EXAMINING AGRICULTURAL AND WETLAND VEGETATION USING ERS-1 IMAGERY

ABSTRACT

This study examined the use of multi-date ERS-1 imagery to identify crops and wetland vegetation within the Rainwater Basin in south-central Nebraska. The Rainwater basin is comprised of a dense cluster of wetlands that have been greatly modified by agricultural production. For conservation purposes techniques are needed to monitor the relationship between existing wetlands and the surrounding agricultural land use practices. The use of multi-date ERS-1 imagery was effective in dinstinguishing between agricultural and wetland vegetation. Temporal curves of radar backscatter were constructed from 150 field observations, which showed a pronounced seasonal variation in backscatter associated with crop phenology. Differences in backscatter among crop types was most pronounced in early and mid-summer images. Statistical analysis suggests that the sensitivity of radar to moisture conditions was great enough to detect a difference between crops grown on hydric versus upland soils. An unsupervised classification was performed to map the spatial distribution of crops and wetlands. Results suggest that classification performance improves when vegetation classes are grouped into structural classes.

INTRODUCTION

The Rainwater Basin is a mosaic of wetlands and agriculture stretching across 17 counties in south central Nebraska. The region provides habitat for waterfowl and is a focus of spring waterfowl migration. During the past century, conversion of wetlands to agriculture has dramatically altered this landscape and resulted in a loss estimated at 80% of the historic wetlands (Figure 1).

Figure 1. Estimated acres of wetlands in Rainwater Basin. Source: Nebraska Game & Parks, 1992.

Most wetlands in the Basin are shallow and small, with size dependent on recent meteorological conditions. For conservation and habitat restoration, new techniques are needed to reliably map such small yet important wetlands. This study examined the use multi-date ERS-1 Synthetic Aperture Radar (SAR) satellite imagery to identify land cover (crop type and wetlands) with radar data alone. Six SAR images were acquired covering the Rainwater Basin during the 1995 growing season (May - September). These images were used to identify typical rural land cover types such as agricultural crops (corn, soybean, milo), wetlands, and grasslands. Further analysis was conducted to test whether the sensitivity of radar to moisture conditions was great enough to detect a differnce between crops grown on hydric versus upland soils. This information would suggest a means for identifying historic wetlands that are currently under agricultural production.

The use of Synthetic Aperture Radar (SAR) to identify agricultural land cover and other vegetation characteristics has been investigated to some extent in the last two decades. Two well-documented advantages of imaging radar are cloud penetration and the capture of information about vegetation canopy structure. The strength of the SAR return is affected by target morphological factors such as surface slope and roughness, moisture content, and molecular structure (Lillesand, et al. 1987). Thus, SAR images capture distinctly different but complementary information compared to traditional V/IR images.

For agricultural applications, the radar backscatter has been found to be dependent on plant species and age, which generally determine plant morphology. For example, sugar beets were found to give a high return due to their high water content and large leaves (Bouman, et al. 1993) In general, broad leaf crops produce a higher signal return than other crops for C band with VV polarization. VV polarized SAR reflectance is predominately caused by the physical structure of the plant canopy rather than understory characteristics (Holmes, 1990).

As with the initial applications using Landsat visible and infrared images, initial ERS-1 SAR researchers have attempted to map various crop types. Bouman and Uenk (1993), used simulated multidate ERS-1 radar to examine SAR backscatter changes over agricultural fields. They reported that a combination of early and late season SAR images provided the best discrimination between crops. Schmullis et al. (1994), reported that ERS-1 SAR returns from different fields of the same crop were more variable in the early season and converged with increasing canopy cover. Early season differences were attributed to differences in soil roughness. Thus, crop identification should be conducted using late summer data when the SAR reflectance is more affected by crop canopy than soil surface effects.

ANALYSIS OF RADAR BACKSCATTER

Six ERS-1 images were acquired between May and September 1995 covering the eastern part of the Rainwater basin in Clay County, Nebraska. These images were rectified into the UTM coordinate system using nearest neighbor resampling to 30 m resolution and assembled into a single multiband image data set where each band is a single date of imagery. Four images were from the descending ERS-1 orbit (approximately 10:30 pm EST) and two images from the ascending orbit (approximately 10:30 am EST).

One problem inherent in most SAR images is speckle, or noise, which is the result of echoes from the target surface (Leberl, 1990). This speckle is usually treated as a random effect and reduced by averaging adjacent pixels using spatial convolution filters. In this study, a 3 x 3 Sigma filter was used to reduce image speckle while retaining edge information such as field boundaries (Lee, 1981). The results of the filter were successful in reducing the heterogeneity on digital number (DN) values within agricultural fields, while preserving field boundaries.

Field observations were made in the study area to record crop type and condition for over 150 fields. The field boundaries were identified on the image data set and digitized to produce training polygons. This information provided the base for constructing temporal profiles of radar backscatter for crops and wetlands in the scene. Temporal signatures of radar backscatter were extracted for each of the 150 fields.

Histograms were used to examine the distribution of backscatter for the entire image and for individual fields. The raw data showed a high variance for the May 6, July 7 and September 23 images, although the distributions were remarkably normal. Graphs produced with the original DN values were influenced by extreme differences in standard deviation, which made comparison of profiles between agricultural fields difficult. To account for the difference in variance between dates the data was standardized to Z-scores and a scalar was added to return all Z-scores to a positive range. Inspection of covariance and correlation matrices indicated a strong positive relationship between brightness values on the July and August images.

Temporal curves were constructed to examine seasonal trends in backscatter. The mean value of backscatter was plotted using normalized Z-scores. The data revealed a strong temporal pattern of backscatter for vegetation throughout the growing season. A separation between agricultural fields, wetlands, and grasslands was apparent for some dates.

The phenological cycle of vegetation becomes important in interpreting the graphs. The first image was acquired in early May during a very wet spring, and prior to most planting. Difference in backscatter from this date are attributed mainly to differences in soil and moisture content. The second image was acquired in mid June. Due to a wet spring the planting was delayed and while agricultural vegetation was emergent, the backscatter was still influenced by the soil background. Winter wheat provided an exception to the general pattern in that the crop was mature in mid - June and close to harvest. The third image taken on July 7th consistently provided the greatest magnitude of backscatter as shown by the peak in the curve for all land cover types. This date also provided the greatest separation between crops. The temporal curves begin to converge during late July as the canopies thicken although there is still distinction between crops. The fifth image acquired in late August represented a full canopy for most crops and showed the least distinction in backscatter. The final image in September shows an increased variation in backscatter which could be attributed to the senescence of vegetation and the variation in harvesting schedules among crops.

In examining the temporal pattern of backscatter among crops a general pattern emerged as well, yet the temporal difference between crops was not as dramatic as the temporal difference between land cover classes. There was high variation between fields during the early image dates that is attributed to the soil background. Once vegetation was emergent, Corn showed the highest backscatter followed by Milo, Soybean, and Pasture. The stubble from winter wheat provided a consistently high backscatter. As the season progressed the separation between crops was most significant in early July. The variation in backscatter among crops was less in late July, and there was little distinction between crops by the late summer.

STATISTICAL ANALYSIS of BACKSCATTER DN VALUES

To evaluate the influence of the soil background on the radar backscatter over time an analysis of variance was conducted using 50 corn fields. Using the digital field boundaries for corn fields a GIS map overlay was done with a digital soils map to identify the soil type associated with each agricultural field. The corn fields were then seperated into those that were grown on predominately hydric or non-hydric soils. The digital soils data was based on county soil surveys for Clay County, Nebraska. Across the study area there are four soils that have a hydric component (Massie, Fillmore, Butler, Scott). Of these, Fillmore soils have been altered most often for agricultural production. Table 1 shows the results of a two way ANOVA using a repeated measures design to test the hypothesis of no significant difference in the backscatter of corn on hydric versus non-hydric soils.

Table 1. Comparison of corn backscatter on hydric versus non-hydric soils using a Two-way ANOVA.

The probability of an F-value > 7.11 is 1% when Ho is true. Thus, we reject the hypothesis and conclude that a difference exists between the mean backscatter of corn fields on hydric versus non-hydric soils. By plotting the mean backscatter on hydric versus non-hydric soils the backscatter was consistently higher on hydric soils. A comparison of means using a protected Fisher's LSD test was then done to assess at which dates a differences among means could be detected.

Prior to the experiment it was thought that a difference in backscatter when comparing crops based on soil type should be most prevalent in the early season images when the soil background is dominant. By late in the summer the full canopy among corn fields should masks the influence of the soil and the backscatter should be similar regardless of soil type. This leads to expectations that the mean DN values should show significant differences in the early dates but not in the later ones. The results suggest that this pattern was generally detected with the available data set. The late season images acquired in August and September showed no difference in mean backscatter, while a differnce was detected in the images in June and early July where the soil background would still be present. The first image taken prior to planting in early May is unusual in that no difference in backscatter was detected for that date either. This image was acquired well before planting occured in this region, when the radar backscatter should be influenced by the soil background and moisture conditions. The failure to detect a difference at this date may be attributed to several factors including: the climatic conditions prior to the satellite pass and the type of statistical test choosen to evaluate means along a temporal profile.

Table 2: Results of Fisher's protected LSD test were used to evaluate the effect of time on mean backscatter. Note the astrik (*) denotes a significant difference for that date.

CLASSIFICATION OF MULTI-DATE IMAGERY

Classification was done using the multi-date data set to determine if landcover classes were easily identified. The research considered two classification schemes the first approach was a traditional unsupervised classification where the image was partitioned into agricultural and major land cover classes that included: corn, wheat, soy beans, alfalfa, milo, pasture, wetlands, and grasslands. The second approach was to emphasize the structural components of these vegetation classes and to group them based on their structural similarities. This approach involved in a reduction in the total number of classes from eight to four. The overall map acurracy for the structural approach was greater than 80%. These results suggest that the map accuracy will improve when considering structural similarities in vegetation classes.

CONCLUSIONS AND FUTURE RESEARCH

This research indicates that multi-temporal radar can be used to track temporal changes in backscatter that are related to differences in crop phenology. Results suggest that the strongest distinction between crops occurs during the early to mid growing stages. Once crops reach a full canopy the ERS-1 C-band sensor appears to provide limited penetration of the vegetation canopy. This trend has been reported in other agricultural studies as well. The extreme variation in backscatter in early season images appears to be influenced by the soil background. Other factors include the differences in planting schedules that occurred due to wet spring weather. A clear distinction was present between major land cover types (crop, wetlands, grassland, etc.). This suggests that multi-date radar has potential for use with regional land cover mapping. Future work will explore the use of multi-date radar for classification of crops and wetlands in this region.

The significant F-value suggests that the ERS-1 data has potential for identifying hydric soils in early growing season before the crop canopy masks the soil background. Within the Rainwater basin where many of wetlands have been drained and the land has been converted to agricultural production, techniques are needed to detect and map historic losses of wetlands. Identifying and estimating the amount of crop acreage that is grown on hydric soils could prove useful as a conservation tool in assessing the impact of agricultral production on the historic extent of wetlands in this region.

ACKNOWLEDGMENTS

ERS-1 SAR images were provided by the European Space Agency (grant USA/124). Additional support was provided by the CALMIT lab at the University of Nebraska, and the Geography Dept. at CSU-Sacramento. Pat Starks, USDA provided Soils data. Additional assistance in field data collection and computer processing was provided by Brian Leavitt and Anthony Militar.

REFERENCES

Lillesand, T.M. and R.W. Kiefer, 1987.

Remote Sensing and Image Interpretation, John Wiley & Sons.

Bouman, B.A.M. and D.H. Hoekman, 1993.

Multi-temporal, multi-frequency radar measurements of agricultural crops during the Agriscatt-88 campaign in The Netherlands. Int. Journal of Remote Sensing. 14(8):1595-1614.

Holmes, M.G., 1990.

"Applications of Radar in Agriculture," Ch. 19 in Applications of Remote Sensing in Agriculture, M.D. Stevens and J.A. Clark, Eds. Butterworth Press.

Bouman, B.A.M. and D. Uenk,1992.

"Crop classification possibilities with radar in ERS-1 and JERS-1 configuration," Remote Sensing of Environment. 40:1-13.

Schmullius, C, Nithack, J. and M. Kern. 1994. Comparison of multitemporal ERS-1 and

E-SAR image data for crop monitoring. Earth Obs. Quaterley. 43:9-12.

Leberl, F.W., 1990.

Radargrammetric Image Processing, Artech House.

Lee, J.S., 1981.

Speckle analysis and smoothing of synthetic aperture radar images. Computer Graphics and Image Processing. 17:24-32.

Keywords: ESA European Space Agency - Agence spatiale europeenne, observation de la terre, earth observation, satellite remote sensing, teledetection, geophysique, altimetrie, radar, chimique atmospherique, geophysics, altimetry, radar, atmospheric chemistry