structure, to previous studies showing how inaccessibility functions as a passive protection for woody biomes [18
]. Our results are especially relevant for environmental management concerned with the maintenance of ecosystem processes and function, which are linked to biophysical properties such as LAI. Curiously, LAI was also found to be higher in regions under greater population pressure. Population pressure is greatest in areas of high rainfall, which are suitable for both crop production and high-LAI forest (of which there is less). Further, associated resource demands on adjacent woodlands and forests mean that, in many areas, these biomes have degraded to bushland and shrubland, which may exhibit higher biome-specific LAI than rainfall-limited bushland and shrubland elsewhere. Thus the observed pattern may reflect historical land use impacts rather than a positive causal relationship between human population pressure and vegetation leaf area. A visual comparison of LAI rainfall isolines along a gradient of increasing population pressure shows primarily flat slopes, supporting the above hypothesis. This pattern is especially apparent in the LAIavhrr
response and, if at all, there seems to be a negative effect of increasing population pressure on LAI in regions with high rainfall.
Validating coarse-scale spatial resolution data with field measurements is inherently difficult due to challenges of spatial mismatch [50
] and differing spatial scales [51
]. LAI distributions inferred from field measurement and satellite retrieval should ideally converge to the true intrinsic distribution of the vegetation class in a given region at a given time [52
]. Whilst our findings are encouraging regarding the capacity of EO data to capture spatial variation in LAI along major environmental gradients, they also highlight the need for further field assessments of inter- and intra-annual LAI dynamics, especially in remote woody biomes, in croplands and in plantations. An important issue, arising from our use of MODIS land cover products in the generation of biome-specific random plots, is that plantations are difficult to distinguish from natural forests, especially if consisting of broadleaved evergreen trees. We find that very few plantations are identified by the MODIS product, despite large areas of forests being converted to plantations in both Kenya and Ethiopia. Similarly, spectral similarities complicate the distinction between “woody savanna” and savanna, as well as between grassland and cropland [30
], particularly when the latter are rotated in seasonal succession rather than being left fallow. It is encouraging, however, that despite these uncertainties in biome-specific LAIavhrr
estimates, the increase in LAI with vegetation height is captured by LAIavhrr
similarly. Furthermore, general additive models identified the same set of environmental predictors having similar impacts on both LAI estimates.
Remotely sensed, global-scale estimates of biophysical structure, such as LAIavhrr effectively capture spatial variation in LAI along major environmental gradients, but some challenges remain. First, LAIavhrr appears to consistently overestimate LAI in woody biomes such as forest. Second, nonlinear responses of LAI to water availability, while captured by both LAIavhrr and LAITrue similarly, differ significantly in their parameterization, such that the remotely sensed product reaches its maximum at higher values of rainfall and moisture index than field-derived LAITrue suggests may be the case. LAITrue increased to 2.23 at MI = 0.76 mm and to 3.19 at PPT = 1400 mm decreasing thereafter, while LAIavhrr increased to 3.13 at MI = 1.58 and to 3.48 at PPT = 2700 mm decreasing thereafter. Third, distance correlation statistics show significant relationships with all environmental drivers tested for both LAITrue and LAIavhrr, although the strength of that correlation varies between LAITrue and LAIavhrr responses.
More generally, we find that the same set of environmental drivers emerges as significant in models of LAI variability. Beside rainfall, temperature and topography, socio-economic correlates such as population pressure, distance to roads, distance to towns and protection status are important for understanding spatial variation of vegetation biophysical structure in the human-modified landscapes of tropical Africa. The responses to these drivers are largely consistent for LAIavhrr and LAITrue except for their responses to potential evapotranspiration. Note though, that protection status, rainfall, temperature range and topography were not significant in general linear models of LAIavhrr variability despite being significant in general additive models suggesting complex underlying relationships.
In particular, higher forest LAI observed within the protected area network suggests a potential mechanism for monitoring efforts to reduce forest degradation (e.g., for carbon conservation or catchment protection). Given the above challenges, targeted field measurements, stratified according to both environmental and socio-economic gradients, will be needed to improve the accuracy of satellite-derived estimates.