Chuanfa Chen

A method of DEM construction and related error analysis

The concept and the computation of terrain representation error (ETR) are investigated and total DEM error is presented as an accuracy index for DEM evaluation at a global level. A promising method of surface modelling based on the theorem of surfaces (SMTS) has been developed. A numerical test and a real-world example are employed to comparatively analyze the simulation accuracy of SMTS and the classical interpolation methods, including IDW, SPLINE and KRIGING performed in ARCGIS 9.1 in terms of sampling and interpolation errors and of total DEM error. The numerical test shows that SMTS is much more accurate than the classical interpolation methods and ETR has a worse influence on the accuracy of SMTS than those of the classical interpolation methods. In a real-world example, DEMs are constructed with SMTS as well as the three classical interpolation methods. The results indicate that, although SMTS is more accurate than the classical interpolation methods, a real-world test indicates that there is a large accuracy loss. Total DEM error composed of, not only sampling and interpolation errors, but also ETRs can be considered as a good accuracy measure for DEM evaluation at a global level. SMTS is an alternative method for DEM construction.

A high speed method of SMTS

In order to improve the computational speed of the method of surface modeling based on the theorem of surfaces (SMTS), a modified Gauss-Seidel (GS) method (MGS) was introduced and a MGS of SMTS (SMTS-MGS) has been developed. Numerical tests show that SMTS-MGS is more than twice as fast as SMTS-GS and much faster than the classical iterative methods provided by MATLAB 7.0. The computing time of SMTS-MGS is proportional to the first power of the total number of grid cells in the computational domain, whereas the classical SMTS computing time is proportional to the third power of the total number of grid cells. A real-world example of constructing a series of DEMs of Dongzhi tableland was employed to comparatively analyze the simulation accuracies of the two versions of SMTS including SMTS-MGS and the classical SMTS, and the well parameterized classical interpolation methods including the inverse distance weighted technique (IDW), kriging, thin plate spline (TPS), regularized TPS, TPS with tension, and ANUDEM Version 4.6.3, at spatial resolutions of 5, 12, 20, and 25m. The real-world example demonstrates that SMTS-MGS with the same accuracy as SMTS is approximately as accurate as the third-order TPS and much better than other classical interpolation methods at almost all spatial resolutions, except for ANUDEM at a spatial resolution of 5m. On average, SMTS-MGS provides the best results with a minimum of computing time.

A robust method of thin plate spline and its application to DEM construction

In order to avoid the ill-conditioning problem of thin plate spline (TPS), the orthogonal least squares (OLS) method was introduced, and a modified OLS (MOLS) was developed. The MOLS of TPS (TPS-M) can not only select significant points, termed knots, from large and dense sampling data sets, but also easily compute the weights of the knots in terms of back-substitution. For interpolating large sampling points, we developed a local TPS-M, where some neighbor sampling points around the point being estimated are selected for computation. Numerical tests indicate that irrespective of sampling noise level, the average performance of TPS-M can advantage with smoothing TPS. Under the same simulation accuracy, the computational time of TPS-M decreases with the increase of the number of sampling points. The smooth fitting results on lidar-derived noise data indicate that TPS-M has an obvious smoothing effect, which is on par with smoothing TPS. The example of constructing a series of large scale DEMs, located in Shandong province, China, was employed to comparatively analyze the estimation accuracies of the two versions of TPS and the classical interpolation methods including inverse distance weighting (IDW), ordinary kriging (OK) and universal kriging with the second-order drift function (UK). Results show that regardless of sampling interval and spatial resolution, TPS-M is more accurate than the classical interpolation methods, except for the smoothing TPS at the finest sampling interval of 20m, and the two versions of kriging at the spatial resolution of 15m. In conclusion, TPS-M, which avoids the ill-conditioning problem, is considered as a robust method for DEM construction.

The smoothness of HASM

To smooth noises inherent in uniformly sampled dataset, the smoothness of high accuracy surface modeling (HASM) was explored, and a smoothing method of HASM (HASM-SM) was developed based on a penalized least squares method. The optimal smoothing parameter of HASM-SM was automatically obtained by means of the generalized cross-validation (GCV) method. For an efficient smoothing computation, discrete cosine transform was employed to solve the system of HASM-SM and to estimate the minimum GCV score, simultaneously. Two examples including a numerical test and a real-world example were employed to compare the smoothing ability of HASM-SM with that of GCV thin plate smoothing spline (TPS) and kriging. The numerical test indicated that the minimum GCV HASM-SM is averagely more accurate than TPS and kriging for noisy surface smoothing. The real-world example of smoothing a lidar-derived Digital Elevation Model (DEM) showed that HASM-SM has an obvious smoothing effect, which is on a par with TPS. In conclusion, HASM-SM provides an efficient tool for filtering noises in grid-based surfaces like remote sensing–derived images and DEMs.

A high-accuracy method for filling voids and its verification

A new method for filling voids is developed by improving the approach to high- accuracy surface modelling (HASM), which is based on the first fundamental coefficients and the second fundamental coefficients of surfaces. The first fundamental coefficients are used to calculate the lengths of curves, angles of tangent vectors, areas of regions and geodesics on the surface. The second fundamental coefficients reflect the local warping of the surface, namely its deviation from the tangent plane at the point under consideration, which can be observed from outside the Earth. Nine regions with different landform complexities in hilly, plateau and mountainous areas are selected for testing the performance of HASM by comparing those ones of the classic methods such as triangulated irregular network (TIN), inverse distance weighted interpolation (IDW), advanced Spline method (ANUDEM), Spline and Kriging. The results demonstrated that the HASM void filling always has the highest accuracy regardless of the landform complexity, void area and auxiliary data.

Surface modeling of DEMs based on a sequential adjustment method

A sequential adjustment (SA) method is employed to decrease the computational cost of high-accuracy surface modeling (HASM), and the SA of HASM (HASM-SA) is being developed. A mathematical surface was used to comparatively analyze the computing speed of SA and the classical iterative solvers provided by MATLAB 7.7 for solving the system of linear equations of HASM. Results indicate that SA is much faster than the classical iterative solvers. The computing time of HASM-SA is determined by not only the total number of grid cells but also the number of sampling points in the computational domain. A real-world example of surface modeling of digital elevation models (DEMs) with various resolutions shows that HASM-SA is averagely more accurate and much faster than the commonly used interpolation methods, such as inverse distance weighting (IDW), kriging, and three versions of spline, namely regularized spline (RSpline), thin-plate spline (TPS), and ANUDEM in terms of root mean square error (RMSE), mean absolute error (MAE), and mean error (ME). In particular, the ME of HASM-SA at different spatial resolutions is averagely smaller than those of IDW, kriging, RSpline, TPS, and ANUDEM by 85%, 83%, 83%, 53%, and 19%, respectively. The high speed and high accuracy of HASM-SA can be due to the absence of matrix inversion computation, combined with the perfect fundamental theorem of HASM.

A robust weighted least squares support vector regression based on least trimmed squares

In order to improve the robustness of the classcial LSSVM when dealing with sample points in the presence of outliers, we have developed a robust weighted LSSVM (reweighted LSSVM) based on the least trimmed squares technique (LTS). The procedure of the reweighted LSSVM includes two stages, respectively used to increase the robustness and statistical efficiency of the estimator. In the first stage, LTS-based LSSVM (LSSVM-LTS) with C-steps was adopted to obtain robust simulation results at the cost of losing statistical efficiency to some extent. Thus, in the second stage, the results computed in the first stage were optimized with a weighted LSSVM to improve efficiency. Two groups of examples including numerical tests and real-world benchmark examples were respectively employed to compare the robustness of the reweighted LSSVM with those of the classical LSSVM, the weighted LSSVM and LSSVM-LTS. Numerical tests indicate that the reweighted LSSVM is comparable to the weighted LSSVM, and more accurate than the classical LSSVM and LSSVM-LTS when the contaminating proportion is small (i.e. 0.1 and 0.2), whereas with the increase of contaminating proportion, the reweighted LSSVM performs much better than other methods. The real-world exmaple of regressing seven benchmark datasets demonstrates that the reweighted LSSVM is always more accurate than other versions of LSSVM. In conclusion, the newly developed method can be considered as an alternative to function estimation, especially for sample points in the presence of outliers. A robust weighted LSSVM was presented.LTS-based LSSVM with C-steps was used to obtain robust results.A reweighted LSSVM was adopted to improve efficiency.The new estimator was validated with two groups of examples.

Sensitivity studies of a high accuracy surface modeling method

The sensitivities of the initial value and the sampling information to the accuracy of a high accuracy surface modeling (HASM) are investigated and the implementations of this new modeling method are modified and enhanced. Based on the fundamental theorem of surface theory, HASM is developed to correct the error produced in geographical information system and ecological modeling process. However, the earlier version of HASM is theoretically incomplete and its initial value must be produced by other surface modeling methods, such as spline, which limit its promotion. In other words, we must use other interpolators to drive HASM. According to the fundamental theorem of surface theory, we modify HASM, namely HASM.MOD, by adding another important nonlinear equation to make it independent of other methods and, at the same time, have a complete and solid theory foundation. Two mathematic surfaces and monthly mean temperature of 1951–2010 are used to validate the effectiveness of the new method. Experiments show that the modified version of HASM is insensitive to the selection of initial value which is particular important for HASM. We analyze the sensitivities of sampling error and sampling ratio to the simulation accuracy of HASM.MOD. It is found that sampling information plays an important role in the simulation accuracy of HASM.MOD. Another feature of the modified version of HASM is that it is theoretically perfect as it considers the third equation of the surface theory which reflects the local warping of the surface. The modified HASM may be useful with a wide range of spatial interpolation as it would no longer rely on other interpolation methods.

A robust estimator for the accuracy assessment of remote-sensing-derived DEMs

A robust estimator based on the M-estimation principle (REMP) has been developed for digital elevation model (DEM) accuracy assessment. Adaptive weights were employed to respond to a broad class of DEM error distributions, and an iterative procedure in terms of REMP starting from robust initial estimates with a high breakdown point was introduced. Original DEMs with the resolution of 2 m were obtained by means of light detection and Ranging (LiDAR) from two study sites. DEM errors in each study site were, respectively, calculated based on 100 checkpoints captured by real-time kinematic (RTK) in terms of stratified random sampling strategy. Each group of DEM errors was, respectively, contaminated by five groups of outliers from different distributions. Thus, ten groups of simulated DEM errors were employed to comparatively assess the estimation accuracies of REMP and the classical estimators. The results indicated that under the non-normal distribution of DEM errors, the classical non-robust estimators are seriously influenced by the non-normality. Some robust estimators such as 10%-trimmed or Winsorized mean and normalized median absolute deviation (MADN) are not very robust to resist the influence of outliers. REMP, slightly affected by the non-normal distribution of DEM errors, is more accurate than the classical estimators. The robust methodology can adapt to the DEMs, especially the ones derived from remote sensing such as LiDAR or digital photogrammetry in non-open terrain.

A review of recent developments in HASM

Ground observation is able to obtain highly accurate data with high temporal resolution at observation points, but these observation points are too sparsely to satisfy the application requirements at regional scale. Satellite remote sensing can frequently supply spatially continuous information on earth surface, which is impossible from ground-based investigations, but remote sensing description is not able to directly obtain process parameters. In fact, in terms of fundamental theorem of surfaces, a surface is uniquely defined by the first fundamental coefficients, about the details of the surface observed when we stay on the surface, and the second fundamental coefficients, the change of the surface observed from outside the surface. A method for high accuracy surface modeling (HASM) has been developed initiatively to find solutions for error problem and slow-speed problem of earth surface modeling since 1986. HASM takes global approximate information (e.g., remote sensing images or model simulation results) as its driving field and local accurate information (e.g., ground observation data and/or sampling data) as its optimum control constraints. Its output satisfies the iteration stopping criterion which is determined by application requirement for accuracy. This paper reviews problems to be solved in every development stage and applications of HASM.