Accuracy Assessment and Improvement of SRTM DEM based on ICESat/GLAS under the Consideration of Data Coregistration over Jiangxi Province

At present, ICESat/GLAS has become the main data source for large-scale SRTM DEM accuracy assessment. Nevertheless, almost all the existing methods neglected the effective coregistration of the two datasets. In order to evaluate the importance of data coregistration, this paper took Jiangxi Province as the research area and analyzed the overall accuracy of SRTM DEM before and after data coregistration. Results show that after data coregistration, the Mean Bias (ME) of SRTM DEM was eliminated significantly, and the DEM Root Mean Square (RMSE) was reduced by 14.4%. We further analyzed the effect of terrain factors (terrain slope, terrain aspects, terrain relief, elevation) and land use types on the accuracy of SRTM DEM. Specifically, this study area was divided into different sub-regions according to slope ranges (0~5°, 5~10°, 10~15°, 15~20°, >20°), aspect ranges (-1, 0~22.5°, 22.5~67.5°, 67.5~112.5°, 112.5~157.5°, 157.5~202.5°, 202.5~247.5°, 247.5~292.5°, 292.5~337.5°, 337.5~360°), relief ranges and elevation ranges (0~100 m, 100~200 m, 200~300 m, >400 m), and land use types (cultivated land, forest cover, grassland, water area, built-up area, unused land), respectively. Then, the ME and RMSE of each sub-region were computed and analyzed. We found that the terrain aspects with a sine-like shape were strongly related to SRTM DEM errors before data coregistration; however, this relationship basically disappeared after data coregistration. The SRTM DEM errors increased with the increase of terrain relief, slope, and elevation. Among the six land use types, SRTM DEM had different accuracy under different land use types. More specifically, SRTM DEM had the highest accuracy on unused land and the lowest accuracy on forest land. Finally, by incorporating terrain slope, aspect, terrain relief, elevation, land use, and ICESat/GLAS data randomly selected with the proportion of 90% into the revision models, the SRTM DEM was improved by use of Multiple Linear Regression (MLR), Back Propagation Neural Network (BPNN), Generalized Regression Neural Network (GRNN), and Random Forest (RF), respectively. Accuracy evaluation of corrected SRTM DEM by use of the remaining 10% ICESat/GLAS data demonstrated that the four correction models with data coregistration obviously outperform themselves without the coregistration. Among the four corrected models, RF produced the best result while GRNN produced the worst result. The RMSE of RF was about 3.1%, 2.7%, and 11.3% lower than those of MLR, BPNN, and GRNN, respectively. Therefore, RF was finally selected to enhance accuracy of SRTM DEM. © 2021, Science Press. All right reserved.