Davidson.Edited.html

Malcolm Davidson
Roland Steingießer
Institut für Pflanzenbau
University of Bonn
Tel: 0228-733186
Fax: 0228-732870
email: ulp107@ibm.rhrz.uni-bonn.de

1. Introduction

Remote sensing in agriculture has a number of inherent advantages with respect to conventional data gathering methods since it can:

Image a large area in a very short time. Thus the data possess a high degree of actuality and have a high economic value.
Provide area-wise information as compared to the traditional point-wise sampling techniques. The ability to produce crop maps as well as statistics is considered particularly important for crop studies at regional and local levels.
Advance the distribution and timeliness of the thematic information that is essential for the planning, policy making and management of agriculture.
Act as a stratifier in that it can help stratify large regions according to land-use or agroclimatological conditions.

While active radar systems have demonstrated considerable potential in collecting information over agricultural lands, the full-potential of SAR systems has not yet been realized due to the low dimensionality of the data recorded by existing earth-orbiting SAR systems such as ERS-1 and JERS-1. These operate at a single frequency and polarization, which limits their sensitivity to crop geometry and hence crop type, and condition.

The two SIR-C/S-SAR shuttle missions provided the first opportunity to collect multi-frequency multi-polarization images over a number of sites worldwide. While the missions took place at non-optimal times during the growing season (i.e. early and late) for agricultural applications in Europe, a number of images were taken over agricultural areas characterized by distinct climatic conditions and soil types, different crop types or similar crops but at different development stages and subject to varying crop management practices. The research at the Institut für Pflanzenbau has focused so far on using this wealth of data in terms of the two basic aspects of agricultural remote sensing; the recognition of crop types over a large area on the basis of self-similar but distinct backscatter characteristics for each crop of agricultural land-cover type and, the mapping of the biophysical status of certain crop types. Ground truth data was collected in three very different test sites (Oberpfaffenhofen in southern Germany, Oltrepo Pavese in northern Italy, and Matera in southern Italy) during the first mission and in two sites (Oberpfaffenhofen again and Flevoland, Holland) during the second mission. The general aim was therefore to judge the sensitivity of the multi-dimensional SIR-C/X-SAR radar data to the biophysical condition of crops and their utility for large scale agricultural inventories.

2. Work in Crop Classification

So far classification work has focused mainly on data acquired during the April 1994 mission over the Oberpfaffenhofen test site, partially because better crop maps could be compiled for this test site and partially because, until recently, data were lacking for the other test sites. Two supervised classification methodologies have been tested on a single full-polarimetric scene for the area; the maximum likelihood method using the amplitude at the standard linear polarizations (X-VV, C-HH, -HV, -VV and L-HH, -HV, VV) and the maximum contrast method which uses the complex covariance matrices at C- and L-band. Also the information content of each frequency/polarization feature was evaluated by comparing the observed classification accuracy as a function of features used in the classification process. The following list summarizes the main results so far.

Classification accuracies achieved for agricultural plots only (i.e. general land cover categories such as urban areas, forested areas, water and road systems were not considered) were quite high. The main agricultural land cover types at this time were ploughed and seedbed bare soil fields, winter barley, winter wheat, oilseed rape, meadows and set-aside (fallow). While there was some confusion between barley and wheat fields due to their almost identical appearance at this time, average classification accuracies, using the maximum likelihood classifier for the above categories, reached 80 to 90% in a number of cases -- depending on the frequency/polarization features used and the minimum field size considered.
By comparing the classification accuracies as a function of frequency/polarization features used it was possible to derive a figure of merit for each feature and thus assess which instrument configuration was most sensitive to agricultural features. In general cross-pol polarizations contained more information than like-pol features and L-band, with a higher between category dynamic range, was more useful than C-band. Adding X-VV information to an L-band only and C-band only classification also improved classification accuracies by 10.9 and 19.2% respectively.
The maximum likelihood classification methodology, using amplitude only images, proved to be more robust than the maximum polarimetric contrast method, which uses both amplitude images and information provided by the complex correlation coefficient between HH and VV polarizations. A plot of the complex correlation coefficient showed that the latter contains little information on agricultural land cover type, except perhaps to distinguish seedbed from the other categories. This can be explained by the fact that crops at this time have little structure -- they are all more or less at a grasslike stage -- and lack a distinct geometry. It is expected that the HH-VV complex correlation coefficient will play a more significant role when data taken during the second shuttle flight is analyzed since crops with significant structure such as corn, sugar beet and potatoes were fully developed at that time.
The benefits of the added dimensionality of SIR-C/X-SAR data were demonstrated by comparing the average classification accuracies as a function of the number of polarimetric features used for the classification. N.B. that these classification results (from a single image early during the growing season) compare favorably with multi-temporal ERS-1 results, using images taken throughout the growing season.

Number of Pol./Freq. Features	Maximum Likelihood Avg. Class, Accuracy (%)
2	64,3
3	78,2
4	85,4
5	88,4
6	89,3
7	89,3

3. Mapping Crop Status

Much of the recent research has focused on quantifying the sensitivity of various polarimetric features to the biophysical status of certain crop types. A large database was collected during the first and second missions containing the biophysical plant parameters--biomass, height, water content, leaf area index, etc.--and soil parameters--density and humidity--for a selected number of fields in each test area. The suitability of multi-parameter radar instruments such as SIR C/X-SAR was then judged by comparing ground measurements with the backscatter signatures for each field. Initial results are:

No clear correlation between single channel backscatter features (single frequency, single polarization) and crop biophysical parameters was observed.
The use of ratios, especially the L-HH/L-VV ratio, appear to be more promising. The L-HH/L-VV index appears to be sensitive to the presence of vegetation scattering with respect to soil scattering.

For relatively smooth seedbed fields, L-VV is greater than L-HH, as predicted by the small perturbation model. However as more vegetation is added on top (e.g., barely early during the growing season) the interaction component which favors L-HH becomes more important and the L-HH/L-VV ratio tends towards values greater than 1. This is illustrated at the end of this report.

4. Future Work

Over the last few months a nearly complete data set of reach test site has been delivered by JPL and the DLR, posing a tremendous challenge to process and evaluate such large data sets before the end of our project (December 1996). However only a limited number of SIR-C scenes were taken in the full-polarimetric mode and since these allow for a complete analysis of backscatter features, we will concentrate on using these first, and move on to the other ones later. Specific goals for next year are

Inter-test site comparisons of classification results and backscatter signatures for certain crop types.
Analysis of data collected during the second shuttle mission.
Classification of crop and land cover types in an unknown region (i.e., for which no ground truth was collected by our institute). Two candidate areas imaged during the SIR-C/X-SAR mission are:
1. Erdingen (west of Munich) where classification results can be compared to the official agricultural statistics for the region.
2. Straßbourg (France) in collaboration with SERTIT, who would provide us with the crop maps after a classification of the area was carried out.

5. Publications/Proceedings

Docter, K., M. W. J. Davidson, R. Steingießer, and W. Kühbauch, "Multitemporale Pernerkundunglandwirtschaftlicher Nutzflächen," Proceedings of the 1st German-French Colloquium for Remote Sensing , Bonn, January 1995.

Kühbauch, W., M. W. J. Davidson, R. Steingießer, and K. Dockter, "Investigation of the agricultural land use in Italy and Germany by means of the multi-band/multi-frequency SIR-C/X-SAR system," Proceedings of the 1995 International Geoscience and Remote Sensing Symposium (IGARSS) , Florence, Italy, July 1995, pp. 1061 1063.

Davidson, M. W. J., R. Steingießer, and W. Kühbauch, "Exploiting multi-frequency multi-polarization radar images for mapping crop types early during the growing season," Proceedings for the 1st International Symposium of the Retrieval of Bio and Geophysical Parameters from SAR Data for Land Applications , Toulouse, France, October 1995, (in press).

Davidson, M. W. J., R. Steingießer, and W. Kühbauch, "A comparison of two classification methods for mapping crop types early during the growing season using SIR-C/X-SAR data," IEEE Transactions on Geoscience (Special Issue), (in preparation).

In addition, papers are being prepared for IGARSS '96 (USA) and EUSAR '96 (Germany). Other activities have included an information seminar with the regional agricultural office of Fürstenfeldbruck, near Munich, where results from the first mission were presented to staff and those farmers involved with our project. In the future we hope to publish three peer-reviewed papers; one on classification methodology for multi-parameter SAR, one on the analysis of backscattered signatures as a function of crop condition and a final one on inter-test site comparisons and the classification of an unknown agricultural area.

[Ratio L-HH/L-VV]

Table of Contents

Converted to HTML by Alvin Wong, al.wong@jpl.nasa.gov

The Jet Propulsion Laboratory
4800 Oak Grove Drive
Pasadena, Cailfornia 91109