Modelling biomass of mountainous grasslands by including a species composition map

High mountain grasslands offer multiple goods and services to society but are severely threatened by improper land use practices such as abandonment or rapid intensification. In order to reduce abandonment and strengthen the common extensive agricultural practice a sustainable land use management of high mountain grasslands is needed. A spatially detailed yield assessment helps to identify possible meadows or, on the contrary, areas with a low carrying capacity in a region, making it easier to manage these sites. Such assessments are rarely available for remote and inaccessible areas. Remotely sensed vegetation indices are able to provide valuable information on grassland properties. These indices tend, however, to saturate for high biomass. This affects their applicability to assessments of high-yield grasslands.The main aim of this study was to model a spatially explicit grassland yield map and to test whether saturation issues can be tackled by consideration of plant species composition in the modelling process. The high mountain grassland of the subalpine belt (1800 – 2500 m a.s.l.) in the Kazbegi region, Greater Caucasus, Georgia, was chosen as test site for its strong species composition and yield gradients.We first modelled the species composition of the grassland described as metrically scaled gradients in the form of ordination axes by random forest regression. We then derived vegetation indices from Rapid Eye imagery, and topographic variables from a digital elevation model, which we used together with the multispectral bands as predictive variables. For comparison, we performed two yield models, one excluding the species composition maps and one including the species composition map as predictors. Moreover, we performed a third individual model, with species composition as predictors and a split dataset, to produce the final yield map.Three main grassland types were found in the vegetation analysis: Hordeum violaceum-meadows, Gentianella caucasea-grassland and Astragalus captiosus-grassland. The three random forest regression models for the ordination axes explained 64%, 33% and 46% of the variance in species composition. Independent validation of modelled ordination scores against a validation data set resulted in an R² of 0.64, 0.32 and 0.46 for the first, second and third axes, respectively. The model based on species composition resulted in a R² = 0.55, whereas the benchmark model showed weaker relationships between yield and the multispectral reflectance, vegetation indices, and topographical parameters (R² = 0.42). The final random forest yield model used to derive the yield map resulted in 62% variance explained and an R² = 0.64 between predicted and observed biomass. The results further indicate that high yields are generally difficult to predict with both models.The benefit of including a species composition map as a predictor variable for grassland yield lies in the preservation of ecologically meaningful features, especially the occurrence of high yielding vegetation type of Hordeum violaceum meadows is depicted accurately in the map. Even though we used a gradient based design, sharp boundaries or immediate changes in productivity were visible, especially in small structures such as arable fields or roads (Fig. 6b), making it a valuable tool for sustainable land use management. The saturation effect however, was mitigated by using species composition as predictor variables but is still present at high yields.