Thematic mapping, vegetation and DEMs

Digital elevation models are of fundamental importance for a large variety of scientific and commercial applications. For example, precise and up-to date information about the Earth’s topography is required in many geoscience areas, such as geology, forestry, glaciology, oceanography, and hydrology. For reliable navigation using global positioning systems, such as GPS and Galileo, digital elevation maps are of essential importance as well. Up to now, the primary source of elevation data on an almost global scale has been provided by the Shuttle Radar Topography Mission (SRTM), which is characterized by a spatial sampling of 30 m between [-60°,+60°] latitudes. For higher latitudes and over Antarctica, only low resolution DEMs are available, such as GTOPO, RAMP, and GLAS/ICESat.
With the main goal of acquiring a global and consistent DEM with unprecedented accuracy, the TanDEM-X mission (TerraSARX add-on for Digital Elevation Measurements) opened a new era in spaceborne synthetic aperture radar (SAR). Developed in a public-private partnership between the German Aerospace Center (DLR) and Airbus Defence and Space, it is comprised of two almost identical satellites, TerraSAR-X and TanDEM-X, mounting a synthetic aperture radar antenna operating at X-band. Since October 2010, both satellites have been flying in a close orbit configuration at an altitude of around 500 kilometers, acting as a single-pass SAR interferometer and allowing for a flexible selection of baselines and acquisition geometries. Images are nominally
acquired in bistatic configuration, where one satellite transmits and both simultaneously receive the backscattered signal from the Earth’s surface, enabling the acquisition of highly accurate interferograms, which do not suffer from temporal and atmospheric decorrelation.
A dedicated acquisition strategy has been developed and optimized throughout the years, in order to achieve a homogenous performance globally. The satellite formation has been adjusted accordingly during different acquisition phases, depending on the target height performance to be achieved.
This paper reports the final performance of the TanDEM-X global DEM, presenting the developed acquisition and processing strategy, and a detailed analysis of all characteristic parameters. The global performance is then assessed in terms of vertical and horizontal errors, and coverage statistics. The obtained results confirm the outstanding quality of the delivered product, which can be now globally exploited for both scientific and commercial applications.

The quantification, monitoring, and mapping of agricultural biomass are of paramount interest in today’s world economy due to their central importance in the production of bioenergy and the prediction of crop yield. However, the generation of accurate forecasts from crop models requires a wide variety of inputs ranging from environmental conditions (e.g. rainfall, soil moisture) to agronomical information (e.g. crop phenology). In most of the cases, the considerable extension of fields coupled with the wide assortment of farming practices makes arduous to acquire the required inputs originating a lack of information.  In response to these needs, SAR remote sensing offers a valuable tool to deliver timely accurate crop condition data over large areas and at relatively low costs. Past research has confirmed correlations between numerous SAR observables (e.g. backscattering, entropy, linearly polarised channel ratios) and crop biomass for different polarisations, frequencies, and crop types and, at the same time, identifying saturation limits (i.e. points at which SAR observables no longer scale with crop biomass) [1,2].

This study aims at investigating an alternative technique for the estimation of crop biomass based on Polarimetric Differential SAR Interferometry. Routinely applied for the estimation of Earth’s displacement along the line of sight direction, this technique has been recently found to be sensitive to vegetation changes (e.g. vegetation growth, senescence) and soil moisture variations [3,4].

The relationship between the PolDInSAR observables (i.e. magnitude and phase of the PolDInSAR coherences) is empirically investigated using zero-baseline L-band data acquired in the frame of an airborne campaign over the agricultural test site of Wallerfing (Germany) in 2014. The nearly zero-baselines, short revisit times (3-7 days), and a dedicated interferometric processing are expected to reduce the impact of additional influential factors, such as topography and motion compensation errors. With the aid of multiple linear regression techniques, the PolDInSAR observables are described as linear functions of the explanatory variables i.e., soil moisture and crop biomass. Particularly, the latter are assumed to impact the magnitude of the PolDInSAR coherences |γ| in a multiplicative fashion while the referenced differential phase Φ is assumed to be linearly governed by the regression variables. The empirical study is conducted separately for each polarisation channel as the impact of the latter might not be necessarily the same.

The estimated regression coefficients exhibit a predominantly a positive sign while the size of the effects differs with polarisation and incidence angle. The sign of both effects is consistent with the findings reported in [3] i.e., an increase of crop biomass between two SAR acquisitions influences Φ in a similar fashion as an increase of soil moisture enlarging the optical path between the sensor and the scatterers (e.g. soil and/or vegetation constituents) and causing decorrelation.

The regression coefficients for Φ present polarisation discrepancies which revealed to be helpful for providing a further insight into the scattering physics giving rise to the vegetation effect. The crops where this polarisation inconsistency occurs exhibit a predominant vertical orientation of their constituents. For canopies showing such preferential orientation (i.e. wheat and barley), the electromagnetic interaction is expected to be remarkably stronger in VV than in HH polarisation, unless the horizontal extent of the scatterers is not smaller than the wavelength, as in the case of mature corn canopies [5]. The pronounced interaction with the VV polarisation is consistent with the higher magnitude of the regression coefficients modelling changes in crop biomass. Moreover, the reported changes of the effective propagation attributed to the forward scattering in the vegetation layer are traceable in the slope term of Φ with respect to crop biomass. This term is expected to be larger in VV than in HH for vertically oriented crops and this difference is significant at α=0.05 for barley, wheat, and for one sample of rape. On the contrary, the regression coefficients for |γ| exhibit only a modest dependence on the choice of polarisation.

Exploiting the pronounced linear relationship between the referenced differential phase and crop biomass coupled with the weak size of the soil moisture effect, the linear phase model is inverted on the account of the regression coefficients computed in the observational analysis. This approach allows generating relative biomass maps (i.e. mapping variations of crop fresh biomass between two SAR acquisitions) of the analysed crops in different polarisations. The estimated fresh biomass is in good agreement with the collected ground measurements. In particular, for vertically oriented canopies, the correlation with ground measurement presents a R-squared close to 0.93 for biomass maps generated using VV polarisation.

The accuracy of the relative biomass maps obtained with this approach is further assessed in terms of robustness to the implementation of the multilinear regression analysis. The choice of the interferometric phase reference (e.g. corner reflector or persistent scatterer in the surrounding area), parametrization of vegetation component (e.g. in terms of biomass or vegetation height), and the normality of the error terms only exhibit a minor influence on the patterns found in the empirical analysis. Therefore, the conclusions drawn from the latter appear to be robust with respect to these assumptions. 

Synthetic Aperture Radar Interferometry (InSAR) is a powerful tool for large-area topographic mapping due to its capability of all-time all-weather imaging and high sensitivity to terrain relief. However, there is an inherent contradiction between geometric decorrelation and sensitivity of height measurement for topographic mapping with a single InSAR pair. A normal baseline of proper length is required to keep a balance between the two issues. A promising solution to this problem is the so-called multi-baseline InSAR analysis. The basic principle of multi-baseline InSAR is to derive an optimal height estimate by joint analysis of multiple phase measurements from a few interferograms with different normal baselines. Compared with single-baseline InSAR, the major benefit of using multi-baseline observations is the possibility of exploiting redundant topographic phase observations with different height of ambiguities to improve the accuracy of phase unwrapping, or even avoid phase unwrapping. 

Stereo radargrammetry (StereoSAR) is also widely used to reconstruct digital surface models (DSMs) Compared with InSAR, it is less affected by temporal decorrelation. Therefore, StereoSAR is often used for DSM generation in heavily vegetated areas but with lower accuracy. As two major approaches to use SAR remote sensing data for topographic mapping, both StereoSAR and InSAR techniques have their own advantages and disadvantages. Consequently,  it is not easy to obtain accurate and reliable height information in terrain relief areas with vegetation cover by adopting only either technique. Aiming at such a problem, we are going to carry out a study on joint utilization of StereoSAR and dual/multi-baseline InSAR to make full use of advantages of the two techniques to substantially improve our capability of height measurement.

In this study, we developed a maximum a posteriori (MAP) estimation method for multi-baseline InSAR assisted by StereoSAR. According to Bayesian theory, the combination of StereoSAR and InSAR for topographic mapping can be viewed as update of the StereoSAR DSM with InSAR phase observations. At the same time, StereoSAR DSM is also a constraint to InSAR phase observations, which can solve the problem of elevation ambiguity and avoid phase unwrapping process. Under the condition of distributed scattering, the likelihood function of height could be derived from InSAR phase. The prior probability of height is acquired form StereoSAR DSM.

In order to verify the proposed method, we choose Mount Song as the test area, which is one of the five sacred mountains of China. Although the mountain peaks reach only around 1500 m, the slopes are very steep and they are densely vegetated, making Mount Song an difficult area for InSAR DSM generation. In our study two interferometric data pairs acquired by TerraSAR-X with different normal baselines and one stereo pair of TerraSAR-X data in Stripmap mode are used together to reconstruct the DSM of 10 m spatial resolution. In order to evaluate the accuracy of generated DSM, we use a 1 m resolution DSM created by airborne photogrammetry as reference data. The experimental result shows that there is neither systematic error nor large data voids in MAP estimated DSM and the standard deviation of height error σ_h of MAP estimated DSM is less than 10 m with respect to the photogrammetric DSM for the whole area, while in plain areas σ_h is about 5 m.

Research on SAR interferometry is mainly focused on the two main application areas, namely the generation of accurate digital elevation models and change detection, and also on the crucial processing steps like interferogram filtering, phase unwrapping, co-registration etc. This study rather aims to make use of interferometric SAR (InSAR) images, once they are generated, for scene recognition purposes.

In literature, limited research is available on the use of InSAR images for object recognition and scene classification. For those studies, the main trend is in the direction of using the interferometric coherence for mostly binary classification such as forest/non-forest, urban/non-urban or change/no-change classification. On the other hand, available research on multi-class classification is mostly based on the temporal variation of interferometric coherence and/or backscatter intensity.

However, in our studies, we emphasize the use of whole complex-valued InSAR image for multi-class classification, i.e., recognition of more complex scenes such as forest, agricultural fields, water body, different kinds of residential and industrial areas, etc., and also their combinations. Furthermore, a new complex-valued phase-gradient InSAR (PGInSAR) image is defined whose phase represents the magnitude of the phase gradient of InSAR in range and azimuth directions. Interferometric phase being related to the terrain height, phase of PGInSAR image can be considered as a measure of the terrain slope, i.e., how fast the interferometric phase changes over the image.

This review study serves as a comparative assessment of various feature descriptors such as Gabor-based, FrFT-based, BoVW-based and partial derivatives based features extracted from the complex-valued InSAR and PGInSAR images, and used for patch-based classification. For this purpose, an image patch database is generated from bistatic interferometric pairs acquired by the TanDEM-X mission over the test site Toulouse, France. In order to investigate the impact of effective baseline on the classification, 3 datasets are constructed from the bistatic acquisitions with 3 different effective baselines over the same area. Supervised KNN classification is performed on the image patches of 200 x 200 pixels from 8 different scene classes representing different natural and man-made structures on a 400m x 340m terrain (Figure 1).

The classification results can be summarized as follows:

The features extracted from the complex-valued InSAR and PGInSAR images improve the mean accuracy by 15% and 27% compared to detected or complex-valued SAR images, respectively [1][2].

The use of InSAR and PGInSAR images improves the individual class accuracies for scenes with natural structures, such as agricultural area, forest and mixed vegetation [1][2]. 

PGInSAR images are found to be more robust to effective baseline changes than InSAR images, as the phase discontinuities are reduced compared to InSAR images [2].

Although the global Gabor features yield better accuracies than the global FrFT features for InSAR images, by the implementation of BoVW model, FrFT outperforms Gabor features [3].

The spatial discriminability of BoVW model together with the strong interferometric signature of urban structures for larger baseline results in better classification performance [3].

The use of partial derivatives based scale-space representation improves the classification performance for scenes like industrial and urban areas. The presence of many local variations in such scenes is represented quite successfully by these features [4].

Our previous work:

[1] N. D. Cagatay and M. Datcu, “Scene Recognition Based on Phase Gradient InSAR Images,” IEEE International Conference on Image Processing (ICIP), Paris, France, 27-30 October 2014.

[2] N. D. Cagatay and M. Datcu, “FrFT Based Scene Classification of Phase Gradient InSAR Images and Effective Baseline Dependence,” IEEE Geoscience and Remote Sensing Letters, vol. 12, no. 5, pp. 1131-1135, May 2015.

[3] N. D. Cagatay and M. Datcu, “Bag-of-Visual-Words Model for Classification of Interferometric SAR Images,” European Conference on Synthetic Aperture Radar (EUSAR), Hamburg, Germany, 6-9 June 2016.

[4] N. D. Cagatay and M. Datcu, “Multi-Scale Feature Extraction Approaches for Classification of InSAR and Phase Gradient InSAR Images,” IEEE International Conference on Image Processing (ICIP), Phoenix, Arizona, USA, 25-28 September 2016.

The Sentinel-1 mission represents a challenging opportunity to develop operational applications for classification and vegetation mapping using periodical interferometric SAR acquisitions. The SinCohMap project under the ESA SEOM framework is intended to develop novel methodologies exploiting the evolution of the interferometric products. To compare with state-of-the-art classification strategies, within the project context a round-robin consultation is defined to involve both internal and external expert groups to test their algorithms using a common set of well-known and in detail monitored scenarios.

The main objective of this research is to develop, analyze and validate novel methodologies for land cover and vegetation mapping using time series of Sentinel-1 data and by exploiting the temporal evolution of the interferometric coherence. Further the project aims on quantifying the impact and possible benefit of using Sentinel-1 InSAR (Interferometric Synthetic Aperture Radar) data relative to traditional land cover and vegetation mapping using optical data (especially Sentinel-2) and traditional intensity-based SAR (Synthetic Aperture Radar) approaches. A careful revision of state-of-the-art classifiers along with the definition of novel methodologies specifically designed for this scenario is being addressed within the project framework.

The most innovative concept is the creation of a robust library based on a forward physical model, which can help identify or separate temporal decorrelation due to physical changes of the scatterers. In such a way, the models can be derived from expected physical changes towards time (e.g. vegetation growth, crops evolution) to identify the classes and to separate from spike measurements or outliers. Machine learning algorithms are also being analyzed due to their high performance and their ability to deal with complex data such as high-dimensional imagery and multi-source data sets.

The main classes sought after are Forests, Agricultural areas (e.g. Crops), Artificial surfaces (e.g. Urban), Water Bodies, Scrub and Herbaceous Vegetation, Open or bare land with little to no vegetation and Wetlands. Three different reference test site areas are defined with very accurate ground truth data are defined for performing quantitative assessment and validation in Spain, Italy and Poland. Additional evaluations to detect the ability to extrapolate the extracted ideas will be performed over boreal forest.