While the GEOBIA paradigm has led to significant improvements in the analysis and understanding of remote sensing images thanks to the processing of objects (i.e. regions) instead of pixels, it still requires to identify the objects (or segment the image into regions) before applying rules for classifying the extracted objects. This segmentation step is not straightforward and relies on user expertise or empirical tuning to be adapted to each new scene to be processed. Thus, it cannot be used for Big GeoData where large-scale analyses require methods that are both very efficient and robust to the wide variety of scenes to be observed. We address here these multiple issues by relying on a multiscale image representation that embed in a tree structure, with no need of parameter tuning, the different (nested) objects to which a pixel can belong. Computation of such a stack of segmentations benefit from some recent scalable implementations that make realistic their very fast extraction from large-scale image datasets. Once the tree structure has been extracted, further image analysis is conducted at a very low computational cost, and relies on Differential Attribute Profiles (DAP). These state-of-the-art features describe a pixel by the properties of the objects it belongs to. We apply these profiles to both the original multispectral bands and some derived features such as NDVI or textural information. We then apply the Random Forest classifier on the extracted DAP features, or more precisely on only the most relevant features. We benefit from the efficiency of the different steps (tree construction, feature extraction, training, prediction) to propose a semi-supervised strategy where we retrain the model for each kind of landscape, thus allowing to tackle the great variety in appearances of objects at a very large-scale (e.g. VHR imagery at Pan-European scale). Due to the low computational cost (e.g. a few minutes for a Pléiades or WorldView-3 scene), a user can then interactively improve the classification by updating the reference samples used for training the model. The proposed scalable solution fully relies on open source components (Orfeo ToolBox, Boost, GDAL, Shark, Triskele OTB remote module) and so can be used in any GEOBIA applications. To illustrate our methodology, we consider here the Mapping of Small Woody Features (SWF), that is to be included as part of a new High-Resolution Layer (HRL) covering the whole of Europe from Iceland to Turkey within the Copernicus pan-European component of the land monitoring service. SWF represent some of the most stable vegetated linear and small landscape objects providing numerous ecological and socio-cultural functions related to soil and water conservation, climate protection and adaptation, biological diversity and cultural identity. Extracting these objects over such a large area (almost 6 million sq.km) from VHR imagery brings numerous challenges: large amount of data (greater than 100TB), large number of individual image scenes (greater than 25,000), diversity of the European landscapes, and the need to process these data in a timely manner whilst ensuring a satisfactory degree of precision.