The creation and editing of digital terrains are a fundamental step in the development of virtual worlds for various domains ranging from scientific simulations to entertainment. Despite advances in terrain editing tools, fully meeting the expressive requirements of artists remains a challenge, whose main aspects are interactivity and control.

This thesis is divided into two parts. First, we explore a deep learning-based approach to learn a rich latent representation of terrains. This model not only enables generating terrains with varied styles but above all offers new capabilities for localized style control, coherent detail amplification, and editing guided by sketches or low-resolution terrains.
 
The second part introduces an approach based on a continuous vector representation, defining the terrain as a sum of primitives. This method includes an efficient algorithm to vectorize raster terrains and proposes an editing system based on a structural skeleton derived from ridge lines. This interface enables direct, semantic, and intuitive manipulation of major geomorphological features, providing precise local and global structural control.
 
These two approaches offer complementary capabilities, ranging from style control to structural editing. Particular effort was dedicated to their practical integration through the development of dedicated user interfaces, facilitating their evaluation and potential adoption. By enriching the toolkit available to artists with methods specifically designed for style or structure control, this thesis contributes to making terrain creation and editing more expressive, efficient, and better aligned with artistic creative processes.