Evaluation of DSMs generated from multi-temporal aerial photographs using emerging structure from motion–multi-view stereo technology

• We generate DSMs from aerial photographs by using SfM–MVS.
• The DSM accuracy is assessed by comparison with leveling data.
• GNSS measuring and auto-leveling were carried out for ground control and evaluation.
• The DSM accuracy is similar to the resolution of aerial photographs.

An accuracy assessment of digital surface models (DSMs) generated from archived aerial photographs using the structure from motion–multi-view stereo (SfM–MVS) technique was carried out. A four-step accuracy-assessment procedure was adopted using aerial photography from eight periods, as follows. Step 1: generate a DSM and orthophoto from digital aerial photographs taken in 2013 and ground control points (GCPs) measured by GNSS. Step 2: assess the accuracy of the DSM by comparison with altitude measured by leveling survey. Step 3: generate other historical DSMs and orthophotos from historical aerial photographs using GCPs extracted from the DSM of 2013. Step 4: assess the accuracy of all historical DSMs by comparing with the leveling survey. Then re-calculate the accuracy of historical DSMs by reducing the inherent error in the 2013 DSM. The DSM based on the aerial photographs taken in 2013 was generated with a resolution of 48.2 cm. The residual height error of the GCPs was 15.4 cm. Validation against the altitudes of 171 points revealed that this DSM has a height root-mean-square-error (RMSE) of 24.1 cm and is 9.2 cm lower than the leveling data on average. Even using US military photos with unconfirmed detailed specifications, the model can measure the altitude with an RMSE value of 121.5 cm. It appears therefore that analysis by SfM–MVS can give comparable measurement accuracy to traditional aerial photogrammetry. The low cost and high accuracy obtained with archived aerial photographs are worthy of special mention.