Accuracy Assessment of Digital Surface Models Based on WorldView-2 and ADS80 Stereo Remote Sensing Data

Martina L Hobi,Christian Ginzler

doi:10.3390/s120506347

Abstract

Digital surface models (DSMs) are widely used in forest science to model the forest canopy. Stereo pairs of very high resolution satellite and digital aerial images are relatively new and their absolute accuracy for DSM generation is largely unknown. For an assessment of these input data two DSMs based on a WorldView-2 stereo pair and a ADS80 DSM were generated with photogrammetric instruments. Rational polynomial coefficients (RPCs) are defining the orientation of the WorldView-2 satellite images, which can be enhanced with ground control points (GCPs). Thus two WorldView-2 DSMs were distinguished: a WorldView-2 RPCs-only DSM and a WorldView-2 GCP-enhanced RPCs DSM. The accuracy of the three DSMs was estimated with GPS measurements, manual stereo-measurements, and airborne laser scanning data (ALS). With GCP-enhanced RPCs the WorldView-2 image orientation could be optimised to a root mean square error (RMSE) of 0.56 m in planimetry and 0.32 m in height. This improvement in orientation allowed for a vertical median error of −0.24 m for the WorldView-2 GCP-enhanced RPCs DSM in flat terrain. Overall, the DSM based on ADS80 images showed the highest accuracy of the three models with a median error of 0.08 m over bare ground. As the accuracy of a DSM varies with land cover three classes were distinguished: herb and grass, forests, and artificial areas. The study suggested the ADS80 DSM to best model actual surface height in all three land cover classes, with median errors <1.1 m. The WorldView-2 GCP-enhanced RPCs model achieved good accuracy, too, with median errors of −0.43 m for the herb and grass vegetation and −0.26 m for artificial areas. Forested areas emerged as the most difficult land cover type for height modelling; still, with median errors of −1.85 m for the WorldView-2 GCP-enhanced RPCs model and −1.12 m for the ADS80 model, the input data sets evaluated here are quite promising for forest canopy modelling.

Highlights

Digital surface models (DSMs) depict the elevation of surfaces visible from the sensor, such as building tops, tree tops, or unoccluded bare ground [1]
It shows that the WorldView-2 ground control points (GCPs)-enhanced rational polynomial coefficients (RPCs) DSM represents the whole picture in good quality but the ADS80 DSM is able to retrieve more details and finer-scale variations of the forest canopy
We were able to show that: (1) using GCPs the RPCs defining the orientation of the WorldView-2 images can be improved; (2) the WorldView-2 DSM with GCP-enhanced RPCs achieves much higher accuracy measures than the WorldView-2 DSM where only the RPCs are used; (3) the accuracy of the WorldView-2 GCP-enhanced RPCs DSM is similar to the ADS80 DSM (2*Ground sample distance (GSD)); (4) the accuracy of a DSM varies with land cover type; and (5) forested areas are the most challenging areas for surface height modelling among the land cover types evaluated here

Summary

Introduction

Digital surface models (DSMs) depict the elevation of surfaces visible from the sensor, such as building tops, tree tops, or unoccluded bare ground [1]. In forest science DSMs are used to model the canopy surface of forests and analyse its vertical structure [2,3]. DSMs enable the 3D modelling of the forest canopy, which allows assessments of tree cover [4], estimation of crown structure [5], measurements of canopy heights [6,7] and the detection of canopy gaps [8], including the monitoring of these properties over time. In general the preferred data source option for digital surface modelling is a balance between the desired accuracy of the DSM, the costs involved in its creation and the availability of the input data [9]

Objectives

Methods

Results

Discussion

Conclusion