Concept & Comparison

Iteration 1 — Classical image processing. Ten iterations of an NDVI + Otsu + watershed pipeline on hessian DOP20 orthophotos. Works on industrial/rail sites with isolated trees, fails structurally on residential areas: at 40–60 actual trees the pipeline plateaus at 34 detections. The missing information is not parameter-level — it is not in the RGB image.

Iteration 2 — Pretrained AI model. DeepForest (RetinaNet, University of Florida, trained on US NEON forest plots) was the obvious shortcut. On the same residential test: 115 detections instead of ~50, false positives on house roofs, sandboxes, parked cars, max confidence 0.67 instead of the usual 0.85–0.95. Diagnosis: domain gap. Five other pretrained models (Ventura Urban, SAM2, Detectree2, ArcGIS Tree Detection, BAMFORESTS) showed the same pattern.

Iteration 3 — Own consensus platform. The unavoidable consequence: a dataset of our own — and a mechanism that secures its quality. Multi-annotator blind labeling, Hungarian matching for circle assignment, five quality levels Q0–Q4, audit trail, full DSGVO-compliant on-premise stack.

Iteration 4 — TrustScore against the elderberry dilemma. A pure consensus has a known weakness: if 25 children consistently mark the same elderberry shrub as a tree, the consensus is high but the annotation is wrong. Four trust stages (novice/experienced/verified/expert), gold-tile based promotion, weighted aggregation, Q4 only with expert involvement. A hundred-fold school consensus can no longer overrule an expert judgment.

Iteration 5 — The missing dimension: height data. From RGB alone you cannot tell a 12 m tree from a 3 m shrub — and for the downstream temperature prognosis that matters. The hessian land survey (HVBG) provides DGM1 (terrain) and DOM1 (surface) as open data since February 2022. The difference — the nDSM — gives the local object height per square meter. Used in a dual-stream DeepLabv3+ (ISPRS 2022) and as didactic warning in the school PWA.