Thematic mapping, vegetation and DEMs

We propose an interferometric SAR (InSAR) denoising filter based on the nonlocal filtering principle. The goal is to generate a digital elevation model (DEM) from TanDEM-X interferograms with a higher resolution and accuracy than the standard product, which relies on simple boxcar averaging. In our previous research [1], [2] we noticed the potential of nonlocal filters, but also some of their flaws, which the proposed filter alleviates.
The nonlocal filtering principle exploits that images have an inherent redundancy so that similar patterns are found multiple times. To denoise a pixel a nonlocal filter searches for similar pixels in its vicinity, the so-called search window. Similar pixels are not just identified by themselves but by comparing their surrounding patches. The logic being that similar pixels have similar neighborhoods, leading to a more robust estimate which also preserves textures and details. After all pixels in the search window have been assigned a weight based on their degree of similarity, the actual pixel value is estimated by their weighted average.
We present in brief the most significant features of the proposed algorithm.
The local fringe frequency caused by the topography is estimated and taken into consideration when searching for similar pixels. This deterministic phase component would otherwise diminish the denoising performance for hills and mountains.
As the noise level hampers the search for similar pixels our method adopts a two-stage approach. The first step consists of regular nonlocal filtering, whereas the second step uses the filtered output of the first to search for similar pixels.
The similarity criteria employed in both steps are taken from [3]. The proposed method adaptively selects the patch size depending on the phase heterogeneity index derived in [4]. By doing so the filtering performance increases in homogeneous, where a larger patch size leads to a more accurate estimate, as well as in heterogeneous areas since smaller patch sizes combat the rare patch effect of nonlocal filters.
Just like existing nonlocal filters our proposed method provides an improved noise  reduction and detail preservation of conventional local filters. Taking into account the phase heterogeneity and topography enhances the filtering performances and can also be incorporated in other nonlocal filtering algorithms.
We are also currently in the process of a more exhaustive analysis of the final DEM’s accuracy.
References
[1] X. X. Zhu, R. Bamler, M. Lachaise, F. Adam, Y. Shi, and M. Eineder, “Improving TanDEM-X DEMs by Non-local InSAR Filtering,” in EUSAR 2014; Proc. of, Jun. 2014.
[2] G. Baier, X. X. Zhu, M. Lachaise, H. Breit, and R. Bamler, “Nonlocal InSAR filtering for DEM generation and addressing the staircasing effect,” in EUSAR 2016; Proc.
of, Jun. 2016.
[3] C.-A. Deledalle, L. Denis, and F. Tupin, “NL-InSAR: Nonlocal interferogram estimation,” IEEE Transactions on Geoscience and Remote Sensing, vol. 49, no. 4, pp. 1441–1452, Apr. 2011.
[4] J. S. Lee, K. P. Papathanassiou, T. L. Ainsworth, M. R. Grunes, and A. Reigber, “A new technique for noise filtering of SAR interferometric phase images,” IEEE Transactions on Geoscience and Remote Sensing, vol. 36, no. 5, pp. 1456–1465, Sep. 1998.

Interferometric repeat-pass coherence is studied for clear-cut mapping in a Sentinel-1 time-series.  The hypothesis is that forest coherence is continuously low before logging while after logging there are image pairs that under favourable acquisition conditions show significantly higher values in areas logged.

In an earlier work in the EU/FP7 project North State, Sentinel-1 amplitude data was shown to produce meaningful clear-cut detection of patches larger than 1 hectare in the conditions characterized by thick snow cover during image acquisitions.  A pair of images were used before (2014-12-08 and 2015-01-13) and after (2016-01-20 and 2016-02-25) the loggings (Rauste et al, 2016).  These acquisition dates were selected based on high contrast between open and forested areas in these images.  The same time-series as in project North State, will be used also in the interferometric follow-on study.

The experiments to be made include:

the coherence between 2014-12-08 and 2015-01-13 compared to the coherence between 2016-01-20 and 2016-02-25, and

a time-series of 12-day coherence over a verified clear-cut area.

Conclusions are drawn and recommendations made for further studies.

The SNAP/Sentinel-1 toolbox of ESA will be utilized in interferometric computations.

References

Rauste, Y., Antropov, O., Mutanen, T., and Häme, T. 2016. On clear-cut mapping with time-series of Sentinel-1 data in Boreal forest, Proceedings of Living Planet Symposium 2016, Prague, Czech Republic, 9-13 May 2016 (ESA SP-740, August 2016), 9 p.

Remote sensing images from the space have always been an obvious and promising source of information for deriving crop maps. This is mainly due capabilities to timely acquire images and provide repeatable, continuous, human independent measurements for large territories. Crop mapping and classification of agricultural crops is extremely valuable source of information for many applied problems in agricultural monitoring and food security.

Taking into account that free optical satellite data was available for many past years and the weather-independent synthetic-aperture radar (SAR) images were very expensive, a lot of studies on crop classification tasks had been done using only optical data. In the same time, for time-series based on optical data there are some issues, such as clouds and shadows effect, and as a result different number of observation for the study area. There are different techniques to deal with this issues: methods for clouds and shadows restoration [1], feature extraction methods [2] etc.

Thanks to the launching Sentinel-1A (S1A) SAR satellite by European Space Agency (ESA) in 2014, we have access to the free high resolution weather-independent SAR images. It allows us to solve the problem with clouds, to equalize number of observation for the all study area and to increase the number of observation.

In this study, we compare three data sources for crop classification maps derivation: Sentinel-1A data, Sentinel-2A data and data fusion from Sentinel-1A and Sentinel-2A. For Sentinel-1A SAR series, only pre-processing to produce geocoded imagery is required before classification, for which we use the SNAP Toolbox. For Sentinel-2A time-series we use only 4 bands with 10m spatial resolution (Blue, Green, Red and near infrared (NIR)). Ground truth data were collected within along the road surveys in 2016 and were randomly divided for training and test samples in equal proportions. Test set was using for independent result validation. For this crop classification investigation ensemble of neural networks had been utilized [3].

Detailed experimental results in term of overall, user accuracy, producer accuracy and crop classification maps for Sentinel-1A, Sentinel-2A and fusion of Sentinel-1A and Sentinel-2A will be presented.

Keywords: agriculture, image processing, data fusion, Sentinel-1, Sentinel-2.

[1] N. Kussul, S. Skakun, A. Shelestov, M. Lavreniuk, B. Yailymov, and O. Kussul, “Regional Scale Crop Mapping Using Multi-Temporal Satellite Imagery,” Int. Arch. Photogramm. Remote Sens. Spatial Inf. Sci., XL-7/W3, pp. 45–52, 2015.

[2] N. Matton, et al., “An automated method for annual cropland mapping along the season for various globally-distributed agrosystems using high spatial and temporal resolution time series,” Remote Sensing, vol. 7, no. 10, pp. 13208-13232, 2015.

[3] S. Skakun, N. Kussul, A. Y. Shelestov, M. Lavreniuk, O. Kussul, “Efficiency Assessment of Multitemporal C-Band Radarsat-2 Intensity and Landsat-8 Surface Reflectance Satellite Imagery for Crop Classification in Ukraine,” IEEE Journal of Selected Topics in Applied Earth Observations and Remote Sensing, 2015, DOI: 10.1109/JSTARS.2015.2454297.

The introduction of Sentinel-1B in September 2016 provides a unique and novel opportunity to generate C-band interferometric coherence over very large areas for 6-day temporal baselines. In Europe, large areas are covered by both ascending and descending orbits, and, especially at higher latitudes, by adjacent orbits. Thus, coherence information can be collected for an even denser temporal sequence, though always interleaved for 6 day intervals. The interest of using coherence in agricultural use contexts has been relatively limited, mostly due to the fact that, during the growing season, vegetation cover often leads to very low coherence, especially if the temporal baselines are too long (e.g. 24 days for Radarsat, 35 days for ENVISAT). Although commercial SAR systems (e.g. TerraSAR-X, CosmoSkyMed) may be able to achieve better temporal baselines, though with a significantly higher frequency (X-band), their use for wide area monitoring is prohibitively expensive. The “full, free and open” data license of the Copernicus programme and the extraordinary performance of the Sentinels are essential to vastly scale up the use of SAR in crop monitoring.

Currently, the agricultural user community is abuzz with the more traditional application examples of SAR, and hybrid SAR and optical, data use, for instance, in crop classification. The joint JRC-ESA-SZIF experiment Czech-Agri (part of the Sen2Agri project) has demonstrated that country-wide consistent crop maps can be derived from combination of Sentinel-1 with Landsat-8 (2015) and Sentinel-1 with Sentinel-2 (2016) combined with existing reference parcel information and targeted surveying. Similar results are available, in other use contexts, in the United Kingdom, the Netherlands, Finland and Ukraine. The GEO Global Agricultural Monitoring (GEOGLAM) community of practice now clearly recognizes the need to integrate SAR in crop classification beyond the well-established use in rice monitoring. Classification accuracies in the examples typically reach the 85-90% overall accuracy range for a significant set of crop types. None of these activities include coherence analysis, however.

Coherence of agricultural surfaces is strongly related to the stability of the surface geometry. This is the key reason why crop canopies, with [moving] vegetation structures that are in the order of the C-band wavelength, exhibit low coherence. Undisturbed bare soil, however, tends to show high coherence. The key word here is “undisturbed”, implying that disturbances that significantly change the surface structure (e.g. ploughing, seedbed preparation, erosion) is detectable as a loss in coherence. Also, emergence of vegetation will lead to a, gradual, loss of coherence. Integrating the changes in backscattering intensity, and partial polarimetric decomposition, provides further clues about the direction of change, for instance, from a smooth surface to a rough surface. For the latter, the use of meteorological records is essential to understand the separate impacts of soil moisture, which may lead to coherent change in backscattering, and incoherent surface structure change. Apart from the potential to use coherence to further refine crop classification products, we expect the greatest added value in the analysis of crop phenology and variation within crop groups, as a contribution to refined crop yield modeling and crop production estimates.

We will demonstrate the use of a time series of combined Sentinel-1 coherence and backscattering intensities over the Netherlands, where we have a complete reference data set and detailed weather information. We highlight the effects of agricultural management practices and how their detection are input to [very] early delineation of crop type probability maps. Our time series will cover the initial growing season of 2016/2017 for which we will integrate available Sentinel-2 information to determine how sensitive the signatures are to crop emergence and canopy closure. We will discuss requirements for large area generation, most of which is automated, and the relevance of our work to European Common Agricultural Policy management and control, with selected examples.

2.03.c   Thematic mapping, vegetation and DEMs