The knowledge of the ground morphology and how it changes is a fundamental information in many applications, such as urban planning and the design and the environmental impact evaluation of new infrastructures. In the past, the only instrument used by designers was cartography, with height information given by contour lines or points. With the development of technologies the situation has quickly changed and nowadays not only numerical maps are available, but also 3D digital terrain or surface models can be created. These models can be generated in many ways, thanks to the quality achieved in the acquisition of 3D information using technologies such as LIDAR, inSAR, GPS, photogrammetry and remote sensing, even if with a lot of differences in terms of precision, costs and time needed for models creation. Among all these techniques, the use of remote sensing is increasing rapidly for all the cartographic applications, because it could bring advantages in terms of costs reduction, fast data acquisition and elaboration also for big areas. Everything started in 1999, when IKONOS II, the first civil satellite offering a spatial resolution of 1m, was launched. Since then other high resolution satellites were launched, and now they ensure a GSD of less than 1m, a revisiting time of the same area of just few days, multispectral data and along-track stereo acquisition (with imagery being processed using the well known photogrammetric algorithm). However, the possibility of using high resolution satellite imagery strictly depends on several factors: the aim of the work (in terms of accuracy needed), sensor’s characteristics (geometric and radiometric resolution), available products, software used to perform the elaboration. Firstly, images must be oriented and corrected both geometrically and radiometrically to remove all the distortions related to the image acquisition process, to the acquisition system (imaging sensor optical-geometric characteristics, orbital characteristics, attitude parameters, platform position and movement during the acquisition in respect to the Earth’s one) and to the atmosphere refraction. As for the geometric correction, a new image is created, which has scale and projective properties typical of a cartographic representation, in order to assure the metric accuracy of all the data contained. So, two different types of orientation models are generally adopted: the rigorous or physical models, which are linked to the physical-geometrical image acquisition process (for each kind of satellite) and are based on the collinearity equations; the polynomial and rational methods, based on transformations which link terrain and image coordinates without any information about the characteristics of the used satellite. A rigorous model is based on the collinearity equations and reconstructs the orbital segment during image acquisition through the knowledge of the acquisition mode, orbital parameters, attitude angles, interior orientation and self-calibration parameters, atmospheric refraction: all the approximate values are known (some of them are in the metadata file) and their correction could be estimated using the Least Square Method. Rational methods were developed principally because sensor managing companies didn’t reveal technical characteristics of the platforms, so it was impossible to implement a correct rigorous model. Nowadays they approximate very well physical models, so very often also those companies use this solution to give information about sensor geometry without releasing neither the internal orientation parameter nor the physical features related to image processing. The most important method is the Rational Polynomial Function (RPF), which is based on a ratio of two polynomials of various order to model the relation between ground and image coordinates, so representing a generic formulation of the rigorous collinearity equations. The link is created by the Rational Polynomial Coefficients given in the metadata file or calculated directly by the final user thanks to some known ground points (even if this solution shouldn’t be used if high accuracy is needed). However, the knowledge of the coordinate of some ground control points is fundamental both with rigorous and rational models: after the orientation is performed, an accuracy assessment checking the Root Mean Square Error of this dataset must always be performed. As far as radiometric correction is concerned, it is useful to solve all the distortions in the acquired signal due both to sensor’s improper functioning (detectors’ relative brightness, striping, no data pixels and lines) and to external conditions (scene illumination and atmospheric conditions). Next step is related to the height information extraction, which is usually done with automatic procedures (image matching): analyzing all the differences between possible targets and background information, homologous points in the two images are recognized. Different elements could be used as search parameter: the radiance value of each image pixel, geometric features, spatial relations between various kind of features in the imagery. The first case is called Area Based Matching (ABM): a template containing the radiance values of a small area of one image is compared with the values of a moving window in the other image until the highest correlation is found. Region growing is very similar, but it starts the matching from some seed points well distributed on the imagery. Only using also a Least Square Matching it is possible to achieve sub-pixel accuracy, because this technique takes into account also the radiometric and geometrical distortions between template and search window, which always afflicts the data due to the different point of view of the imagery. The second possibility is the Feature Based Matching (FBM): various mathematical operator recognize simple features (such as corners or edges) and compare the created dataset among imagery. The last method is the relational matching, which is the less accurate one: it considers at the same time different relations between all the features in the images (geometrical ones, combinations of simple features, … ). Once the matching is finished, it is possible to calculate 3D coordinates of the recognized points using the parameters of the orientation step; finally, the model is extracted generating a regular grid which can then be used as input data for many applications. In fact, digital elevation models can be generated in three different ways: irregular points, which don’t contain any height information in the areas among them; TIN, Triangulated Irregular Network: the irregular points are liked using triangular faces which create a continuous surface; matrix models (GRID): the space is divided into a regular grid, so for each cell the height value is calculated interpolating the irregular dataset. The GRID model is typical when using imagery as input data and it’s also the most used for digital representation; the most important parameter is the grid size, because its value depends both on input data spatial resolution and on the detail needed by the final model. After the whole process is finished, the extracted model must be validate in order to define its accuracy and so determine its possible field of application (or if it is as precise as requested by the customer). Quality analysis must be performed together with comparisons with check datasets (having accuracy at least of one order better than the one expected for the model) to evaluate standard deviation on residuals between the model and the check data (control points, reference model, breaklines). In this thesis the entire generation process (from imagery orientation to DSM extraction) to generate a very accurate model was evaluated; both commercial and scientific software were tested in order to optimize human, economic and computational resources. Two different in-track high resolution satellite stereo pairs were processed: the first one was acquired over the area of Ferrara in 2004 by IKONOS II (spatial panchromatic resolution of 1m), the other covers the area of Argenta (FE) and it was taken in 2010 by GeoEye-1 (spatial panchromatic resolution of 0.5m). In both cases not only the panchromatic imagery (generally the only input data for DSMs generation) were used but also the multispectral ones (on the contrary generally useful to generate thematic maps). During the analysis different software have been tested, because with available commercial ones good results are achievable concerning stereo orientation, while some drawbacks are inherent to image matching: the commercial software Orthoengine (PCI Geomatics) and the scientific one SISAR (Software per Immagini Satellitari ad Alta Risoluzione) developed by Prof M. Crespi and the team of Geodesy and Geomatic Area - La Sapienza University of Rome and designed for the orientation of high resolution satellite imagery. Some tests inherent to IKONOS II imagery were performed also with ENVI 4.6.1(which gave unsatisfactory results) and the scientific one DPCOR, developed by prof. K. Jacobsen and his team at Leibniz University of Hannover. As far as the orientation is concerned, results with rigorous models show that accuracy of about 0.50.8 pixel in East and North direction and of about 0.81 pixel in Up direction is achievable; SISAR seems to be more stable varying GCPs number and distribution (620 GCPs are sufficient for an area of about 100Km2). The most important thing is the correct identification of the GCPs on imagery, because both software manifest higher errors during test run with point not well recognized (small errors of about 0.50.75 pixel for some points). Test performed with SISAR using rational functions show comparable results; moreover, RPCs generation and refinement don’t give better results, confirming a good pre-processing of the sensor managing companies while generating the RPCs. Finally, Tie Points recognition don’t improve the orientation accuracy. After that, image matching and DSM extraction were performed in small homogeneous areas with scientific software, while with OrthoEngine a full extent model was created. With regard to IKONOS II stereo pair, commercial software had some problems: matching wasn’t satisfactory in dense urban areas and many pixel had wrong height values (peaks and holes far from real ones) due to noise and difficulties during correlation step. Further tests were run with multispectral bands: NIR imagery gave good results, so using GIS software no data and noisy zones of the panchromatic models were edited and corrected merging data with NIR model. Derived DSMs generated with scientific software were less noisy, even if grid extraction is too approximate because it just creates a TIN. Accuracy evaluation was performed both with a set of points collected by kinematic and stop&go GNSS survey with 3D accuracy of 0.3m and with a reference grid of 2x2m with accuracy of 0.6m derived by a photogrammetric flight: calculated RMSEs show gaps of about 2 pixel in rural areas and 4 pixel in urban and vegetation zones. Even if features and infrastructures are well identified, these models aren’t suitable for high accuracy applications. With GeoEye-1 stereo pair only 40% of the whole scene was matched: built-up areas and streets were recognized whereas the major part of the fields were too homogeneous, because in march (when the imagery were acquired) there were just uncultivated fields and wheat plant were too small. This problem was much more important with OrthoEngine, so manual and automatic procedures to solve it were implemented, also using GIS software. Accuracy evaluation, performed as for IKONOS II stereo pair, show errors of about 1.5 pixel in rural area and slightly better than 4 pixel in urban one. The results are comparable to previous ones in terms of pixel residuals, showing an accuracy limit of models derived from high resolution satellite imagery; the great advantage is so represented by pixel size, which is halved in few years let DSMs to reach accuracy of about 1 meter. One of the main goals of this study was the creation of models to be used in the forecast and management of environmental emergencies, because the importance of this field of application is increasing day by day. In fact, the forecasting hydraulic models, which study the outflow of water after, for example, river floods or downpours, need as preliminary input a very good model of the studied area. The DSMs used for this purpose need to reach a very high precision, especially with regards to the correct identification of all the man-made and natural elements, like riverbanks or roads in relief, which may interfere with the outflow of water (deviating or stopping it) or, on the contrary, of all the crosses, which are a preferential way to water’s flow. Furthermore, this demand becomes more important for flat areas, because little errors in height definition could cause important variations in the identification of flooded areas. According to CISIS’s technical guidelines about elevation models, results obtained with both the stereo pairs show that the extracted DSMs aren’t suitable for this purpose, because height values must have an accuracy better than 0.5m (the DSMs extracted are at level 3, while they should be at least at 5). With regard to data derived from GeoEye-1 imagery, they seems to be useful, even if merged with other surveying technologies to acquire in a better and more precise way the characteristics of all the barriers and the crosses. This fact was demonstrated also with simulations run on the same area by the technician of the land reclamation consortium of Ferrara (Consorzio di Bonifica Pianura di Ferrara): in very flat areas (such as the one investigated), in order to obtain a correct identification of flooded areas (both with geographic position and temporal sequence), the correct implementation of barriers and drains network is much more important than the height information of areas in the middle of the hydraulic cell. So, using scientific software it would be possible to obtain a refinement of the matching parameters and procedures in order to create accurate DSMs also starting from imagery with GSD of 0.5 meter; in any case, an accurate editing done by the user and a correct point cloud interpolation is fundamental. On the contrary, because the user can’t generally set all these parameters using commercial software, extracted DSMs should be used only for applications where lower accuracy is needed. In the first part of the thesis all the theoretical aspects are described: after some information about photogrammetry and remote sensing (Chapter 1) and the main characteristics of platforms (Chapter 2), Digital Terrain and Surface Models are discussed (Chapter 3), focusing the attention to the whole generation process performed using high resolution satellite imagery (Chapter 4). The second part presents all the procedures adopted and the results obtained both with the IKONOS II stereo pair (Chapter 5) and the Geoeye-1 imagery (Chapter 6).
Valutazione dell'accuratezza di DSMs estratti da stereocoppie satellitari ad alta risoluzione per territori a carattere fortemente pianeggiante
FURINI, Alessio
2011
Abstract
The knowledge of the ground morphology and how it changes is a fundamental information in many applications, such as urban planning and the design and the environmental impact evaluation of new infrastructures. In the past, the only instrument used by designers was cartography, with height information given by contour lines or points. With the development of technologies the situation has quickly changed and nowadays not only numerical maps are available, but also 3D digital terrain or surface models can be created. These models can be generated in many ways, thanks to the quality achieved in the acquisition of 3D information using technologies such as LIDAR, inSAR, GPS, photogrammetry and remote sensing, even if with a lot of differences in terms of precision, costs and time needed for models creation. Among all these techniques, the use of remote sensing is increasing rapidly for all the cartographic applications, because it could bring advantages in terms of costs reduction, fast data acquisition and elaboration also for big areas. Everything started in 1999, when IKONOS II, the first civil satellite offering a spatial resolution of 1m, was launched. Since then other high resolution satellites were launched, and now they ensure a GSD of less than 1m, a revisiting time of the same area of just few days, multispectral data and along-track stereo acquisition (with imagery being processed using the well known photogrammetric algorithm). However, the possibility of using high resolution satellite imagery strictly depends on several factors: the aim of the work (in terms of accuracy needed), sensor’s characteristics (geometric and radiometric resolution), available products, software used to perform the elaboration. Firstly, images must be oriented and corrected both geometrically and radiometrically to remove all the distortions related to the image acquisition process, to the acquisition system (imaging sensor optical-geometric characteristics, orbital characteristics, attitude parameters, platform position and movement during the acquisition in respect to the Earth’s one) and to the atmosphere refraction. As for the geometric correction, a new image is created, which has scale and projective properties typical of a cartographic representation, in order to assure the metric accuracy of all the data contained. So, two different types of orientation models are generally adopted: the rigorous or physical models, which are linked to the physical-geometrical image acquisition process (for each kind of satellite) and are based on the collinearity equations; the polynomial and rational methods, based on transformations which link terrain and image coordinates without any information about the characteristics of the used satellite. A rigorous model is based on the collinearity equations and reconstructs the orbital segment during image acquisition through the knowledge of the acquisition mode, orbital parameters, attitude angles, interior orientation and self-calibration parameters, atmospheric refraction: all the approximate values are known (some of them are in the metadata file) and their correction could be estimated using the Least Square Method. Rational methods were developed principally because sensor managing companies didn’t reveal technical characteristics of the platforms, so it was impossible to implement a correct rigorous model. Nowadays they approximate very well physical models, so very often also those companies use this solution to give information about sensor geometry without releasing neither the internal orientation parameter nor the physical features related to image processing. The most important method is the Rational Polynomial Function (RPF), which is based on a ratio of two polynomials of various order to model the relation between ground and image coordinates, so representing a generic formulation of the rigorous collinearity equations. The link is created by the Rational Polynomial Coefficients given in the metadata file or calculated directly by the final user thanks to some known ground points (even if this solution shouldn’t be used if high accuracy is needed). However, the knowledge of the coordinate of some ground control points is fundamental both with rigorous and rational models: after the orientation is performed, an accuracy assessment checking the Root Mean Square Error of this dataset must always be performed. As far as radiometric correction is concerned, it is useful to solve all the distortions in the acquired signal due both to sensor’s improper functioning (detectors’ relative brightness, striping, no data pixels and lines) and to external conditions (scene illumination and atmospheric conditions). Next step is related to the height information extraction, which is usually done with automatic procedures (image matching): analyzing all the differences between possible targets and background information, homologous points in the two images are recognized. Different elements could be used as search parameter: the radiance value of each image pixel, geometric features, spatial relations between various kind of features in the imagery. The first case is called Area Based Matching (ABM): a template containing the radiance values of a small area of one image is compared with the values of a moving window in the other image until the highest correlation is found. Region growing is very similar, but it starts the matching from some seed points well distributed on the imagery. Only using also a Least Square Matching it is possible to achieve sub-pixel accuracy, because this technique takes into account also the radiometric and geometrical distortions between template and search window, which always afflicts the data due to the different point of view of the imagery. The second possibility is the Feature Based Matching (FBM): various mathematical operator recognize simple features (such as corners or edges) and compare the created dataset among imagery. The last method is the relational matching, which is the less accurate one: it considers at the same time different relations between all the features in the images (geometrical ones, combinations of simple features, … ). Once the matching is finished, it is possible to calculate 3D coordinates of the recognized points using the parameters of the orientation step; finally, the model is extracted generating a regular grid which can then be used as input data for many applications. In fact, digital elevation models can be generated in three different ways: irregular points, which don’t contain any height information in the areas among them; TIN, Triangulated Irregular Network: the irregular points are liked using triangular faces which create a continuous surface; matrix models (GRID): the space is divided into a regular grid, so for each cell the height value is calculated interpolating the irregular dataset. The GRID model is typical when using imagery as input data and it’s also the most used for digital representation; the most important parameter is the grid size, because its value depends both on input data spatial resolution and on the detail needed by the final model. After the whole process is finished, the extracted model must be validate in order to define its accuracy and so determine its possible field of application (or if it is as precise as requested by the customer). Quality analysis must be performed together with comparisons with check datasets (having accuracy at least of one order better than the one expected for the model) to evaluate standard deviation on residuals between the model and the check data (control points, reference model, breaklines). In this thesis the entire generation process (from imagery orientation to DSM extraction) to generate a very accurate model was evaluated; both commercial and scientific software were tested in order to optimize human, economic and computational resources. Two different in-track high resolution satellite stereo pairs were processed: the first one was acquired over the area of Ferrara in 2004 by IKONOS II (spatial panchromatic resolution of 1m), the other covers the area of Argenta (FE) and it was taken in 2010 by GeoEye-1 (spatial panchromatic resolution of 0.5m). In both cases not only the panchromatic imagery (generally the only input data for DSMs generation) were used but also the multispectral ones (on the contrary generally useful to generate thematic maps). During the analysis different software have been tested, because with available commercial ones good results are achievable concerning stereo orientation, while some drawbacks are inherent to image matching: the commercial software Orthoengine (PCI Geomatics) and the scientific one SISAR (Software per Immagini Satellitari ad Alta Risoluzione) developed by Prof M. Crespi and the team of Geodesy and Geomatic Area - La Sapienza University of Rome and designed for the orientation of high resolution satellite imagery. Some tests inherent to IKONOS II imagery were performed also with ENVI 4.6.1(which gave unsatisfactory results) and the scientific one DPCOR, developed by prof. K. Jacobsen and his team at Leibniz University of Hannover. As far as the orientation is concerned, results with rigorous models show that accuracy of about 0.50.8 pixel in East and North direction and of about 0.81 pixel in Up direction is achievable; SISAR seems to be more stable varying GCPs number and distribution (620 GCPs are sufficient for an area of about 100Km2). The most important thing is the correct identification of the GCPs on imagery, because both software manifest higher errors during test run with point not well recognized (small errors of about 0.50.75 pixel for some points). Test performed with SISAR using rational functions show comparable results; moreover, RPCs generation and refinement don’t give better results, confirming a good pre-processing of the sensor managing companies while generating the RPCs. Finally, Tie Points recognition don’t improve the orientation accuracy. After that, image matching and DSM extraction were performed in small homogeneous areas with scientific software, while with OrthoEngine a full extent model was created. With regard to IKONOS II stereo pair, commercial software had some problems: matching wasn’t satisfactory in dense urban areas and many pixel had wrong height values (peaks and holes far from real ones) due to noise and difficulties during correlation step. Further tests were run with multispectral bands: NIR imagery gave good results, so using GIS software no data and noisy zones of the panchromatic models were edited and corrected merging data with NIR model. Derived DSMs generated with scientific software were less noisy, even if grid extraction is too approximate because it just creates a TIN. Accuracy evaluation was performed both with a set of points collected by kinematic and stop&go GNSS survey with 3D accuracy of 0.3m and with a reference grid of 2x2m with accuracy of 0.6m derived by a photogrammetric flight: calculated RMSEs show gaps of about 2 pixel in rural areas and 4 pixel in urban and vegetation zones. Even if features and infrastructures are well identified, these models aren’t suitable for high accuracy applications. With GeoEye-1 stereo pair only 40% of the whole scene was matched: built-up areas and streets were recognized whereas the major part of the fields were too homogeneous, because in march (when the imagery were acquired) there were just uncultivated fields and wheat plant were too small. This problem was much more important with OrthoEngine, so manual and automatic procedures to solve it were implemented, also using GIS software. Accuracy evaluation, performed as for IKONOS II stereo pair, show errors of about 1.5 pixel in rural area and slightly better than 4 pixel in urban one. The results are comparable to previous ones in terms of pixel residuals, showing an accuracy limit of models derived from high resolution satellite imagery; the great advantage is so represented by pixel size, which is halved in few years let DSMs to reach accuracy of about 1 meter. One of the main goals of this study was the creation of models to be used in the forecast and management of environmental emergencies, because the importance of this field of application is increasing day by day. In fact, the forecasting hydraulic models, which study the outflow of water after, for example, river floods or downpours, need as preliminary input a very good model of the studied area. The DSMs used for this purpose need to reach a very high precision, especially with regards to the correct identification of all the man-made and natural elements, like riverbanks or roads in relief, which may interfere with the outflow of water (deviating or stopping it) or, on the contrary, of all the crosses, which are a preferential way to water’s flow. Furthermore, this demand becomes more important for flat areas, because little errors in height definition could cause important variations in the identification of flooded areas. According to CISIS’s technical guidelines about elevation models, results obtained with both the stereo pairs show that the extracted DSMs aren’t suitable for this purpose, because height values must have an accuracy better than 0.5m (the DSMs extracted are at level 3, while they should be at least at 5). With regard to data derived from GeoEye-1 imagery, they seems to be useful, even if merged with other surveying technologies to acquire in a better and more precise way the characteristics of all the barriers and the crosses. This fact was demonstrated also with simulations run on the same area by the technician of the land reclamation consortium of Ferrara (Consorzio di Bonifica Pianura di Ferrara): in very flat areas (such as the one investigated), in order to obtain a correct identification of flooded areas (both with geographic position and temporal sequence), the correct implementation of barriers and drains network is much more important than the height information of areas in the middle of the hydraulic cell. So, using scientific software it would be possible to obtain a refinement of the matching parameters and procedures in order to create accurate DSMs also starting from imagery with GSD of 0.5 meter; in any case, an accurate editing done by the user and a correct point cloud interpolation is fundamental. On the contrary, because the user can’t generally set all these parameters using commercial software, extracted DSMs should be used only for applications where lower accuracy is needed. In the first part of the thesis all the theoretical aspects are described: after some information about photogrammetry and remote sensing (Chapter 1) and the main characteristics of platforms (Chapter 2), Digital Terrain and Surface Models are discussed (Chapter 3), focusing the attention to the whole generation process performed using high resolution satellite imagery (Chapter 4). The second part presents all the procedures adopted and the results obtained both with the IKONOS II stereo pair (Chapter 5) and the Geoeye-1 imagery (Chapter 6).File | Dimensione | Formato | |
---|---|---|---|
408.pdf
accesso aperto
Tipologia:
Tesi di dottorato
Licenza:
Non specificato
Dimensione
17.95 MB
Formato
Adobe PDF
|
17.95 MB | Adobe PDF | Visualizza/Apri |
I documenti in SFERA sono protetti da copyright e tutti i diritti sono riservati, salvo diversa indicazione.