An innovative approach to improve SRTM DEM using multispectral imagery and artificial neural network

Although the Shuttle Radar Topography Mission [SRTM) data are a publicly accessible Digital Elevation Model [DEM) provided at no cost, its accuracy especially at forested area is known to be limited with root mean square error (RMSE) of approx. 14 m in Singapore's forested area. Such inaccuracy is attributed to the 5.6 cm wavelength used by SRTM that does not penetrate vegetation well. This paper considers forested areas of central catchment of Singapore as a proof of concept of an approach to improve the SRTM data set. The approach makes full use of (1) the introduction of multispectral imagery (Landsat 8), of 30 m resolution, into SRTM data; (2) the Artificial Neural Network (ANN) to flex its known strengths in pattern recognition and; (3) a reference DEM of high accuracy (1 m) derived through the integration of stereo imaging of worldview‐1 and extensive ground survey points. The study shows a series of significant improvements of the SRTM when assessed with the reference DEM of 2 different areas, with RMSE reduction of ∼68% (from 13.9 m to 4.4 m) and ∼52% (from 14.2 m to 6.7 m). In addition, the assessment of the resulting DEM also includes comparisons with simple denoising methodology (Low Pass Filter) and commercially available product called NEXTMap® World 30™. Key Points: Significant improvement of SRTM beneficial to patch areas with missing reliable data Improvement in heterogeneous forest, where SRTM accuracy is low and survey is costly Applicable when missing areas share similar land cover with training reference