Accurate geo-referencing and DSM generation with HRSI

Earth-observation techniques attract currently a substantial amount of interest and open fresh research and business opportunities for the geospatial community. They offer nowadays a broad spectrum of diverse platforms, sensors and products. UAV systems are paving their way into novel applications. The development of large format digital aerial camera systems has triggered a renewed interest in aerial photogrammetry. This is all in line with a general trend of turning the attention “back-toearth” in science and development.

With the advent of very high-resolution satellite imagers (VHRSI) this development is strongly supported and enforced. We are witnessing presently worldwide many activities in space, fueled by environmental, resource management, security and also military concerns. Many countries are getting involved, setting up their own earth-observation programs, others are strengthening their existing ones. With WorldView-1 and GeoEye-1 we have reached the 50 cm footprint level and the development will not stop here.

VHRSI sensors differ from each other. They are characterized by a number of geometric, radiometric, spectral and operational specifications. This raises questions concerning the appropriate use of those devices. 3D mapping is a hot topic under discussion. But whatever application may be envisioned, at the beginning of every value-adding procedure stands the problem of geo-referencing and (in many cases) DTM generation.

Figure 1: Workflow of the SAT -PP software system

Processing of VHRSI with SAT -PP

All high-resolution satellite sensors acquire panchromatic and/or multispectral images in pushbroom mode for photogrammetric and remote sensing applications. They use Linear Array CCD technology for image sensing and are equipped with high quality orbital position and attitude determination devices like GPS, IMU systems and/or star-trackers.

For the full exploitation of the potential of the Linear Array CCD sensors’ data, the “classical” satellite image analysis methods must be extended in order to describe the imaging geometry correctly, which is characterized by nearly parallel projection in alongtrack direction and perspective projection in cross-track direction.

We have developed a full suite of new algorithms and the related software package SAT-PP (Satellite Image Precision Processing) for the accurate processing of high-resolution satellite image data. The software can accommodate images from IKONOS, QuickBird, SPOT5 HRG/HRS, Cartosat-1, ALOS/PRISM, WorldView-1 and sensors of similar type to be expected in the future.

The software package SAT-PP consists of the following components (Figure 1):

(a) User interface for project and data management, image format conversion and pre-processing (with an edge-preserving smoothing filter) and image display / roaming in mono and stereo modes

(b) Sensor and trajectory models (rigorous and generic ones such as the rational polynomial function model – RPF), designed for the high-resolution Linear Array sensor geometry

(c) Orientation of single stereo models and triangulation of larger units.

On-line quality control and error analysis via interaction with graphics elements. Ground control point (GCP) and tie point measurement in manual and semi-automated modes

(d) Derivation of quasi-epipolar images for stereo mapping and feature collection

Figure 2: DSM of Piemont testfield, derived from an ALOS/PRISM triplet analyses 0.5-0.8 pixel planimetric with SAT -PP

(e) Automated generation of Digital Surface Models (DSMs) by using a precise and robust image matching approach, combining area-based, feature-based and relational matching techniques. Stereoscopic checking of the automatically matched features

(f) Generation of orthorectified images

(g) Mono-plotting functions with existing or automatically derived DTMs. Stereoscopic measurement and collection of objects with particular emphasis on 3D city modeling by using the semiautomatic 3D modeling software CyberCity ModelerTM

(h) Pan-sharpened image generation to enhance the visual information of multispectral imagery by fusing it with the detailed spatial information of panchromatic imagery. Fully automated sub-pixel image registration between multispectral and panchromatic imagery

Geo-referencing

The increased spatial resolution of the HRSI and the stereo capabilities demand accurate methods for geo-referencing. The methods used for geo-referencing can be categorized in two main groups: rigorous (“physical”) and generic models (also called “parameterized” and “non-parameterized” models, respectively). These approaches have several advantages and disadvantages against each other. In order to make full use of the sensors’ geometrical potential in terms of geolocation accuracy, a rigorous modeling of the sensor and trajectory geometry is to be favoured. Physical sensor models require precise knowledge of interior and exterior orientation parameters. The information may be acquired through laboratory calibration and/or on-thejob self-calibration /orientation. On the contrary, the generic models, such as the Rational Polynomial Function (RPF) Model, Direct Linear Transform, 2D/3D affine transformation, etc. do not require a-priori knowledge of the sensor location, altitude, and interior geometry for the user. They basically provide an approximate solution and the model parameters do not correspond fully to the physical state of the sensor. However, these models are easier to implement, generic for various sensors, easy to transfer from one software package to another, and more suitable for inexperienced users. However, depending on the procedures used for the generation of RPF-parameters there is usually the need to post-correct the orientation with some GCPs and low-order polynomials. This technique may also be interpreted as self-calibration (of RPF models).

The image processing level, which is often denoted in product levels (1A, 1B, etc.), should be considered when choosing a particular geo-referencing technique. For example, the rigorous models should be applied to geometrically raw data. A number of pre-processing methods are always applied in order to correct the radiometric and geometric systematic errors, at least partially. The raw (original from the acquisition) satellite images are usually processed internally and not provided to the users. A high processing level increases the data costs inevitably. The user needs depend on the required geometric accuracy and intended applications. Advanced users can handle raw data for sophisticated analysis, while end-users and non-photogrammetrists prefer the high-level processed and already geo-referenced images.

Recently, as Member of the Calibration/ Validation Team of JAXA’s ALOS/ PRISM mission we have validated our geo-referencing and DSM generation approaches using ALOS/PRISM images over several testfields (Saitama, Okazaki, Sakurajima, all Japan; Piemont, Italy; Haiphong, Vietnam; Bern/Thun and Zurich/Winterthur, both Switzerland;, Wellington, South Africa; Adana, Turkey).

The results are very consistent over all testfields. We achieve from checkpoint analyses 0.5-0.8 pixel planimetric accuracy and 0.3-0.8 pixel in height. This corresponds quite well with earlier results which we have obtained with SPOT-5, IKONOS and Quickbird images (0.3 pixel in planimetry and 0.5 pixel in height). The ALOS/PRISM results are a bit inferior, especially in planimetry, because PRISM images are suffering under low image quality.

All results are based on a sufficiently high developed sensor model and just a few (1-5) ground control points (the required number depends on the used sensor model and the number and type of self-calibration parameters).

DSM/DT M generation

This is a key issue in many applications. If produced in manual mode this does not constitute a problem, it only needs time– a lot of time. Therefore automated DSM generation by image matching becomes a relevant topic. Image matching – in its essence – is still an unsolved problem. With our software SAT-PP, which includes an advanced matching module, we obtain height accuracies between 1 and 5 pixels from high-resolution satellite images, depending on the type of terrain, land cover, image texture and image quality. In the worst cases – very steep terrain with strong tree coverage and shadows in the valleys – we may have to expect an accuracy of only 5 pixels. Validating the performance of image matchers with VHRSI data requires very high accuracy DSM reference data (ideally it should be accurate to at least 1/3 of a pixel). We obtain this from laserscans and aerial images.

While the RMS errors in such tests show usually quite good results we must note that in all these cases substantial blunders (10 times the RMSE and more) still exist in the data. This is not acceptable to the cartographer. This can only be solved by substantial and time-consuming post-editing of the DSM. Therefore, the avoidance and/or detection of blunders in the automatically generated DSM is a critical point for future research and development. The next problem we are faced with is the reduction of the DSM, produced by the image matcher, to the DTM, as represented in the landscape model. Although there are some attempts available to automatically perform the reduction, the results are not convincing, because these algorithms are purely based on geometrical considerations. What is needed however is an image or point cloud interpretation approach which lets us understand what kind of objects we are dealing with in the particular areas of images or point clouds.

3D mapping

3D mapping entails much more than just geo-referencing and DSM generation. Since it is generally assumed that highresolution satellite images with stereo capabilities provide for an interesting data source for topographic mapping we are currently doing a sequence of tests in order to find out to what extent sensors like IKONOS, WorldView-, etc. are suited for topographic mapping.

Our test with manual measurements of IKONOS stereo images for 1:25,000 topo-mapping in Switzerland revealed many problems. Small terrain features did get lost in contours and many buildings could not be interpreted and mapped appropriately. Also, the classification of roads turned out to be difficult in many cases.

Modern technologies allow for the new concept of 3D mapping, shifting the concept of map generation as the primary product to the generation of 3D digital landscape models. However, many issues of truly 3D modeling are not yet well understood.

However, one often occurring question can be answered already now:

Fully automated processing techniques are far away from delivering reliable results.