Comparison of Height Models from High Resolution Aerial Images and from LiDAR

A digital Surface Model (DSM) has been generated with an Optech Gemini laser scanner with approximately 4 points/m2 in an area also covered by an UltraCam Eagle with 5cm GSD. Based on the digital images DSM and digital terrain models (DTM) were automatically/interactively produced with BAE SYSTEMS “NGATE” (New Generation of Automatic Terrain Extraction) and also SimActive’s “Correlator3D” with 50cm point spacing. A direct comparison of the overlapping LiDAR strips as well as the comparison of the LiDAR data with the height models based on the aerial data showed differences especially in areas covered by vegetation as well as at building limits, requiring a filtering for the correct comparison. As for all height models it is not so simple to specify the accuracy figures correctly. The standard deviation as well as the normalized median absolute deviation (NMAD), LE90 and LE95 by theory is based on normal distributed discrepancies. Normal distributed height discrepancies require corresponding point definition in compared height models and within the height models similar conditions as e.g. flat terrain. Operationally it is very difficult to filter height models in a manner leading to normal distributed values, so the frequency distribution of height discrepancies has to be investigated. The standard deviation is very sensitive for higher number of larger discrepancies which usually cannot be avoided. NMAD is not as sensitive as the standard deviation and describes the frequency distribution of height discrepancies in a better manner. The threshold values LE90 and LE95 are even depending upon larger discrepancies as the standard deviation and should be avoided. LiDAR and image matching both have advantages and disadvantages, with both methods qualified height models can be generated.