GausScape: Evaluating 3D Gaussian Splatting for Spatial Understanding in Greenhouse Environments

Commercial greenhouses are spatially complex systems of systems, that integrate technical infrastructure, with strict operational constraints in terms of access, safety, and/or maintenance. Thus, accurate spatial understanding is essential for tasks such as navigation and technical inspection. In practice, growers, engineers, and sellers frequently rely on photographs or videos when accessing greenhouse information remotely. Prior research shows that such 2D representations often fail to convey depth, and scale, leading to misinterpretations of layouts and operational constraints. Recent advances in 3D reconstruction offer potential alternatives. In particular, 3D Gaussian Splatting (3DGS), which generates photorealistic, view-dependent 3D representations from multi-view photographs. Compared to conventional mesh-based reconstructions, 3DGS preserves visual detail to a greater extent. As a result, 3DGS is being increasingly adopted for 3D representation-applications for practical uses. Yet, its actual contribution to users’ spatial understanding has not yet been empirically validated. This lack of empirical evidence creates a knowledge gap relating to the responsible use of immersive technologies in horticulture, specifically when assessing how 3D representations support spatial understanding of greenhouse systems compared to 2D representations (e.g., videos or photos). Accordingly, this project treats immersive realism as a variable to be critically examined rather than a goal in itself. To ensure methodological independence and avoid technology self-validation, the study is structured around a clear separation of roles. Splatto Works contributes real-world greenhouse cases and 3D representations. The experimental design, data collection, analysis, and interpretation are led by the Social Creative Technologies Lab at Wageningen University. Aryzon provides virtual world building expertise, enabling the creation of consistent high-quality 3DGS datasets. This separation ensures that findings are evidence-driven, and reproducible in other domains.