Master s Thesis: ANAMELECHI, FALASY EBERE Analysis of a Raster DEM Creation for a Farm Management Information System based on GNSS and Total Station Coordinates Duration of the Thesis: 6 Months Completion Date: April, 2014 TUTORS 1: Dr.-Ing. Martin Metzner (STUTTGART) TUTORS 2: M.Sc. C.Jaeger-Hansen (HOHENHEIM) EXAMINER 1: Dr.-Ing. Martin Metzner (STUTTGART) EXAMINER 2: Prof. Dr.-Ing. Hans W. Griepentrog (HOHENHEIM) Analysis of a Raster DEM Creation for a Farm Management Information System based on GNSS and Total Station Coordinates In today s world of precision farming, more machines and agricultural field tractors are greatly relying on the set-up of autonomous platforms for navigation and localization in farmlands and orchards. And with great advancements made in these machineries, the need for structured agricultural fields to be better modeled and optimized to provide continuous surface (x, y, and z) coordinates and datasets to these machines during field missions becomes a necessity. Moreover, with the unreliability of GNSS receivers under tree leaves and the headland turning problems encountered during field operations, it becomes more necessary to have a better picture and terrain pattern-analysis of the field area in three-dimension, in order to have an improved defined mission and maneuver for these agricultural field tractors by use of the terrain pattern. However, this problem can be addressed if a contour profile and terrain elevation dataset is designed and modelled for the agricultural field. Thus, is the need to ascertain the performances and reliability of the different data sources and techniques used for the creation of raster-based DEMs that solves the visibility and topographic problems of the field, unlike the graph map. This master thesis focuses on the analysis of a raster DEM creation for a farm management information system based on two data sources (1) GNSS and, (2) Total station coordinates. In order to create these elevation models, raw geometric data was collected in real time from the field using a remotely controlled technology that was mounted on an agricultural field tractor, comprising of GNSS and also through direct field survey using a Total station. With these elevation details of the soil, the data was analyzed and evaluated using various interpolation methods to create a contour profile and a digital elevation model for the field in the software SURFER 10. With the elevation model filtered and optimized, the geo-referenced field could thus provide continues surface (x, y, z) coordinates which can be used in a farm management information systems (FMIS) and other planning purposes like contour farming, soil erosion prediction and seed routing of crops. The site for the study is the Meiereihof test-field at Schwerzstraße, University of Hohenheim. The study site has an elevation range from 392 m to 402 m and thus, a vertical relief of 10 m. It has an average slope of 3.67 and about 14 000 square meters. The tracked field for the elevation from both data sources is shown below: 1
Figure 1: Tracking from GNSS Figure 2: Tracking from Total Station By analyzing the 3D elevation tracked data used in the generation of Digital Elevation Models (DEMs) from both data source, this study investigated the reliability of its measurement methods and interpolation algorithms. Moreover, the analysis tried solving the spatial disparity for every grid post so that the modelled field corresponds to the agricultural field. Figures 3 and 4 highlight the modelled height differences between these interpolators. Figure 3: Elevation Differences between Interpolators 2
Figure 4: Mean Elevation between Interpolators Tables 1 and 2 present the statistical results of the different interpolators for the two data sources. Table 1: Interpolation Statistics for GNSS GNSS Evaluation Statistics Types of Interpolator Total Nodes Used Nodes Mean Elevation [m] Vertical Relief [m] Accuracy Measure [m] Kriging 6800 6800 397.333 11.530 0.042 TIN Linear 6800 3892 398.768 11.453 0.047 Natural Neighbors 6800 3789 398.796 11.309 0.046 Table 2: Interpolation Statistics for Total Station Total Station Evaluation Statistics Types of Interpolator Total Nodes Used Nodes Mean Elevation [m] Vertical Relief [m] Accuracy Measure [m] Kriging 6800 6800 395.311 9.445 0.037 TIN Linear 6800 3688 396.410 9.371 0.040 Natural Neighbors 6800 3591 396.425 9.192 0.039 Comparing the interpolation statistics of both tables above for the three interpolators used, the following can be deduced: 3
In examining the results of the total grid nodes used, it notable that the kriging algorithm generated extra surfaces for both data sources by extrapolating those surfaces from the field data obtained (see Figures 6 and 7). This is a disadvantage from the interpolator s artefact since undesired surfaces (marked black) were generated as though it was part of the tracked field due to using all the grid nodes which is 6800. However, for both the TIN Linear and Natural Neighbors algorithms, about half the total grid nodes were used for the generation of the elevation model. This difference is accounted for in both algorithms artefacts since in forming a network of triangles, the TIN Linear and Natural neighbors uses elevation points from the contour lines. And in areas where the three chosen points lie on the same contour lines, the point is reported flat and thus deleted from the total grid nodes. Nevertheless, this action is a disadvantage since bench mark points could be lost from such algorithm. Furthermore, the vertical relief for the kriging in the Total station statistics performed better than the rest two interpolators since the value is closer to the vertical relief of the field (see Figure 1). However, the GNSS plots were far away from the threshold. Also, the kriging in GNSS statistics returns a closer result to the field s threshold or mean elevation from the control point measurements than those of the rest two interpolators (see Figure 2). Moreover, looking at the different contour images generated for each of the interpolation routines reveals that there are common trouble zones for interpolation across each method especially along their edges (Figures 6 to 11). However, the different accuracy measures of both tables indicate that Kriging out-performed the rest two interpolation methods, TIN Linear and Natural Neighbors, statistically (see Tables 1 and 2). However, the Natural Neighbors has the best representation spatially. This shows that its method is much more capable of dealing with a sparse set of points to interpolate. This advantage could be because the tessellation underlying the interpolation routine extends a little beyond the boundary of the eventual field, allowing for better interpolation around the boundary of the test-field relative to the other methods. (Harman & Johns., 2012). But since the statistical difference between the Kriging and Natural Neighbors is in the millimeter range, it can then be deduced that both statistically and spatially, the Natural Neighbors interpolator is to be preferred than the rest two interpolation methods. Although the overall system setup for the Total station seems tedious, its accuracy measures for all three interpolators are better than those of the GNSS. Moreover, the uncertainty of digital elevation models resulting from ground surveying techniques is related to two aspects: the sampling and measurement error, and the interpolation process. This assessment certifies the quality of a DEM. Table 3: Error Statistics for the two Elevation data sets based on a scale of 1:10000 GNSS Total Station No. of Points 12447 8300 Grid Spacing 0,5 [m] 0,5 [m] Grid Size 6800 6800 Propagated Height RMSE 0,071 [m] 0,065 [m] Table 3 shows that the DEM created from the Total station using the SURFER 10 software has the better accuracy because of its lower value for RMSE which also reveals the closeness of the interpolated surface to the reality. Furthermore, two major terrain attributes (slope and aspect) were estimated from the elevation models generated from the two data sources under investigation. These parameters constitute a topographical and hydrological characterization of the study area, which exposes the soil slope and possible water-log spots on the agricultural 4
field orchard respectively. Figure 5 shows the cumulative distribution of slope values within the study area for the two elevation data. Although the trends are quite similar (note the same near absence of slopes around the 8000 points for the two elevations), it is clear that the coarser resolutions show a larger contribution of lower slope angles and fewer short steep slopes. It should however be noted that the slope of an elevation model is an elevation model itself. Moreover, in Figure 5, the slope map statistics (stochastic though), derived from the Total station have the lowest minimum and maximum slope value which indicates that the Total station derived terrain is flatter, while the GNSS elevation data with the highest minimum and maximum slope values shows that the terrain is steeper but farther to the reference terrain. This is because the lower the slope value, the flatter the terrain; the higher the slope value, the steeper the terrain. Furthermore, the aspect identifies the steepest down slope direction at a point in an elevation model of the earth surface. Figures 6 to 11 below, also shows the visual Aspect map (slope s direction) statistics from the various surface maps. The elevation values from the elevation bar indicate that the steepest down slope is in the direction of North-East. Figure 5: Elevation Spots data from both data sources Also, from the figure above, it is observed that the GNSS data and Total station data do not overlap each other at respective projected surfaces. This variation of about 3 meters between both elevation data sources could have being as a result of the height off-set from the data platforms on the tractor and the configuration set-up for the GNSS height which was set at a plus 3 meter height for every elevation point. The effect of this height off-set is seen affecting values of the mean elevation and projected vertical reliefs especially from the GNSS measurement data. Moreover, Figure 5 also indicate that there were points from the tracking data sets were the Total station lost sight of the prism and this affected the full area coverage from the 3D elevation points. Thus, from this it could be 5
deduced that the GNSS data is more covering over the tracked field than that the Total station (see Figures 1 and 2). For the Meiereihof test-field, the following surface views were generated using all three interpolators on both data sources. I. Kriging Surface Maps for both data sources Figure 6: Surface Map from GNSS data (kriging) Figure 7: Surface Map from Total station data (kriging) 6
II. TIN Linear Surface Maps for both data sources Figure 8: Surface Map from GNSS data (TIN Linear) Figure 9: Surface Map from Total station data (TIN Linear) 7
III. Natural Neighbors Surface Maps for both data sources Figure 10: Surface Map from GNSS data (Natural Neighbors) Figure 11: Surface Map from Total station data (Natural Neighbors) 8
Visual inspection and examination of surface maps (see Figures 6 to 11) obtained from the extracted elevations was used to evaluate qualitatively the various interpolation models when compared to that obtained from the ground survey of the two data sources. Comparing visually between the various surface map representations, Figures 6 and 7 displays a different spatial pattern from the rest surface maps (Figures 8 to 11) which may be as a result of the artefacts from the interpolator. For instance, the area marked black in the kriging interpolators gives a poor surface representation of the test site since undesired surfaces were generated as though it was part of the tracked field. Moreover, the visual impression was confirmed by comparing the interpolation models (Figures 6 to 11) and the various statistical measures of Tables 1 and 2. The interpolation models for the specific upslope area in Figures 6 to 11 clearly showed the difference in distribution among the interpolators with narrowing and increasingly skewed distribution with decreasing information content and increasing grid size. Conclusively, it was observed that the reference DEM from ground survey using total station proves to be a very efficient method for generating DEMs but requires much field work in capturing detailed terrain data. Also, the DEM derived from GNSS points does perform well in obtaining DEM data but is slightly reliable and also gives a good result but its quality compared to the Total station is poor relatively (see Tables 1 to 3). The significance of these findings can then be integrated and utilized in the farm management information systems and for other planning purposes in the catchment area of the test-field like in seed routing, contour farming and even creating a field yield frequency map as a post-harvest operation. Based on the findings of this study, it is to be recommended that a more extensive set of comparison between these interpolation models be carried out to meet and improve the respective needs of DEM applications since variation in interpolation parameters may significantly improve or worsen the DEM accuracy. Also, since the accuracy of spatial interpolation of elevations is subject to input data point density and distribution, future analysis should also be extended into the impact of different data collection patterns (e.g., random vs. systematic; significant points vs. contouring). Equally important is the problem of how to take the complexity of terrain into account to determine point density and distribution. Comparative studies may be conducted by using different input data density for different degrees of roughness in a surface or by taking break lines. Furthermore, further studies on the uncertainty of measurement data should be investigated upon so as to enhance DEM reliability. Moreover, it is important to point out that the accuracy of elevation data sets for generation of DEMs be properly understood and defined before they are utilized in varying applications. Thus, in view of the results obtained herein there is the need to validate all available global elevation data set, in order to ascertain their suitability or otherwise. 9