Rice field inventory using AVHRR data

Rice field inventory using AVHRR data

D. Bachelet
Mantech Environmental Technology Inc.
US EPA Environmental Research Laboratory, Corvallis, Oregon, USA

J.L.Mailander
Department of Geosciences, Oregon State University, Corvallis, Oregon, USA

Abstract
Time series Normalized Difference Vegetation Index (NDVI) data, computed from Advanced Very high Resolution Radiometer (AVHRR) data, were used in pilot study to locate area for rice cultivation in the United States of America (USA). The large size of rice fields and the relative phonological homogeneity of the rice growing regions in the US make them ideal sites for a pilot study. NDVI dynamics were examined using 16 km global are coverage satellite data from 1988. Unsupervised classification was used to distinguish rice fields form other vegetation cover types. The technique was used for California where the contrast between irrigated and natural vegetation is the most pronounced and later applied to Louisiana, Arkansas and Texas. Identical methods were used to classify the vegetation in China where the field size is much smaller and the cropping season more extended. The rice NDVI dynamics was most obvious where only one crop is grown and the growing season is limited by low winter temperatures. Areas where several crops are grown each year were more difficult to identify. The effectively assess rice locations, the seasonal fluctuations of the crop, which are only partially dependent on seasonal precipitation because of irrigation, must be isolated from characteristics associated with natural vegetation and other irrigated corps.

Introduction
Global and regional scale ecological research has so far relied upon simple interpretations of land cover and surface properties and coarse resolution global land cover databases (Matthews 1983, Olson and Watts 1982, Wilson and Henderson Sellers 1985). Higher resolution data with greater precision for classification would clearly improve the quality of the research (IGBP 1991, Loveland et al 1991). During the last decade, substantial progress has been made in using National Oceanic and Atmospheric Administration (NOAA) Advanced Very High Resolution Radiometer (AVHRR) data for land cover characterization (Goward et. 1985, Tucker et al. 1986, Roller and Colwell 1986, Townshend et. al 1991, Lloyd 1991, Loveland et al 1991). AVHRR data have moderate spatial resolution (1 km) but excellent temporal resolution (twice a day) which increase the likelihood of finding cloud-free observations during specific periods and makes it possible to monitor changes inland cover over periods as short as a growing season (Kennedy 1989, Johnson et al. 1987, Malingreu 1986. Justice et. al. 1985, Goward et al. 1985). Applications of AVHRR data have focused on the analysis of vegetation greenness which is most often characterized using a vegetation index such as the normalized difference vegetation index (NDVI) (Goward et al. 1985). Because of limitations in data availability, regional and global scale analyses have used resampled data in which either 4 or 16 km pixels aer combined (Global Vegetation Index or GVI). High Resolution (1 km) data are readily available for the USA but not yet available for Asia.

Rice is unique among grain crops in that it is generally flooded for a length of time at the beginning of the growing season. The NDVI signal should thus be low during flooding conditions, and much greater as the rice plant emerges from the standing water and eventually reaches peak vegetative growth. Malingreau (1986) plotted GVI signature of wetland rice in Asia using time series (1982-1985) AVHRR imagery. His results showed the applicability of coarse –scale imagery to describe the dynamics of an agricultural landscape. The objective of this paper was to determine the adequacy of AVHRR to identify rice growing regions at large scale. We used low resolution AVHRR data to identify rice growing areas in California and the Gulf States of the united Stated where locations and timing of the rice cultures are well known. We then attempted to extend the technique to the rice growing regions in the Hunan and Yangtze-delta of China where we had only few indications of the exact location of rice paddies.

Materials and methods

Area of study
The area of study included California and the Gulf states of Mississippi, Missouri, Louisiana, Arkansas and Texas, which are the primary rice growing states of the USA. Rice fields in California are flooded between April and early May before planting. The fields remain flooded for the entire growing season and are only drained about 10 days before harvest (Chang 1971). The rice plants emerge about 10 days after planting, flower about 80 days later, and are harvested in October approximately 150 days after planting (Nuttonson 1981). Rice fields in the Gulf states and Arkansas are grown under more varied water regimes. Rice fields in the Gulf states and Arkansas are grown under more varied water regimes. Commonly rice is planted between April and June in flooded fields. The fields are drained a few days later and flooded again following the emergence of the rice plants, some 10 to 12 days later. Harvest usually occurs late August or September.

The area of study also included the Jiangsu, Anhul and Zhejinag provinces which are located in the Yangtze river delta in Western China (from 35°30’ to 26°50’ North latitude, and from 114°30’ to 122°40’ East longitude) and the Hunan province). Two types of rice crops are cultivated there. In a double cropping system, the first crop of rice is planted between March and April and harvested between June and August; the second crop of rice is planted between June and July and harvested between October and November. Alternatively, a single crop is planted between April and June harvested between August and October (USDA 1987). This single crop is typical of the northern Yangtze delta area (USDA 1987).

AVHRR data
The global vegetation index (GVI) satellite imagery that was used in this study was obtained from the US Army Corps of Engineers Construction Engineering Research Laboratory for the year 1988 at a weekly time step. The data was originally from the National Oceanic and Atmospheric Association (NOAA) satellite series which carries the advanced very high resolution radiometer (AVHRR) and generates daily coverage. The GVI is essentially identical to the normalized difference vegetation index (NDVI), the “greenness” index, brought up to a 16km spatial resolution and one week time step using a maximum value compositing procedure (Holben and Fraser 1984, Holben 1986). Weekly GVI data were composited using the maximum value procedure to produce monthly and tri-weekly maxima.

Image Classification
Data processing was done in GRASS, a geographic information systems software package (Army Crops of Engineers 1988). The First step of our classification method was to collect raster map layers in an imagery group which enabled the analysis on any combination of the raster map layers in that group. The maps we chose were time series of AVHRR images. Since we did not a priori ground-truth information, we performed an unsupervised classification whereby the computer analyzed the image data and grouped pixels which were spectrally similar.

Unsupervised classification partitions the data space in the absence of any ground truthing information. The method used in the GRASS software is a clustering algorithm. The program first reads the entire dataset and sequentially identifies the natural groupings based on a measure of the Euclidean distance calculated for each pixel using the spectral properties of the data (i.cluster). Secondly, a maximum likelihood classifier uses the cluster means and covariance matrices previously generated to determine to which spectral class each pixel in the image has the highest probability of belonging (i.maxlik). This maximum likelihood classification technique ensures that all the pixels are classified. Subsequently, ground truthing can be used to identify the landcover type for each group identified.

Results
California and the Gulf states :
Using AVHRR monthly composite images, we ran the clustering and maximum likelihood algorithms to classify the monthly images. After seven iterations (out of 30 possible). 15 classes were identified. The NDVI signature for each class was then plotted against time and compared with rice phonology. One of the classes (class 10 in figure 1) seemed the most likely to represent paddy rice with low NDVI early (January to June) and late (November and December) in the year. The same analysis was performed using 3 week composites (not shown here) and the signature was clearer with a sharper peak in the 11th week (around July 30) and values less than. 1 from week 14 (October 1) on. This would agree well with a peek greenness period from July 20 to August 2 and the post harvest bare soil from October 12-25 on.

Figure 1 1988 NDVI development curve for the 15 classes resulting from the unsupervised classification of 12 monthly 8.6 min. resolution AVHRR composites for California, USA. Units are NDVI*100. The region analyzed extends from 42:30N to 31:55N latitude and from 124:45W to 113:45W longitude.
Sutter, Colusa, Butte and Glenn counties in California contain the largest acreage of rice in California (Rice Journal 1992, California State Statistical Office Pers. Comm.). The field sizes are relatively large and the skies are generally clear. Assuming that class 10 was representative of rice, we calculated that slightly more than 50,00 acres are planted to rice which agrees well with reported acreage (Rice Journal 19920. unfortunately, this classification greatly overestimated the area planted to rice in Merced, Fresno and Kern counties, probably because of the presence of other irrigated crops. We also used tri-weekly instead of monthly NDVI averages to try to better separates the rice signature from other crops with similar Phenology. Again we overestimated rice acreage rice in the southern counties of Fresno, Kern, Tulare and Kings.

We plotted NDVI signatures obtained for the Gulf states against time and one of the classes (class 7 on figure 2) appeared to match known rice phenology with low NDVI from January to July (flooding occurs in June), peak greenness in August dropping rapidly to a value less than. 15 in October (the bulk of the harvest occurs in September). However, this signature could also apply to other irrigated cultures. Loveland et. al (1991) classified the land cover corresponding to our class 7 as mixed cultures with rice.

Figure 2 1988 NDVI development curve for the 15 classes resulting from the unsupervised classification of 12 monthly 8.6 min. resolution AVHRR composites for the Gulf Stages (Missouri, Arkansas, Louisiana, Mississippi, Texas) USA. Units are NDVI*100. The regions analyzed extends from 38:00 to 2500N latitude and from 100:00W to 87:00W longitude.
Hunan province and Yangtze delta region:
In 1988, the Hunan province had the greatest area planted to rice (4.3 106 ha for a total area of 21 106 ha ie. 20%) and the greatest rice production (23 106 t) in China (IRRI 1991). We performed the image analysis for this area, identifying 15 classes. Overlaying the results of the unsupervised classification on to the map a elevation (Edwards 1986) for the province (figure 3) showed that classes 11 through 15 correspond to high elevation areas and are probably forest vegetation. Classes 1 and 2 correspond to the Xangliang river valley and are probably irrigated crops. Several irrigated crops are grown along the river and it remains difficult to separate the rice NDVI signature from these low elevation areas (figure 4). The same analysis was performed using tri-weekly composite images instead on monthly composites (not shown here). Results did not show any significant improvement in separating the rice signal from other irrigated cultures.

Figure 3 Classified map for the Hunand province, China obtained using the unsupervised classification of 12 monthly 86. min. resolution AVHRR NDVI composites for 1988. elevation is used to display the map in 3 dimensions. The region analyzed extends from 30:50N to 24:42N latitude and from 114:00E to 108:67E longitude. The highest peak on the left of the map is 2797m high and on the forefront, Baaimaashan reaches 1963m.

Figure 4 1988 NDVI development curve for the 15 classes resulting from the unsupervised classification of 12 monthly 8.6 min resolution AVHRR composites for the Hunan Province, China. Units are NDVI*100. the region analyzed extends from 30:50N to 24:42N latitude and from 114:00E to 108:67E longitude.
In the Yangtze delta, the unsupervised classification highlighted 5 regions (figure 5). First, it identified a high elevation vegetation type corresponding to the Kuocangshan range (in the southeastern part of the Zhejiang province). These classes, 12 through 15, probably correspond to wooded areas. Secondly, the Xingjiang and the Yangtze valleys correspond to class 1 probably representing water (NDVI <.1) which is most evident in the Pyanghu area. Thirdly, the southern part of Jiangsu where rice culture is most extensive, corresponds to class 9. Fourthly, the southern part of Anhui (classes 4 and 5) and finally the northern part of Anhui (classes 7 and 8). Again if broad pattern emerge, ground truthing would be essential to separate irrigated cultures from pure rie (figure 6). Triweekly images display greater variability in NDVI dynamics but again the rice signal does not appear more clearly.

Figure 5 Classified map for the Yangtze delta region, China, obtained using the unsupervised classification of 12 monthly 8.6 min. resolution AVHRR NDVI composited for 1988. Elevation is used to display the map in 3 dimensions. The region analyzed extends from 35:50N to 26:83N latitude and from 114:50E to 122:67E longitude. The highest peak on the right side of the map is 1849m high.

Figure 6 198 NDVI development curve for the 15 classes resulting from the unsupervised classification of 12 monthly 8.6 min. resolution AVHRR composites for he Yangtze delta region (Anhui, Zhejiang and Jiangsu, china. Unites are NDVI*100. The region analyzed extends from 35:50N to 26:83N latitude and from 114:50E to 122:67E longitude.
Discussion
The effect of aggregation and sampling has been described in detail by Malingreau (1986, 1980). He observed that “despite such drastic sampling (10m for the size of the rice field and 15,000m for a pixel), the GVI data closely and consistently reflect the dynamics of a paddy landscape”. Because he concentrated his study on single pixels, where he knew rice was grown, he obtained a relatively clear signal reflecting the phytodynamics of the landscape. He recognized that “because of ground resolution constrains, the NDVI curves integrate various aspects of a paddy landscape. Not all fields are planted at the same time and undergo the same development rates”. Using a simple classification method over large areas, we were able to identify broad patterns such as irrigated vs dryland agriculture, woodlands but were unable to identify pure rice cultures.

Malingreau (1986) further observed that tri-weekly NDVI seems most appropriate “to monitor fast maturing crops in the tropics”. We used both monthly and tri-weekly images. We thought higher resolution images (1.1 km) would produce more reliable estimates of 1988 rice acreage in California. We compared our results with Loveland et al. (1991) descriptions of land cover types. They used high resolution AVHRR (1.1 km) images and identified irrigated cultures in California that matched the location of our class 10. We believe that we over estimated rice acreage because we could not separate other irrigated cultures from rice. Since the resolution was coarse (the minimum size of a pixel is around 13,500 acres) only high density of rice fields could be picked up by the classification method. We were hoping to be able to clearly identify the high density of rice fields along the Mississippi river and in the Yangtze delta region. NDV dynamics appeared more jugged with higher amplitudes and more frequent and pronounced drops due to the presence of clouds especially in the Hunan province using tri-weekly images. It did not enable us to clearly identify the rice signal form the other vegetation signals. This was particularly true in China, where several planting dates are only separated by a few weeks and thus complicate the signal.

Since existing global databases are coarse, AVHRR data could be used to refine our understanding of vegetation distribution over large regions. These data greatly improve our ability to identify areas of interest, whether they are woody mountainous vegetation or valley bottom irrigated crops. These data also enable us to distribute better the acreages of a particular vegetation type or crop into the spatial extent of a given landscape.

Conclusions
We believe that Malingreau (1986) may have been overly optimistic about the usefulness of low resolution AVHRR. However, these data are clearly to determine broad patterns of vegetation distribution. Groundtruthing and/or a good knowledge of the area of study would also greatly improve the ability of the image analyst to identify a particular vegetation type. Smaller time intervals between images sharpens and NDVI dynamics and should, in principle, allow for better recognition of a particular pattern (of ex. rice phenology). But in a complex agricultural setting such as China where several crops overlap and where the rice growing season corresponds to the rainy season (high cloud cover), satellite imagery does not solve all the problems. Even a landcape where fields are large and skies are clear, good knowledge of the ground cover is till necessary to separate rice from other irrigated locations where individual crops, such a rice, are grown. However, global coverage is unavailable and in Asia low resolution AVHRR remains the best tool available.

References

Army Corps of Engineers, 1988. GRASS Users and Programmers Manual. US Army Corps of Engineers, Construction Engineering Research, Laboratory, Campaign, Illinois.
Chang, S.S.T 1971. The role of water resources in the wet rice agricultural system of the Sacramento valley, Ph.D Dissertation, U.C. Los Angeles, 285pp.
Edwards, M.H. 1986, Digital Image processing of local and global bathymetric data. MS Thesis, Washington University, Department of Earth and Planetary Sciences, St. Louis, Missouri, 106pp.
Goward, S.N. Tucker, C.J., Dye, D.G. 1985. North American Vegetation Patterns observed with the NOAA-7 advanced very high resolution radiometer, Vegetation 64:3 – 14.
Wilson, M.F. and Henderson-Sellers, A. 1985. A. Global archive of land cover and soils data for use in general circulation models. Journal of Climatology 5:119-143.
Holben, B. 1986 Characteristics of maximum value composite images from temporal AVHRR data. International Journal of Remote Sensing 7:1417-1434.
Holben, B., and Fraser, R.S. 1984 Red and bear-infrared response of off-nadir viewing. International journal of Remote Sensing 5:145-160.
International Geosphere-Biosphere Programme 1990. Global Change : Report No. 12, Stockholm, Sweden.
International Rice Research Institute (IRRI). 1991. World Rice Statistics 320 pp
Johnson, G.E van Dijk A., Sakamoto, C.M. 1987. The use of AVHRR data in operational agricultural assessment of Africa, Geocarto International 1:41-60.
Justice, C.O., Townshend, J.R.G. Holben, B.N. and Tucker, C.J. 1985 Analysis of the phenology of the global vegetation using meteorological satellite data. International Journal of Remote Sensing 6:1271-1318.
Kennedy, P. 1989. Monitoring the vegetation of Tunisian grazing lands using the normalized difference vegetation index Ambio 18L119-123.
Lloyd, D. 1991. A Phenological Classification of Terrestrial Vegetation using shortwave Vegetation Index Imagery. International Journal of Remote Sensing 11:2269-2279.
Loveland, T.R., Merchant, J.W. Ohlen, D.O., Brown J.F. 1991. Development of a Land-cover characteristics of database of the conterminous U.S. Photogrammetric Engineering and Remote Sensing 57:1453-1463
Malingreau, J-P 1986 global vegetation dynamics : satellite observations over Asia. International Journal of Remote Sensing 7: 1121-1146
Malingreau, J.P. 1980. The wetland rice production system and its monitoring using Remote Sensing in Indonesia, Ph.D. Dissertation, U.C., Davis 350.pp.
Matthews, E.1983 Global Vegetation and land-use: new high resolution data bases for cliame studies. Journal of applied Meteorology 22:474-487.
Muttonson, M.Y. 1981. A study of the consistency of the phenology of rice with special reference to the central valley of California. A Report-submitted to Dr. John Estes, Director, geography Remote Sensing Unit 42 pp.
Olson, J.S and Watts, J.S 1982, Major World Ecosystem Complexes Map, Scale = 1:30,000,000 : Oak Ridge National Laboratory, Oak Ridge, Tennessee.
Rice Journal, 1992, Rice industry Guide 199. 95:33-39.
Roller, N.E.G. and Colwell, J.E. 1986. Coarse resolution satellite data for ecological surveys. Bio Science 36: 468-475.
Townshend, J., Justice C., Li. W. Gurnery, c., and McManus, J. Global Land cover Classification by Remote Sensing: Present Capabilities and future possibilities. Remote Sensing of environment 35:243-255.
Tucker, C.J. Justice, C.O. and Prince, S.D. 1986. Monitoring the grassland of the Sahel 1984-1985 international Journal of Remote Sensing 7:1571-1581.
United States Department of Agriculture (USDA). 1987 Major World Crop Areas and Climatic Profiles. Agriculture Handbook No. 664. 159 pp.