Main content area

A rule-based approach for mapping macrophyte communities using multi-temporal aquatic vegetation indices

Villa, Paolo, Bresciani, Mariano, Bolpagni, Rossano, Pinardi, Monica, Giardino, Claudia
Remote sensing of environment 2015 v.171 pp. 218-233
Landsat, autumn, biogeochemical cycles, carbon, data collection, freshwater ecosystems, lakes, macrophytes, reflectance, remote sensing, spring, temperate zones, vegetation index, wetlands, China, Hungary, Italy
Macrophytes are important components of freshwater ecosystems, playing a relevant role in carbon and nutrient cycles. Notwithstanding their widespread diffusion in temperate to subtropical shallow lakes, little effort has been performed so far in extensively mapping macrophyte communities at regional to continental scale. A rule-based classification scheme was implemented for mapping four macrophyte community types (helophyte, emergent rhizophyte, floating, and submerged-floating association). Input features were selected among multi-spectral reflectance and multi-temporal vegetation indices, based on Landsat data acquired over four test sites: Lake Taihu (China), Kis-Balaton wetland (Hungary), Lake Trasimeno and Mantua Lakes system (Italy). The best performing features were derived from Water Adjusted Vegetation Index (WAVI) computed at: early spring, maximum growth, and late autumn conditions. Overall accuracy (OA) and Kappa coefficient (k) of macrophyte maps produced with our approach over the ensemble of four sites were 90.1% and 0.865, respectively, with best performance in European temperate areas (OA=93.6–94.2%, k=0.887–0.916), and lower scores for subtropical Lake Taihu (OA=82.8%, k=0.762). Per-class accuracies were higher than 80% for all target classes, except for the submerged-floating association, with misclassifications concentrated in Taihu site. The robustness of the approach was tested over two independent validation cases: a different site (i.e. Lake Varese, Italy), and a different input dataset (i.e. AVNIR-2 data, for Mantua Lakes system). Consistent accuracy results were achieved: OA=94.3% (k=0.922) and OA=85.6% (k=0.766), with some misclassification due to spatial resolution of AVNIR-2 data.