LAUSR

Abstract The anthropic activities carried out in the vicinity of the water body trigger various harmful effects on flora, fauna, and human health. The release of effluents and nutrients renders the aquatic environment increasingly eutrophic, favoring the proliferation of macrophytes, leading to loss of biodiversity, in addition to degradation of water quality. Since land use and land cover (LULC) around water bodies strongly influence macrophyte proliferation, it is necessary to simultaneously monitor the water body and its surroundings to manage and control the overabundance of aquatic vegetation. Due to the spectral similarity between aquatic and terrestrial vegetation, it is a challenge to simultaneously monitor these targets, especially when macrophyte mapping is placed within the context of LULC mapping. The inclusion of auxiliary data that promotes an increase the aquatic vegetation discrimination can even improve the accuracy of a LULC mapping. However, an excessive number of features can increase the complexity and reduce the mapping accuracy. We assume that it is necessary to select a set of suitable features, including spectral and textural data extracted from multispectral images, in addition to a DEM, allows the macrophytes discrimination in a water body and increases the classification accuracy in a simultaneous mapping of LULC classes, but only with the addition of a DEM, there is better discrimination of aquatic macrophytes in a simultaneous mapping of LULC classes. To achieve this, four feature selection methods (Correlation, Gain Ratio, Relief, and out-of-bag Random Forest) were applied to define the best subsets for a LULC mapping that integrates the water body with the watershed landscape. The various subsets of selected features were classified using a supervised Artificial Neural Network (ANN) algorithm. The results show that features derived from OLI/Landsat 8 combined with DEM data provided an effective contribution to increase the accuracy classification and reduce confusion between aquatic and terrestrial vegetation. The best kappa coefficient was obtained from the Correlation features set, achieving a 10% improvement after the inclusion of DEM. Thus, the results of this study showed that the DEM inclusion can improve the aquatic vegetation discrimination in LULC classifications, even when different feature selection methods do not indicate that variable as relevant to machine learning-based classification.