Due to the large number of technological developments in recent years, UAS systems are now used for monitoring purposes and in projects with high precision demand, such as 3D model-based creation of dams, reservoirs, historical monuments etc. These unmanned systems are usually equipped with an automatic pilot device and a digital camera (photo/video, multispectral, Near Infrared etc.), of which the lens has distortions; but this can be determined in a calibration process. Currently, a method of “self-calibration” is used for the calibration of the digital cameras mounted on UASs, but, by using the method of calibration based on a 3D calibration object, the accuracy is improved in comparison with other methods. Thus, this paper has the objective of establishing a 3D calibration field for the digital cameras mounted on UASs in terms of accuracy and robustness, being the largest reported publication to date. In order to test the proposed calibration field, a digital camera mounted on a low-cost UAS was calibrated at three different heights: 23 m, 28 m, and 35 m, using two configurations for image acquisition. Then, a comparison was made between the residuals obtained for a number of 100 Check Points (CPs) using self-calibration and test-field calibration, while the number of Ground Control Points (GCPs) variedand the heights were interchanged. Additionally, the parameters where tested on an oblique flight done 2 years before calibration, in manual mode at a medium altitude of 28 m height. For all tests done in the case of the double grid nadiral flight, the parameters calculated with the proposed 3D field improved the results by more than 50% when using the optimum and a large number of GCPs, and in all analyzed cases with 75% to 95% when using a minimum of 3 GCP. In this context, it is necessary to conduct accurate calibration in order to increase the accuracy of the UAS projects, and also to reduce field measurements.

We studied the applicability of point clouds derived from tri-stereo satellite imagery for semantic segmentation for generalized sparse convolutional neural networks by the example of an Austrian study area. We examined, in particular, if the distorted geometric information, in addition to color, influences the performance of segmenting clutter, roads, buildings, trees, and vehicles. In this regard, we trained a fully convolutional neural network that uses generalized sparse convolution one time solely on 3D geometric information (i.e., 3D point cloud derived by dense image matching), and twice on 3D geometric as well as color information. In the first experiment, we did not use class weights, whereas in the second we did. We compared the results with a fully convolutional neural network that was trained on a 2D orthophoto, and a decision tree that was once trained on hand-crafted 3D geometric features, and once trained on hand-crafted 3D geometric as well as color features. The decision tree using hand-crafted features has been successfully applied to aerial laser scanning data in the literature. Hence, we compared our main interest of study, a representation learning technique, with another representation learning technique, and a non-representation learning technique. Our study area is located in Waldviertel, a region in Lower Austria. The territory is a hilly region covered mainly by forests, agriculture, and grasslands. Our classes of interest are heavily unbalanced. However, we did not use any data augmentation techniques to counter overfitting. For our study area, we reported that geometric and color information only improves the performance of the Generalized Sparse Convolutional Neural Network (GSCNN) on the dominant class, which leads to a higher overall performance in our case. We also found that training the network with median class weighting partially reverts the effects of adding color. The network also started to learn the classes with lower occurrences. The fully convolutional neural network that was trained on the 2D orthophoto generally outperforms the other two with a kappa score of over 90% and an average per class accuracy of 61%. However, the decision tree trained on colors and hand-crafted geometric features has a 2% higher accuracy for roads.

High-resolution stereo and multi-view imagery are used for digital surface model (DSM) derivation over large areas for numerous applications in topography, cartography, geomorphology, and 3D surface modelling. Dense image matching is a key component in 3D reconstruction and mapping, although the 3D reconstruction process encounters difficulties for water surfaces, areas with no texture or with a repetitive pattern appearance in the images, and for very small objects. This study investigates the capabilities and limitations of space-borne very high resolution imagery, specifically Pléiades (0.70 m) and WorldView-3 (0.31 m) imagery, with respect to the automatic point cloud reconstruction of small isolated objects. For this purpose, single buildings, vehicles, and trees were analyzed. The main focus is to quantify their detectability in the photogrammetrically-derived DSMs by estimating their heights as a function of object type and size. The estimated height was investigated with respect to the following parameters: building length and width, vehicle length and width, and tree crown diameter. Manually measured object heights from the oriented images were used as a reference. We demonstrate that the DSM-based estimated height of a single object strongly depends on its size, and we quantify this effect. Starting from very small objects, which are not elevated against their surroundings, and ending with large objects, we obtained a gradual increase of the relative heights. For small vehicles, buildings, and trees (lengths <7 pixels, crown diameters <4 pixels), the Pléiades-derived DSM showed less than 20% or none of the actual object’s height. For large vehicles, buildings, and trees (lengths >14 pixels, crown diameters >7 pixels), the estimated heights were higher than 60% of the real values. In the case of the WorldView-3 derived DSM, the estimated height of small vehicles, buildings, and trees (lengths <16 pixels, crown diameters <8 pixels) was less than 50% of their actual height, whereas larger objects (lengths >33 pixels, crown diameters >16 pixels) were reconstructed at more than 90% in height.

No abstract