Abstract: This paper focuses on machine learning in intelligent photogrammetry:the elevation difference fitting neural network method.The limitations of observation technologies and processing methods lead to the lack of global high-accuracy seamless DEMs,which further restrict DEMs’ application in the fields of hydrology,geology,meteorology,military and other applications.In this paper,we propose a multi-source DEM fusion method using the neural network model trained based on elevation difference.The proposed method is employed to generate a high-quality seamless DEM dataset blending SRTM1,ASTER GDEM v2,and ICESat GLAS.At first,the ICESat GLAS data were filtered according to the relevant parameters and the elevation differences with DEMs.The threshold of elevation difference adaptively varied with terrain slope to remove the abnormal points effectively.The neural network was then applied to learn the error distribution of ASTER GDEM v2,using the ICESat GLAS data as the control points.We constructed the network input composed of slope information,latitude and longitude coordinates,while the elevation difference of ICESat GLAS and ASTER GDEM v2 were set as the target output.The corrected ASTER GDEM v2 results can be obtained by adding the predicted output to the original elevation values.At last,the corrected ASTER GDEM v2 values were utilized to fill the voids of SRTM1,where the vertical bias between the datasets were dealt with TIN delta surface method to blend the seamless DEM.Randomly selected data were used for actual experiments,and the proposed model was evaluated by comparing with other methods and DEM datasets through quantitative evaluation and visual discrimination.Experiment results show that the proposed method has lower value of RMSE than compared methods both in void or the whole area,which can effectively overcome the influence of outliers in ASTER GDEM v2,and generate seamless DEM.

Key words: multi-source DEM fusion, neural network, TIN delta surface, slope adaption