Built from scratch rather than inherited, the system emerged from large-scale agriculture constrained by unreliable connectivity and power, limited budgets and specialist capacity, few opportunities for repeated flights, and repetitive landscapes where feature matching is difficult.

Observation quality is controlled before computation. Solar time, weather, illumination, GSD, overlap, and area–time rotation form a shared data contract for photogrammetry and machine learning, reducing image volume and limiting sources of variability that can contribute to reconstruction failure and downstream false negatives.

The architecture manages uncertainty rather than eliminating it prematurely. Coordinates guide candidate selection and initial placement as priors, while image evidence, robust multi-view geometry, and bundle adjustment determine reconstruction. A high-recall GPU matching path preserves redundancy in repetitive scenes, and performance is evaluated end-to-end because faster components do not necessarily produce a better system.

Offline-first is an execution contract. Raw imagery is processed locally without network access, and only validated products are transferred when connectivity becomes available. The containerized workflow repeatedly produces orthophoto COGs, point clouds, tree-level layers, height indicators, and machine-learning reports across datasets, operators, and sites while remaining ready for later cloud-native use.