Multi-temporal Classification for Crop Discrimination using TerraSAR – X Spotlight images

Within the past years investigations have been carried out on the usefulness of ENVISAT ASAR dual polarimetric data for environmental mapping of agricultural areas. The performance of such approach is limited by the spatial resolution of the data, which is firstly too coarse for sufficient sampling of small fields often found in Europe and secondly useful features related to agricultural treatments (irrigation, fertilization, soil and plant treatment) are blurred. However, it was shown that proper selection of images out of a time series according to the crop-calendar of that region is beneficial and gives in general more accurate results than using all of the images. This is due to the fact that some fields are covered by different types of crops during the year and such sequence is often hard to model because the decisions of individual farmers are governed by a variety of phenological, ecologic, and economic constraints. The latter might be influenced either from sudden change of global or national economic conditions (e.g., oil prize, taxes, and subsidies), the farmer’s personal entrepreneurial strategies, or both. In this paper, results obtained from multi-temporal classifications of TerraSAR-X image pairs (HH and VV) covering a whole season are presented. A standard pixel based Maximum Likelihood classification technique was applied, which has been amended by the use of regional crop calendar to account for seasonal variations of specific cultivations with respect to permanent crops. Results obtained have been compared to ground truth, which have been gathered in-situ in parallel to the satellite measurements. It is shown that even when using all images of the year (i.e., not considering crop calendar) a considerable classification accuracy of more than 75% can be achieved. This accuracy can be improved by pre-processing, which is demonstrated for different filtering techniques, such as Lee speckle filter, a multi-temporal averaging, and anisotropic non-linear diffusion. The best accuracy is obtained by using appropriate sets of images out of the time-series according to the crop calendar available or by performing a multivariate analysis of the covariance matrix (Factor analysis) of the different acquisitions. Some remaining discrepancies in some species are caused by the structural behaviour of the plants on ground. This is discussed by comparison with close range photos being taken during ground truth collection. As could be demonstrated the use of time-series of images from TerraSAR-X despite of frequent cloud cover offers an excellent tool for monitoring crops and serve as indicator for the estimation of the amount of fertilizers used within that area. Using this information, farmers could improve their efforts in establishing good agricultural practice, as being claimed by recent legal and environmental jurisdiction.