Science Service System

Summary of Proposal LAN0167

TitleThe application of TerraSAR-X data for land cover mapping, crop inventories and agricultural practices check.
Investigator MROZ, Marek - University of Warmia and Mazury in Olsztyn, Department of Photogrammetry and Remote Sensing
Team Members
Dr Chmiel, Jerzy - Warsaw University of Technology, Laboratory of Remote Sensing and GIS
MSc Szumilo, Malgorzata - University of Warmia and Mazury in Olsztyn, Photogrammetry and Remote Sensing
Dr Perski, Zbigniew - Delft University of Technology, Delft Institute of Earth Observation and Space Systems
SummaryThe project will start in the spring of 2007. TerraSAR-X data will be acquired in repeat cycle of nominal 11 days in SM Dual Polarization mode, both as detected products (GEC)and SSC complex images. The objectve is to investigate the usage of agricultural parcels in terms of the crops cultivated and the practices applied for the maintenance of good parcel conditions to production (eligibility for subsidiary). Additionally other than agricultural types of land cover will be also mapped. Ground truth data including crop identification and its state description will be gathered during the dates of imaging. The main deliverables should be the detailed crop and land cover maps providing the information useful for the declared crop control. At this moment the funding from the university budget is assured but the PI will apply for the financial support from the Polish Ministry of Science and Informatization.
Detailed reportScenarios of radar and optical images acquisitions for the test area Malbork. TerraSAR-X images were programmed on the basis of system scheduling and available configuration at given day in 11-days interval. Strip Map mode at single and dual polarizations configuration were chosen. The area of interest (AOI) of 20x20 km large is located in Northern Poland in typically agricultural region, partially on flat alluvial soils (west) and postglacial moraines (east) with moderate relief. The TSX acquisitions started on 25th of April 2008 when the winter crops (rape, cereals, …) were developed above the ground but the spring crops were just emerged or not yet sown. All images were taken on descending pass and pre-processed to EEC (Enhanced Earth Corrected) product variant using precise orbit parameters (SCIEntific). All products were ordered in RE (Radiometrically Enhanced) resolution variant giving less speckle constraints. Two strips, R013 and R007, having different looking angles were chosen. Incidence angle at the scene centre for R013 Near_strip is about 42 deg. and 31 deg. for R007 strip respectively. This difference of about 11 degrees could be examined as an additional potential discriminative factor of crops classes for different surface roughness situations resulting from crops development stages. For both strips the images were gathered in VV polarization. For R013 strip in dual-pol mode the second polarimetric component was cross-pol channel VH for 2008 season and the like-polarised HH for 2009 season. Creating a vector database of agricultural parcels was the initial stage of work. For this purpose the data of land and building registration were used. Each parcel is assigned to a new identification number and the attribute specifying whether the polygon is non-agricultural or agricultural area. In this way, database of agricultural parcels and mask of agri and non-agri area were created and then used for further analysis. Several hundred parcels have been identified during about fifteen field survey called RFV - Rapid Field Visits in 2008 and 2009. Type or species of plants, stage of growth and the average height of plants have been documented. Representative parcels, as well as abnormal, different development phase have been identified. Digital photographs of the parcels visited have been taken during the visit, and stored in a database with their location. Methodology for the processing of radar images TSX/EEC and optical images for the series from the years 2008 and 2009. The main aim of the methodological procedure was to create the subsets of data which enable to extract maximum of information in the shortest multitemporal dataset. Field visits were made on days when satellite was flying or very close these days, while the photographic material has always come from the previous flight. Thus it was possible to "visit" every time the same fields in order to document the changes in the appearance of crop and its growth stages but also to expand the database in cases where discovered parcel was distinguished by color on another color composition. Multitemporal color and dual-pol compositions were the basic working material. They ensure the creation of reference parcels set, which take into account the diversity of color of parcels inside the thematic class so that the classes generated during classification were not too narrow. As all registered single and dual polarization bands were delivered in EEC precise geometry, with sub-pixel overlaying accuracy, deGrandi multitemporal filtering was applied for each strip (R007 and R013) separately. Within this multitemporal filtering an optimum weighting filter is introduced to balance differences in reflectivity between images at different times. The filtering was done with SARScape software. For stronger speckle filtering smaller values of number of looks were chosen: 1.3 looks for R013 dual-pol VV/VH strip and 2.2 looks for R007 single-pol VV strip. DeGrandi filtering was followed next by moderate smoothing filter - median 5x5. The images representing backscatter intensity were next converted from 32-bits values (real-float) to 8-bit values with very slight histogram stretching on the right side: 0.15 % for VH bands and 0.30 % for VV bands in both strips. For more precise crops distinguishing an agricultural / non-agricultural areas mask was used. The following subsets of gathered data were analysed separately: -TSX R013 VV/VH - time series from longest to shortest -TSX R013 VV only - time series from longest to shortest -TSX R007 VV - time series -IKONOS + TSX R013 VV/VH shortest time series. The same analysis was done for the same time series of R013 dual-pol acquisitions but neglecting cross-polarisation component VH. The results could give us an indication on the use only VV polarisation, at the same incidence angle of about 42 degrees (R013 or R014 strips) but in single polarisation mode. The first steps in time series image analysis were calculations of Jeffries-Matusita Distance (JMD) - class separability measure. The analysis of class separability lead to their redefinition and the “per pixel” classification was made using maximum likelihood classifier on the “spectral/temporal” signatures extracted from all reference parcels. For each classification routine the same high probability level of 0.9 was set up in the classifier parameters set. The last step in classification procedure was calculation of Confusion Matrix and Kappa Index of Agreement. The last part of processing consisted in statistical analyses (JMD measure and Confusion matrix) of joint sets of data (optical+TSX) acquired on the same day or close each to other in time in order to check how the TSX imagery can improve classification accuracy based on optical data alone. The TSX acquisitions started on 15th of April for the 2009 season, with a nominal 11-day interval. Based on the results of the 2008, only strip R013 in VV/HH polarization was gathered. A sequence of five acquisitions was completed on June 28. Thus in both seasons radar images were taken before the harvest. Processing a series of TSX / SSC / 2008 images to form a map of coherence. Coherence has become a feature of the dynamics of land cover change. Imaging product called ILC / ILU - Interferometric Land Cover / Land Use Interferometric was made. Series of 18 maps of coherence based on the images TSX / SSC from the 2008 season was generated in the project. They show that, in accordance with the theory, the coherence between the VH components is almost zero for natural objects (uniform distribution) and is higher only in urban areas and railroads. Much more dynamic coherence is noted for a series of components of VV. The high coherence is noted for the plots of crops, which at the beginning of the growing season are bare soils like plots of sugar beets, potatoes and corn, both the 11-day interval and with the 22-day interval. Very low coherence with the same approximately low values characterize the remaining parcels of land (cereals, grasslands, rape). Parcels whose coherence changes from low to high at a later date, have been identified in the analysed series of images, but they are beyond the main AOI limits and they were not reported during the field visits. Generating coherence implies a choice of window size, which examines the stability phase, then the choice of pixel size, which will be adopted to change the image geometry from original SSC (slant range) into the target system of cartographic projection. Coherence images were generated with a pixel size of 7x7 m using resampling function of 3rd degree, and the next stage was transformation them according to precise formulae of cartographic reprojection to UTM34N with 3.5x3.5m pixel size. Coherence calculations were performed using an algorithm implemented in SARScape software. The results obtained in the project can be summarized as follows: TerraSAR-X images in StripMap mode have a high suitability for mapping of crops in the framework of CwRS/CAP and for GAEC’s check in some aspects. Dual polarimetric images acquired with incidence angle 42 deg.and with resolution 3.5 m comparing to images acquired with angle 31 deg. with resolution 2.75m are more useful. Selection only the VV channel to carry out an automatic classification also demonstrates the increased utility of imaging angle 42 than 31 deg. This assessment is incomplete because the R007 analysed images were recorded with a time delayed in relation to the R013 series VV / VH up to three weeks, with a shorter interval. The first image was recorded on May 22, the third on June 24, and the last-four on 29 August. Crops like rape, sugar-beet, cereals, potatoes, maize can be automatically classify with very good accuracy using the best series of registration. Winter wheat can be characterized by diversity of colours on color compositions resulting from the presence of different varieties, the density of cereals and the sowing date. The values of separability indicators and confusion matrix highlight the need to include registration of the spring, when part of the crop is still in emerging or germination phase. Applied X-band does not allowed even on a five images VV/VH, correctly identify and separate grasslands from arable crops. It is an important obstacle in agricultural landscape mapping because protection of permanent meadows and pastures is an important element of the Common Agricultural Policy. The results also indicated that the band "X" relevance VV and HH components for classification and diagnosis of crops is very similar. Separability indicators are somewhat better for vertical polarization VV than horizontal HH, but not enough to result also in the overall classification accuracy and Kappa coefficient. Differences in overall accuracy for both cases of polarization varies from 2 to 7%, even for the benefit of horizontal polarization. Using two polarization in the longest series of images improves the quality of the classification of 4-5.5% (compared to single pol.), but the accuracy at the shortest series deteriorates more than 2%. It follows that the system polarization VV / HH is less favourable to the mapping of agricultural crops than the system polarization VV / VH. In the latter case, the benefit of an cross polarization VH component obtained in the 2008 season, measured by the index "overall accuracy" was about 25%, compared to a single vertical polarization. The use of basic 11-day or 22-day revisit interval increases costs of acquisition of satellite images TSX (5-7 images) but allows the use maps of coherence in agricultural monitoring. The results show that high coherence is achieved in areas without vegetation in the advanced growth stage. Thus it is a way to detect aberrant parcels, parcels with a crop change in a shorter cycle than in other plots. Increased spatial resolution of TerraSAR-X and reduced observation interval to 11 days allowed to obtain additional element in crop recognition process. Sugar-beet and potato crops in the early growth stages were parcels of high coherence in the site. The results obtained in the study indicate that the classical multispectral images acquired at medium resolution, several times during the growing season, can actually be replaced by a series of TSX images for the classification of the above mentioned groups of crops. An important argument for using integrated optical and radar images is the ability to improve the identification and delimitation of significant losses in crops caused by storms or other factors like chemical contamination. An important element in terms of the GAEC requirements is the management of stubble. Applying an TSX image taken at the beginning of September shows that through visual interpretation it can be identified on which parcel were left post-harvest stubble (unfavourable due to the loss of soil water) and where stubble were ploughed after harvest of cereals.

Back to list of proposals

© DLR 2004-2016