3D Reconstruction Using Single Multi-Spectral Imagery Under Limited Conditions

Digital Terrain Model (DTM in short), as an important data, is widely used in mapping, agriculture, forestry, industrial and other applicable filed. Photogrammetry is the main method to generate the DTM, which is based on the principle of feature extraction and feature matching. It is aim to find the pairs of matched points and get the parallax information. However, in some areas seriously lack of texture, such as deserts, waters, mountains and so on, it is unable to quickly and efficiently extract the feature points and edges, under this condition, the matching accuracy will be declined. Meanwhile, some areas often fail to obtain stereo image pairs of high spatial resolution due to the bad weather such as cloud and fog, so terrain reconstruction can’t be done by stereo method.Since single remote sensing image is more accessible compared with stereo pairs, if we can acquire the terrain details from single image and consider it as the auxiliary data in surface reconstruction; it is possible to generate an accurate dense surface model by using sparse ground control points with the help of the single imagery. Shape from Shading technology (we can call it SFS in short), is aim to obtain surface shape information from a single image, which is based on the inverse process of the physical imaging. It is an important tool for3D surface reconstruction and very potential in terrain surface reconstruction of a single remote sensing image. In this paper, SFS is introduced in geoscientific research. Interpolating the surface model according to the information about the surface details recovered from SFS, can not only generate the DTM for the areas where the stereo image pairs can’t be obtained, but also take the real trend of the internal surface of the grid into account. It is of great significance for quickly creating surface model for wide areas. The main works of this paper are as follows:(1) Solving the limitation of the reflectance in SFS theory under Lambertain surface assumption:described the Lambertian SFS theory and the two main types of solution method:minimization and propagating method. To solve the limitation in SFS theory that reflection should be a constant in the image, a reconstruction method combined the Intrinsic Image theory and SFS theory was proposed. The reconstruction error caused by surface reflectance variation was reduced by image decomposition. Experiments performed on synthetic images show that the intrinsic image decomposition can effectively separate the image reflectance component from the original image and improve the accuracy of the SFS algorithm.(2) Establishing the non-Lambertian surface fast reconstruction method:gave a detailed study of the Oren-Nayar model and the Phong model based3D reconstruction algorithm, rewrote the reflection map equation in the form of the Hamilton-Jacobi equation (referred to as H-J equation), then used the Lax-Friedrichs Hamiltonian in the equation discretization, finally utilized the high-ordered fast sweeping method (referred to as HO-FSM method) for estimating the surface normal vectors.(3) Reconstructing the area where stereo matching method fails to work:Based on the Oren-Nayar reflectance model, a theoretical framework of anisotropic surface reconstruction under the limited condition was proposed. First of all, supervised classification was performed, and the reflectance for each land cover type was estimated; then the reflection map equation was written as the H-J equation, and HO-FSM method was used for solving it; Finally, the least squares was utilized to interpolate the grid of control points.(4) Discussing the important parameters estimation problems in the SFS theory. This paper reviewed the approach how to estimate the illumination direction using image reflection equation. Then three classic methods were compared and analyzed in both synthetic and real image experiments.(5) Giving a combination framework for stereo and SFS algorithm on the basis of the advantages and disadvantages for each one:Firstly the FSM-SFS algorithm was implemented for the feature region reconstruction, and the results were used to form a new depth similarity measurement; then only the stable and reliable points were regarded as the matching points, and other points were reconstructed by SFS algorithm; Finally the results for all the surface points were combined to generate the3D model. The reconstruction experiment was implemented based on the three-line array moon images and the results show that the proposed method could give more accurate result than stereo and SFS each of themselves.In summary, this paper first solved the inherent problems in the current SFS theory, proposed solutions to improve the accuracy of SFS algorithm; then, in the area where stereo matching method fails to work, an DTM generation algorithm using single remote sensing imagery and certain ground control points was proposed; finally, when the stereo pairs are available, the SFS theory and stereo method was combined to improve the accuracy of stereo matching, so the DTM with higher resolution could be generated.