GISdevelopment.net ---> AARS ---> ACRS 1998 ---> Poster Session 1

Combined Use of SPOT and GIS Data to Detect Rice Paddies

Lau, Chi-Chung Shiao, Kao-Hsing
Energy and Resource Laboratories
Industrial Technology Research Institute (ITRI)
W000 ERL/ITRI, Bult. 24, 195-6, Sec. 4, Chung Hsing Rd. Chutung, Hsinchu,
Taiwan 310
E-mail: f820027@erl.irti.org.tw


Abstract
Multi-temporal SPOT images combined with cadastral GIS data were analyzed for detecting rice crops in Chang-Hwa county on central Taiwan. Pixels inside one parcel were assigned to same class by rues using NDVI. Because typical NDVI values of rice paddies change from a low value on transplanting period to high value on reproducing period, and decline again to harvesting period. This pattern prompts a rule-base to distinguish rice paddies from evergreen and other man-make backgrounds. Results show that using one image with GIS makes classification accuracy up to 89%. Using Multi-temporal images increases accuracy to an operational level (92%). Two images, one image may comes from reproducing stage, are preferred for applying the rule-based detecting approach.

Introduction
For food supply and agricultural planning, measure of rice production at an early stage is very important in Taiwan. Crop Bureaus of Taiwan purchases rice to stabilize the food market and maintaining rice inventory in a minimum level. In order to estimate purchasing cost, better measure of rice yield can obtain if he growth of the crops is being monitored during the growing season. For years, the Crop Bureaus has used aerial photography collect the information of area of rice-paddy for calculating rice yield.

Photogrammetric interpretation is a labor-intensive and costly work. The use of remotely sensed data seems a reasonable alternative. However, previous studies show that these data alone do not provide the accuracy or specificity required for rice-paddy detecting. Frequent cloud cover cause causes other problem on data acquisition. There are two cropping in Taiwan and each growing season takes about four months. Good quality image may not available at the best time for detecting rice paddy. This paper investigates the incorporation of multi-temporal spatial data through a geographic information system (GIS) to improve the accuracy and specificity of classification derived from satellite data. The final result is a set of decision rule which define the cropping status of rice paddy.

Background
Detecting of rice paddy on Taiwan faces two problems included vegetation background and size of paddies. Tropical climate causes abundant vegetation that confuses the delineation of rice paddy. The small sizes of paddies cause mixed pixel on satellite image. Both problems induce bad accuracy and specificity of a classification that cannot be accepted by the Crop Bureau. In this study, two approaches were adopted to surmount above difficulties: using multi-temporal images and integrating cadastral GIS data and SPOT imagery.

Crop growth makes the agricultural land exhibiting temporal change that shown in satellite remote sensing images. Tasseled Cap has illustrated the periodic pattern. (Kauth and Thomas, 1976, and Crist and Cicone, 1984). Vegetation indexes, such as normalized difference vegetation index (NDVI) and GREENNESS (from Tasseled Cap Transformation), were derived as a measure for crop growth. During a growing season, paddy's NDVI is low when the field imitated with water at the initial time called transplanting period. When growth stage comes to reproducing period, rice's biomass goes to higher level and NDVI get a maximum level too. After the harvesting period, NDVI returns to a minimum level with paddy turns to bare soil. Comparing with paddy's NDVI, natural vegetation background shows a constant value. Difference between two images can distinguish the two classes.

Introducing GIS information into detection process increases classification and spatial accuracy. GIS data cluster mixed pixels in same category. For example, rice-paddy's cadastre is considered as a boundary confined the is-rice pixel. Accurate spatial distribution of rice-paddy and their production provide necessary census information.

Integration methods can be generalized into three categories : 1. Pre-classification stratification. 2. Classifier modification, and 3. Post-classification (Harris and Ventura, 1995). Cadastral parcel provides base to dividing scene into smaller areas. This enables land covers that are spectrally similar to be classified independently. Classified modification changes the priori probabilities of classes based on a known database. Post-classification allows individual pixels to be refined based on decision rules derived from the GIS data. To simplify the study processes, pre-classification method is presented in this study and the other method may considered in later research.

Data Processing
The Chang-Hwa county covers 98,186 hectares and 73.84% belonged to agricultural land, which makes Chang-Hwa county a major agricultural county of Taiwan. Rice is the county's major crop but it also cultivates other fruits, vegetable and flower. A 5000 hectares test site in the north-western corner of the county was selected where the cultivating pattern is comparatively simple for studying.

According to the definition from Crop Bureau, there are two rice cropping in Taiwan. First cropping covers period between the end of winter and harvested in summer. Second cropping transplants 15-20 days after first cropping and harvests in late autumn. The first cropping takes 110-140 days and the second cropping takes 100-110 days due the difference of temperature and illumination.

Growing stages decide the selection of images. A simplest way divides the whole period into three intervals. 1. Sowing-transplanting period. 2. Growing period, and 3. Fallow period. In Taiwan, the growing period can be separated into (1). vegetative stage, (2). reproductive stage, and (3). ripening stage. For propose of appropriately defining the change of rice reflectance, a five stages scheme was adopted in this study: 1. Transplanting, 2. Growing, 3. Reproducing, 4. Mellowing, and 5. Harvesting. According to the local growing schedule, 1997 second cropping of Chang-Hwa county was transplanted from July 12 to August 8, and harvested from November 18 to December 5. Five SPOT images represented the five stages were selected.

Preprocess work includes spectral normalization to reduce the influences from sun's angle and climate conditions. Three bands data were converted to three indexes for examination. They are Normalized Difference Vegetation Index (NDVI), GREENESS, and GREENNESS

NDVI=(NIR-RED)/NIR+RED)
GREENNESS=-0.30132*XS1-0.4321*XS2+0.86408*XS3
BRIGHTNESS=0.60539*XS1+0.61922*XS2+0.50008*XS3

Paddy Cadastre of the Chang-Hwa county was provide by the Crop Bureau. Data include interpreting result from aerial photography of cadastre's cropping status. There are two kinds of status: is-rice. GIS data were converted from ARC/INFO polygons to raster-based pixels. Data were overlaid with matrix of vegetation index. Some of them were selected for calculating statistics of is-rice parcel and non-rice parcel. Some were served as calibrate data in error matrix calculation. Comparing with the three indexes, NDVI has a great success in accuracy, GREENESS show similar performance and high correlation to NDVI. Overall accuracy of BRIGHTNESS is between 75% to 80% that is worst among the three indexes.

Results and Discussions

1) Single image
Table 1 shows the range of NDVI to define rice paddy and the corresponding overall accuracy. The accuracies are low at both transplanting and harvesting stages and a highest value at the reproducing stage.

Table 1. NDVI ranges and accuracies using single image
Stages NDVI Range Overall Accuracy
Transplanting -0.01<NDVI<0.14 74.06%
Growing 0.25<NDVI<0.45 87.74%
Growing 0.29<NDVI<0.56 89.01%
Mellowing 0.20<NDVI<0.42 84.49%
Harvesting -0.33<NDVI<0.05 76.01%

On transplanting stage, the paddy was flooded with water and cover by rice germ. The NDVI make user confuse paddy with other landuse parcel. On growing and mellow stages, the crop was wither in growing or mature phases. Both of them have similar NDVI values that enable distinguishing paddy for building, paved for building, paved road, and fish pond.

On October 3. the crop was vegetative. Highest NDVI values allow one best rule to find the rice paddies. If only one images can be used, image acquiesced from reproducing stage may be the best choice. On December, the crops have been harvested. However, some paddies did not return status of bare soil but was covered by inter-cropping vegetables. It confuses classification and reduce accuracy.

Table 2 shows the error matrix estimated from results on 1997/10/03 (Reproducing stage). Classes of rice and Other Crops are inside the cadastral parcel and OTHERS (most for non-agricultural use) is outside of cadastral parcel. Data show that a lot of rice class misplaced to other-crop class.

Table 2. Error Matrix for Reproducing stage (1997/10/03) data
  Rice Other Crops Others Total Producer's Accuracy
Rice 107,049 17,901 7 124,957 85.67%
Other Crops 4,636 51,782 0 56,418 91.78%
Others 1,397 5,932 83,024 90,353 91.89%
Total 113,082 75,615 83,031 0271,728  
User's Accuracy 94.66% 68.48% 99.99%    

2)Difference between two images
Classifying NDVI ranges and the corresponding accuracies are shown in Table 3. Simple rule of finding NDVI difference is subtracting a NDVI image by succeeds NDVI image. An additional couple between reproducing and transplanting stages which yields largest difference and best accuracy. Difference between Growing-Transplanting stages ranks secondary in the difference value and accuracy. It might imply that bigger the change, better the detection. Cases of reproducing-growing and mellowing- reproducing show similar NDVI ranges and accuracies. It suggests the difference calculated before or after the reproducing stages lead similar result. Error matrix (Table 4) shows the primary modification of difference approach. 17,901 other-crop pixels were misplaced to rice in single-image case but the number is 7,747 in two-image approach. Consequently, correctly placed non-crop pixels numbered 64,484 compared with 51,782 in single image approach.

Table 3. NDVI ranges and accuracies of using difference between images
Stages NDVI Range Overall Accuracy
Growing-Transplanting 0.19<NDVI<0.50 91.42%
Reproducing-Growing 0.04<NDVI<0.30 83.73%
Reproducing-Transplanting 0.24<NDVI<0.60 92.24%
Mellowing-Reproducing -0.20<NDVI<0.02 82.67%
Harvesting-Mellowing -0.33<NDVI<0.05 76.01%

Table 4. Error Matrix for Reproducing-Transplanting Data
  Rice Other Crops Others Total Producer's Accuracy
Rice 103,140 7,747 7 110,894 93.01%
Other Crops 8,011 64,484 0 72,495 88.95%
Others 1,931 3,384 83,024 88,339 93.98%
Total 113,082 76,615 83,031 271,728  
User's Accuracy 91.21% 85.28% 99.99%    

3) Adding a constraint difference approach.
Constraint means screening the non-rice paddy data before applying classify rule. However, the overall accuracy has no significant improvement (Table 5). Error matrix (Table 6) show that the constraint cannot efficiently screen non-rice parcel. The OTHERS class was pre-screened but it is misplaced to non-crop class.

Table 5. NDVI ranges and accuracies of using difference between images with constrain
Stages NDVI Range Overall Accuracy
Growing-Transplanting -0.14<NDVI<0.05 92.33%
Reproducing-Transplanting-Transplanting 0.24<NDVI<0.60 92.92%

Table 6. Error Matrix for Reproducing-Transplanting Data (with constrain)
  Rice Other Crops Others Total Producer's Accuracy
Rice 103,245 8,076 7 111,328 92.74%
Other Crops 8,414 66,341 0 74,755 88.74%
Others 1,423 1,198 83,024 85,645 96.94%
Total 113,082 75,615 83,031 271,728  
User's Accuracy 91.30% 87.74% 99.99%    

The study has illustrated integration of multi-temporal SPOT data and cadastral GIS data can effectively increasing the classification accuracy. Parcel-based classification yields producer accuracy up to 89% in single image case and 92% in two-image case. For operational propose, a rule base combining NDVI, GREENNESS, BRIGHTNESS, and their stage-difference is under constructed. Users can select of combined rules depend on the available remotely sensed data.

Acknowledgements
This study was funded by the Agricultural Council of the Republic of China under the grants of 87-RS-3-(3). During the course of this study, considerable support of cadastral and census data were provided from the Crop Bureau of Taiwan and the Department of Agricultural Engineering. National Taiwan University. We also acknowledge the advice from Professor Yi-Hsing Tseng, National Cheng Kung University.

References
  • Crop Bureau of Taiwan (1997) : Investigation Report of Rice Inventory on Taiwan. 90 Pages (in Chinese)
  • Crist, E. P., and Cicone, R. C.,. (1984) : A Physically-based Transformation of Thematic Mapper Data-The TM Tasseled Cap, IEEE Trans. On Geoscience and Remote Sensing, 22(3), 343-352
  • Eckhardt, W. David and James, P. Verdin (1990) : Automated Update of an Irrigated Lands GIS Using SPOT HRV Imagery, photogrammetric Engineering and Remote Sensing 56(11)1515-1522.
  • Kauth, R. J., and Thomas, G. S. 1976, The Tasseled Cap- A Graphic Description of the Spectral-Temporal Development of Agricultural Crops As Seen by Landsat, Proc., The Symposium of Machine Processing of Remotely Sensed Data, Purdue University, West Lafayette, Ind.
  • Morse A., Zarriello, T. J. and Wramber, W. J. (1990) : Using Remote Sensing and GIS Technology to Help Adjudicate Idaho Water Rights, Photogrammetric Engineering and Remote Sensing, 56(3)365-370.
  • Wolter, P. T., Mladenoff, D. J., Host G. E., Crow, T. R.(1995) : Improved Forest Classification in the Northern Lake States Using Multi-Temporal Landsat imagery, Photogrammetric Engineering & Remote Sensing, Vol. 61, No. 9.
  • Zhuang, X., Engel, B. A., Baumgardner, M. F., Swain, P. H. (1991): Improving Classification of Crop Residues Using Digital Land Ownership Data and Landsat TM Imagery, Photogrammetric Engineering & Remote Sensing, Vol. 57, No. 11, pp: 1487-1492.