Mapping of next season's crops during the winter using ERS SAR

E. Nezry

PRIVATEERS N.V., Private Experts in Remote Sensing,
Great Bay Marina, P.O. Box 190, Philipsburg, Netherlands Antilles
Phone: +33-5-61545827. Telefax: +33-5-61991724

G. Genovese

JRC/IRSA/AIS, TP440, 21020 Ispra (VA), Italy
Phone: +39-332-785 449. Telefax: +39-332-789 074

S. Rémondiére, G. Aa. Solaas

Eurimage, c/o ESA/ESRIN Data Utilisation Section, Frascati, Italy
Phone: +39-6-94180 686/626. Telefax: +39-6-94180 622

CROP IDENTIFICATION IN EUROPE DURING WINTER

Application of remote sensing techniques for crop identification and crop area estimates in Europe during the winter season is a 'virtual reality' problem, in the sense that no or very low agricultural vegetation cover is actually present on the ground. As a consequence, during most of the season, agricultural targets are principally the soils on which crops will grow later on, and it is not possible to identify directly crop species. For this reason, the acquisition of optical spaceborne imagery in the framework of the MARS project begins only from March on, when crops start developing their plant structure. The first crop area estimates are not available before mid-April, if the weather conditions are favourable.

Nevertheless, earlier estimates can be obtained, taking advantage of the specific sensitivity of the ERS SAR to important soil properties such as surface roughness. These soil properties and their evolution over time are not casual, as far as agricultural surfaces are concerned. Thus, a reasoning and processing methodology has been designed to exploit in a comprehensive way the causal relationships existing between soil properties as observed with ERS during the winter and crop cultivations.

A methodology to identify non-cultivated surfaces and major crop types should be robust regarding influences of meteorological factors on the ERS radar cross-section over time [1], and should be applicable whatever the study site. To this end, it must combine appropriately a multidisciplinary knowledge:

• The sensitivity of ERS radar back-scattering to soil roughness [3] at the beginning of the agricultural season, and, from February on, volume diffusion effects observable on the first growing crops [3,4] should be exploited.

• The agricultural practices and crop calendars at the site scale should be known.

• Both the specific aspects of SAR data processing and the statistical aspects of classification should be mastered.

PRACTICAL ASPECTS

1. Specific SAR data processing (radiometric correction for ground range sampling, filtering [5]) and accurate co-registration.

2. Transformation of the time series into synthetic channels, as decorrelated as possible between each other. In order to preserve the physical and historical information meaningful to the current application, we compute the following parameters for each pixel:

1. Mean backscattered amplitude during the observation period ('backscatter' channel);
2. Range of radar cross-section variation during the observation period ('variation channel);
3. The date at which the backscattered amplitude is at maximum ('history' channel).

These synthetic channels provide us with an understandable picture of the causes and of the history of the ERS radar cross-section of agricultural targets during the observation period. The first two channels provide a representation in which urban, forest, grasslands/bare soils, and agricultural themes can already be separated using an ERS time series [1]. The third channel carries information regarding the history of field preparation and crop phenology, thus allowing crop identification in relation to the crop calendars.

3. Spatial segmentation through isodata clustering of these synthetic channels (a statistical method which can be regarded as one of the most objective and impartial.

4. Final classification is obtained by aggregation of the statistical clusters into crops and land-use classes which are identified through deductive interpretation of the synthetic channels, using simultaneously general knowledge regarding agriculture, physics of interaction, and statistics (examination of the structure of the clusters). In addition, vector masks of forested and urban areas can also be used, either when statistical confusion remains between themes, or to exclude the masked areas from the final classification. Finally, detection of built-up areas, roads and rivers [5] can be done separately.

MARS PROJECT-EXPERIMENT AND RESULTS

A real-time pre-operational ERS experiment has been carried out to provide very early surface estimates of non-cultivated land and major crops over three significantly different study sites (40 x 40 km) of the MARS project during the winter 1994-95.

During ERS-1 Phase F, 19 PRI frames (ascending and descending) were acquired over the three test sites, from 15 Nov. 1994 to 18 Feb. 1995:

- Albacete (Spain): 4 dates, 5 frames
- Bologna (Italy): 7 dates, 7 frames
- Chartres (France): 4 dates, 7 frames.

The ERS SAR data were delivered within an average period of 10 days after acquisition, meeting the time constraints requirements dor data delivery for the Rapid Estimates Activity of the MARS project.

Chartres (France)
This site, located within one of the richest agricultural areas in Europe, is of particular importance for the estimation of cereal production in France. As it is shown in the multitemporal colour composite image (Fig.1), the site is hilly, with altitude variations up to about 400 m, but generally gentle slopes. The synthetic channels carrying the selected information are produced on a per-pixel basis (Fig. 2).

Figure 1. Multidate ERS SAR imagery over the Chartres test site. Processed ERS-1 images: 28 Nov. in red, 4 Jan. in green and 10 Feb. in blue.

Figure 2. Chartres test-site. Synthetic channels, with the 'backscatter' channel in red, the 'variation' channel in green, and the 'history' channel in blues.

The final 8-class classification is presented in Figure 3, where:

• non-cultivated, bare soils and natural vegetation (low in winter) are black;
• forest is green;
• winter wheat on soils prepared before November appears in red;
• winter wheat on soils prepared early in November appears in pink;
• other winter cereals appear in light grey;
• rape seed appears in yellow;
• soil prepared in February for summer crops, as well as field peas and fodder beets, are dark grey;
• urban areas and roads appear in light blue.

Figure 3. Final classification of the Chartres test-site (8-classes).

Examination of the classification shows that it presents a good robustness to the relief characteristics of this site. Table 1 summarises the comparison between surfaces estimates retrieved from this ERS classification and the Spot-based estimates of the MARS project available in May 1995. Since the classes identified by the two classifications are slightly different, classes have been grouped into land-use families. Good agreement is found between the two classifications.

Table 1. Chartres test site. Comparison of early surface estimates using ERS (March 1995) and Spot (May 1995) multitemporal data.

   Crops         ERS-1, 4-dates      Spot, 2-dates
 (areas in ha)   28 Nov.-10Feb.      4 Mar.-5May
--------------  ----------------     -------------
Non cultivated	    14225               11499        
Winter wheat        68178               70622
Other cereals        5357               10015
Rape seed           12834               11360
Summer crops        24368               25784
Non-agriculture     35038               30720

Bologna (Italy)
This test site includes most of the richest Italian agricultural region: Emilia-Romagna. It consists in a flat plain, densely populated, located between the Po river in the north and the Apennines mountains in the south.

Figure 4 is a colour composite image which illustrates the complex land-use fragmentation of this site. In this case, information selection using the synthetic channels simplifies image interpretation for crop identification, and improves the statistical result of the clustering.

Figure 4. Multidate ERS imagery over the Bologna test site. Processed ERS-1 images: 15 Nov. in red, 5 Dec. in green and 11 Jan. in blue.

The final 9-class classification is presented in Figure 5, where:

• non-cultivated soils and fallow are brown;
• natural vegetation, orchards, and vineyards are dark green;
• rice fields are black;
• winter cereals are red;
• sugar beet are pink;
• potatoes, and also corn fields, are light green;
• other summer crops are yellow;
• rivers and main roads are dark blue;
• built-up areas are light blue.

Figure 5. Final classification of the Bologna test-site (9-classes).

Table 2 gives the comparison between surfaces estimates retrieved from this ERS classification and the Spot-based results.

Table 2. Bologna test site. Comparison of early surface estimates using ERS (March 1995) and Spot (May 1995) multitemporal data.

     Crops         ERS-1, 7 dates       Spot, 2-dates
   (areas in ha)   15 Nov.-17 Feb.      21 Mar.-3 May
----------------   ---------------      -------------
Non-cultivated         13029               13844
Rice                     636                 438
Orchards, etc.         23888               29280
Winter cereals         56139               43840
Sugar beets            23005               22295
Potato + maize         10463               10330
Spring crops            8729                7332
Non-agriculture        24109               32640

Although important differences can be noted for non-agricultural surfaces and winter cereals, these results show a good overall agreement for other crops.

Albacete (Spain)
This test site was chosen to evaluate the method in presence of moderate and strong relief. The dominant grey tones in Figure 6 indicate that meteorological conditions were almost constant (presistent drought) during this winter. The coloured areas are mainly due to the effects of tillage and irrigation. Mountainous areas present important confusion due to the changes in look angle (phase F) and in incidence angle.

Figure 6. Multidate ERS imagery over the Albacete test site. Processed ERS-1 images: 6 Dec. in red, 12 Jan. in green and 18 Feb. in blue.

This problem will be of less importance for the future due to the operational scenario of ERS-2.

The final 9-class classification obtained over this site is presented in Figure 7, where:

• non-imaged and masked areas are black;
• non-cultivated soils are brown;
• forest is dark green;
• vineyards and natural vegetation are light green;
• winter cereals (mainly barley) are orange;
• other cereals are in red;
• sunflower (mainly irrigated) is light blue;
• corn fields (mainly irrigated) are yellow;
• urban areas are white.

Figure 7. Final classification of the Albacete test-site (9 classes)

Within mountainous areas, the output of the clustering procedure reflects the misclassification due to strong relief.
Nevertheless, applying the mask used for the Spot data on this site, it is clear that:

• the same problem is met in the same areas using optical remote sensing data;

• although the statistical definition of the clusters is affected by the mountainous areas, the classification procedure performs successfully within areas which present moderate relief.

USER COST

The total user-cost of the project has been traced during the course of the experiment. The overall user-cost for ERS data, personnel and running costs was under 5500 ECU per site.
Nevertheless, in the actual operational ERS phase (repeat-pass orbit), data and processing costs could be considerably reduced if the test sites are chosen to coincide with ERS PRI frames.

CONCLUSIONS

Results show that early estimates of non-cultivated areas and crop areas using ERS data can be carried out in operational conditions, using a methodology which is:

• technically simple: since it does not require technical developments in addition to what is generally available. Moreover, good knowledge and understanding of ERS SAR related physics, and of agronomy, are required;

• robust: the methodology proved robust to the effects not directly related to the agricultural use of soils as well as to the diversity of sites and to the site-landscape complexity. In this view, the selected ERS information relevant to our specific application is stored in the synthetic channels.

• direct: the agricultural targets of interest are identified in the ERS winter time series through their temporal characteristics, using only general knowledge of physics and agriculture. They support an interpretation which is primarily based on a statistical analysis of the remote sensing data, and specifically oriented towards the identification of agricultural themes.

• efficient: since good performances are obtained within a range of geomorphologic conditions which is similar to that on which optical remote sensing is successful.

• the ability of the SAR to acquire images through cloud cover and at night guarantees coverage as planned and thereby improves the operationality of the project as compared with use of optical data.

• the early availability of results allows more efficient use of the emerging agricultural statistics.

Finally, the total user-cost of such a project is particularly competitive, using the actual operational ERS satellites system.

REFERENCES

[1] Kattenborn G, E Nezry & G De Grandi, 1993: High-resolution detection and monitoring of changes using ERS-1 time series, Proc. 2nd ERS-1 Symp., Hamburg 11-14 Oct. 1993, ESA SP-361, Vol.1, pp.635-642.

[2] Kohl HG, E Nezry, M Mróz &H De Groof, 1994: Towards the integration of ERS SAR data in an operational system for rapid estimates of acreage at the level of the European Union, Proc. 1st Workshop on ERS-1 Pilot Projects, Toledo, 22-24 June1994, ESA SP-365, pp.433-441.

[3] Ulaby FT & MC Dobson, 1989: Handbook of radar scattering statistics for terrain, Artech House Inc., Norwood (MA), USA.

[4] Laur H, 1989: Analyse d'images radar en télédétection:Discriminateurs radiométriques et texturaux, Ph.D. dissertation, Univ. Toulouse, 250 p.

[5] Lopes A, E Nezry, R Touzi & H Laur, 1993: Structure detection and statistical adaptive speckle filtering in SAR images, Int. J. Remote Sens., Vol.14, no. 9, pp.1735-1758.