Accurate vegetation cover maps of forested areas are crucial for ecosystems monitoring, as well as for management of water balance, flood and fire risk, and other forest-associated resources. With this regard, remote sensing techniques have been used for land cover mapping worldwide. Here, we propose a vegetation-mapping methodology in a dehesa environment using ultra-high spatial resolution imagery (UHSR) with a spatial resolution of 0.25 m and four bands in the visible and near-infrared spectrum. Land cover categories were defined by their runoff generation capability and considered two levels of disaggregation: among species (macro-class level) and within species (class level). Additionally, we developed a method to reduce field campaigns and manual work by transferring random forest classifiers trained with a group of images (training group) to neighboring images (validation group). The training group was remarkably accurate, achieving an overall accuracy of 91.6% (k = 0.89) at the class level and 95.8% (k = 0.94) at the macro-class level. The results for the validation group were also very high, with an overall accuracy of 78.3% (k = 0.74) at the class level and 86.3% (k = 0.82) at the macro-class level. Moreover, we found that the blue band, soil color index, and texture features have a great influence on species discrimination, especially within shrub species in dehesa environments. Notably, having accurate land cover maps is crucial, given that the use of a global database led to underestimating the potential runoff in the most representative land cover in the dehesa environment. Future research will focus on the automatic generation of new samples extracted from the classified UHSR images, which could be used as training datasets for the supervised classification of other high spatial resolution images (e.g., Sentinel imagery) for regional-scale hydrological models.