Browsing by Author "Itkonen, Pekka"

The indigenous forests of the Taita Hills, SE Kenya, boast a vast biodiversity and provide several vital ecosystem services to local communities. Population growth and land use change pressures have resulted in a significant decrease in indigenous forest cover in the Taita Hills in recent decades. Quantifying the aboveground biomass (agb) and carbon sequestration capacity of the Taita forests provides a strong argument for striving for their more efficient protection in the context of UN-REDD programme. Although the role of tropical forests as global carbon sinks has been widely recognized, their agb and leaf area index (LAI) remain uncertain. Optical remote sensing (RS) provides a cost-effective means of LAI and agb estimation in remote areas, but empirical modeling using remote sensor data has limited certainty in densely vegetated tropical forests. The agb and LAI of the Taita Hills were estimated using empirical regression modeling by relating in situ data (n = 181 for agb, n = 172 for LAI) and spectral vegetation indices (SVIs) derived from SPOT HRVIR optical remote sensing data. Field plots (20 m x 20 m = 0.04 ha) were located in indigenous (n = 80) and exotic (n = 83) forests, woodlands (n = 9) and agroforestry areas (n = 9). In situ LAI was derived from hemispherical photography (HP) using Lang's approach and the foliage clumping correction method by Chen & Cihlar. In situ agb was estimated using allometric equations which relate agb with tree parameters such as tree diameter at breast height. Empirical relations between the response variables (agb, LAI) and SVIs were utilized in predictive regression modeling. The predictor variables were selected using forward stepwise regression based on Akaike Information Criterion (AIC) values. The regression models resulted having only one predictor each due to the redundancy of the SVIs. Also topography-based predictor variables were tested, but they were poorly or not at all related with LAI and agb. The models performed moderately (D2 = 0.62 for LAI model, D2 = 0.41 for agb model). The total agb and carbon sequestration of the study area were estimated as 4.264 Tg and 2.132 Tg C, respectively. Mean agb densities of the indigenous forests and the whole study area were estimated as 463 ± 190 Mg ha-1 and 126 ± 115 Mg ha-1, respectively. Mean in situ LAI of the indigenous forests and all plots were estimated as 3.66 ± 0.44 and 3.12 ± 0.84, respectively. Indigenous plots had the highest mean in situ agb density and LAI values compared to exotic forests, woodlands and agroforestry areas (ANOVA p < 0.001). The RMSE values of the models were 0.59 (18.6 %) for LAI and 376.85 Mg ha-1 (82.9 %) for agb. The agb model was negatively biased (bias: -107.1 Mg ha-1, 23.6 %), but there was no statistically significant bias in the LAI model. The resulting agb estimates are rather high due to high in situ agb values, partly resulting from the emphasized contribution of very large trees to biomass on small plots. LAI values are quite low for dense tropical forests due to indirect estimation method using HP, but still comparable with other similar studies. As expected, the modeling performance was impaired by SVI saturation effect in relation to LAI and agb. The agb model was biased most likely due to the use of transformed variables in linear regression. The predictive models are not transferable to other regions as such, for the relative prediction performance of SVIs is case-specific and the model parameters have to be estimated using in situ data for each site. In order to improve the model credibility, a more extensive dataset based on a random or a systematic sample should be used, having larger plot size and containing more observations with low LAI and agb values.