| GISdevelopment.net ---> AARS ---> ACRS 1999 ---> Agriculture |
Determination of Rice Planting Area in Japan using Remote Sensing Data
Ogawa, Sigeo1 and Saito, Genya2
National Institute of Agro-Environmental Sciences
3-1-1, Kannondai, Tsukuba, Ibaraki 305-8604, JAPAN
1: Tel. +81-298-38-8225 Fax. +81-298-38-8199
E-Mail:sogawa@niaes.affrc.go.jp
2: Tel. +81-298-38-8192 Fax. +81-298-38-8199
E-Mail:genya@niaes.affrc.go.jp
Ogawa, Sigeo1 and Saito, Genya2
National Institute of Agro-Environmental Sciences
3-1-1, Kannondai, Tsukuba, Ibaraki 305-8604, JAPAN
1: Tel. +81-298-38-8225 Fax. +81-298-38-8199
E-Mail:sogawa@niaes.affrc.go.jp
2: Tel. +81-298-38-8192 Fax. +81-298-38-8199
E-Mail:genya@niaes.affrc.go.jp
Abstract
In Japan, Ministry of Agriculture, Forestry and Fishery (MAFF) is planning to modify of determination system of crops planting areas. MAFF expected high accurate results and work saving procedures using remote sensing data. We determine the paddy field areas in selected regions with very high accuracy using RADARSAT at transplanting season combined with Landsat TM data of past observation. The method thus developed in this study is very useful not only in estimating the crop area very early in the growing season also it can be used at all weather conditions.
We also studyed the relationship between rice growth and backscatter coefficients of RADARSAT/SAR.
1. Introduction
Rice is one of the important crops in Japan. Therefore it is vital to monitor the growth conditions of rice and estimate the area of paddy field as accurately as possible. This also helps design a good food supply plan. Accurate estimation of rice fields, using Landsat TM data (Okamoto, et al. 1996) as well as using ERS-1 data (Panigrahy, et al. 1997) have been well studied. Kurosu, et al. (1997) has also published, recently, of monitoring of rice fields using ERS-1 data. However, the size of paddy fields in Japan is small and there is mixture of crops in the fields with many land use covers. Further more, Japan is located in monsoon area and there are only one or two optical sensor data sets become available during the planting terms. It is virtually impossible to acquire optical sensor data every year during the rice planting seasons.
In Japan, paddy fields are flooded during the rice-planting season and when we use RADARSAT data, water covered areas; paddy fields and mountain shades have low backscatter coefficients. This makes the water-covered areas, paddy fields and mountain shades easy to distinguish then from other land covers. The defaults of SAR data is single banded and contains speckle noise. On the other hand, additional use of Landsat TM data combined with RADARSAT, water covered areas could be distinguished from the mountain shade areas and water covered areas.
The objective of this study, is to estimate the paddy field area annually and accurately mainly using the RADASAT data by monitoring the change in backscatter coefficients in the rice fields. First, we tested some filtering methods to reduce the speckle noise of RADARSAT data, as it is possible to be acquiring RADASAT data sets at any time by changing the sensor angles. We also combined the classified image data from Landsat TM with RADARSAT data, to estimate the paddy field areas with high precision, under all weather conditions.
2. Test Site and using Data
Test site for monitoring paddy field in this study was Kantoh plain Tochigi, Japan. Test site for estimating paddy field area was Ishikari plain Hokkaido (Fig. 1). Tochigi area is located in the center of Japan, while Hokkaido is located in the northern part of the Island.
There are 18 districts in the test site, Ishikari plain where paddy fields occupy large areas. All paddy fields are flooded in planting season and after planting, from the end of May to middle of June. Two Landsat TM data sets were used to select the paddy field candidates. Another TM data set was used to compare the paddy field estimated from this method. RADARSAT data set was used to estimate paddy field areas in the paddy field candidates made from two Landsat TM.
Fig. 1 Test site location
Three RADARSAT data sets were used to monitor the paddy fields and three Landsat TM combined with one RADARSAT data sets were used to estimate the paddy field areas (Table 1).
In Tochigi area, paddy field is flooded in early May, and rice is also planted during this season. Therefore, RADASAT data sets obtained on 17 May, when rice was in its early growth stage of about 15 - 20 cm height and on 27 June, when rice was grown to 50 - 60 cm height were used. Ground survey provided the information about the status of rice growth in the field. The other RADASAT data used was collected on 28, July, just prior to flowering, when rice was grown up to 70 - 80 cm height. ,In Hokkaido area, paddy field is flooded by the end of May. Planting season in northern area is later than in the center of Japan, Tochigi area. Rice in this area is in its early growth stage by 5 and 17 of June.
Table 1. Landsat TM and RADARSAT data sets.
3. Method of Transforming the Digital Number into Backscatter Coefficient
The following procedures are used to extract the value of the radar brightness for pixels in processed image for detected product of path image. If DNj is the digital number, which represents the magnitude of the jth pixel from the start of a range line in the detected image data, then the corresponding value of radar brightness for the pixel is given by;
b0j = 10*log10 [(DNj2 + A3) / A2j] db (1)
Where A2j is the scaling gain value for the j_th pixel, and A3 is the fixed offset. These values are obtained from Radiometric Data Records.
Radar brightness data may be converted to radar backscatter coefficient using the following equation;
s0j = b0j + 10*log10(sin Ij) db (2)
Where Ij is the incidence angle at the j_th range pixel. This formula assumes that the earth is a smooth elllipsoid at sea level. Following procedures were then conducted to complete the analysis.
The verification of the accuracy of estimation of paddy field areas was conducted as follows.
5.1 Monitoring for Paddy Field
Characteristics of backscatter coefficient were shown in Fig. 2. In general, urban areas had high coefficient throughout. High backscatter urban areas were urban1 and urban2 and in descending data set (July 28) they were high (about 5 dB) while in ascending data sets (May 17 and June 27) they were low (about -5dB). On the other hand, urban3 and urban4 had high backscatter in ascending data sets and while low backscater in descending data. This might have occurred due to the direction of houses or buildings.
Fig. 2 Backscatter coefficient of land use in 1998. Paddy1 and urban4 show standard deviation extents.
Although, backscatter coefficients of paddy field were increased from May 17 to June 27 it remain almost unchanged from June 27 to July 28. River and pond had low backscatter coefficients in all three data sets. The C band of SAR data showed that it is easy to distinguish paddy field from other land covers, when we use data obtained during flooded seasons and growing periods.
Fig. 3 Relationship between Backscatter and rice growth
Backscatter coefficient and rice growth index is related, especially backscatter and LAI is observed to lineal relationship (Fig3).
The study fields had complex land usage due to a mixture corps and growing at various stages. Backscatter coefficient of field area were decreased from May to July as the crop grows, ploughed field would be covered with crop covers.
5.2 Estimating for Paddy Field Area
Fig. 4 Comparison of filter processing to reduce speckle noise.
Paddy field and water area are were distinguished from other land covers using TM data set of 12 May, 1995 as it was the flooded season,. Then paddy field were distinguished form water area using TM data set of 8 July 1993. Combing these two classified data of Landsat TM, image of paddy field candidates were created.
Fig. 5 Histogram of main land use
Table 3 Comparision statistical paddy field area and the estimated value. <
Digital number of radar brightness for pixels contains speckle noise. Fig. 4 shows that digital number of original pixel was varied in same land use. Among three filtering methods that were tested MAP refined filter yielded the best result. Its result shows that homogeneous area was averaged and that edges of different areas were detected sharply. This method was adopted in this analysis.
Histogram of land use after filter processing, are shown in Fig. 5. Digital number of paddy field and water area were lowest in them. Only riverbed arranged and covered with grass was not distinguished from paddy field. Threshold was decided to be 34.
After overlaying paddy field candidates of RADARSAT and paddy field candidates of Landsat TM, paddy field was determined. Paddy field area of 18 districts was totaled from the number of pixels. The result was shown in table 3 and Fig.6. The total difference between statistical area and estimated area was -2.3 %. As overlaying RADARSAT data and Landsat TM data, center of Ishikari plain were correspond to each other, but paddy field that is distributed in high elevation area was in a different position because of fore shortening phenomena. As a result, total paddy field area was under estimated. This could have been avoided if it was geocoded using DTM data.
Fig. 6 Comparision statistical paddy field area and estimated value.
6.Conclusions
Using seasonal RADARSAT data, characteristics of backscatter coefficient in agricultural area were investigated. As the result, paddy field has low backscatter in flooded season. Backscatter of paddy field increases as rice grows. However, it is impossible to monitor rice fields precisely due to speckle noise of SAR data. Combining geographical information system with SAR data might help solve this problem.
For an accurate estimation of paddy field area under any weather conditions, use of RADARSAT data in combination of paddy field candidates made from Landsat TM data are very useful. Merely using RADARSAT data, which contains speckle noise results in low accuracy of estimation of paddy field area. There are no created paddy fields from forest or grassland in Japan. It is possible to use paddy field candidates made from Landsat TM every year. However in this study it became possible for us to estimate paddy field area with 97.7% accuracy. Moreover, this method can be used to estimate crop area very early on the growing season. One problem is to correct the fore shortening phenomena. Further studies are being carried out to improve the problem caused by fore shortening.
Acknowledgments
This study was performed as the part of collaboration project of Agriculture Remote Sensing by National Space Development Agency and Ministry of Agriculture, Forestry and Fishery in Japan. The Crops Statistics Section, Statistics and Information Department, Economics Affairs Bureau and Sapporo Statistics and Information Office of the Ministry of Agriculture, Forestry and Fisheries cooperated in collecting paddy field area data and in conducting field studies for this work. Dr. Shoji Takeuchi, Masahiko Honzawa and Miss Chinatsu Yonezawa, researcher at Research Department, Remote Sensing Technology Center of Japan, cooperated in analyzing the RADARSAT data. Mr. Lukman Thalib suggested statistical analysis. We would like to express our deep gratitude to all the members.
References
In Japan, Ministry of Agriculture, Forestry and Fishery (MAFF) is planning to modify of determination system of crops planting areas. MAFF expected high accurate results and work saving procedures using remote sensing data. We determine the paddy field areas in selected regions with very high accuracy using RADARSAT at transplanting season combined with Landsat TM data of past observation. The method thus developed in this study is very useful not only in estimating the crop area very early in the growing season also it can be used at all weather conditions.
We also studyed the relationship between rice growth and backscatter coefficients of RADARSAT/SAR.
1. Introduction
Rice is one of the important crops in Japan. Therefore it is vital to monitor the growth conditions of rice and estimate the area of paddy field as accurately as possible. This also helps design a good food supply plan. Accurate estimation of rice fields, using Landsat TM data (Okamoto, et al. 1996) as well as using ERS-1 data (Panigrahy, et al. 1997) have been well studied. Kurosu, et al. (1997) has also published, recently, of monitoring of rice fields using ERS-1 data. However, the size of paddy fields in Japan is small and there is mixture of crops in the fields with many land use covers. Further more, Japan is located in monsoon area and there are only one or two optical sensor data sets become available during the planting terms. It is virtually impossible to acquire optical sensor data every year during the rice planting seasons.
In Japan, paddy fields are flooded during the rice-planting season and when we use RADARSAT data, water covered areas; paddy fields and mountain shades have low backscatter coefficients. This makes the water-covered areas, paddy fields and mountain shades easy to distinguish then from other land covers. The defaults of SAR data is single banded and contains speckle noise. On the other hand, additional use of Landsat TM data combined with RADARSAT, water covered areas could be distinguished from the mountain shade areas and water covered areas.
The objective of this study, is to estimate the paddy field area annually and accurately mainly using the RADASAT data by monitoring the change in backscatter coefficients in the rice fields. First, we tested some filtering methods to reduce the speckle noise of RADARSAT data, as it is possible to be acquiring RADASAT data sets at any time by changing the sensor angles. We also combined the classified image data from Landsat TM with RADARSAT data, to estimate the paddy field areas with high precision, under all weather conditions.
2. Test Site and using Data
Test site for monitoring paddy field in this study was Kantoh plain Tochigi, Japan. Test site for estimating paddy field area was Ishikari plain Hokkaido (Fig. 1). Tochigi area is located in the center of Japan, while Hokkaido is located in the northern part of the Island.
There are 18 districts in the test site, Ishikari plain where paddy fields occupy large areas. All paddy fields are flooded in planting season and after planting, from the end of May to middle of June. Two Landsat TM data sets were used to select the paddy field candidates. Another TM data set was used to compare the paddy field estimated from this method. RADARSAT data set was used to estimate paddy field areas in the paddy field candidates made from two Landsat TM.
Fig. 1 Test site location
Three RADARSAT data sets were used to monitor the paddy fields and three Landsat TM combined with one RADARSAT data sets were used to estimate the paddy field areas (Table 1).
In Tochigi area, paddy field is flooded in early May, and rice is also planted during this season. Therefore, RADASAT data sets obtained on 17 May, when rice was in its early growth stage of about 15 - 20 cm height and on 27 June, when rice was grown to 50 - 60 cm height were used. Ground survey provided the information about the status of rice growth in the field. The other RADASAT data used was collected on 28, July, just prior to flowering, when rice was grown up to 70 - 80 cm height. ,In Hokkaido area, paddy field is flooded by the end of May. Planting season in northern area is later than in the center of Japan, Tochigi area. Rice in this area is in its early growth stage by 5 and 17 of June.
Table 1. Landsat TM and RADARSAT data sets.
| Sensor | Observation date | Use |
| Landsat TM | 12-Jun-96 | Removing areas of open water, mountainous shadows and selecting paddy field candidates |
| Landsat TM | 8-Jul-93 | |
| Landsat TM | 17-Jun-97 | Verifying paddy field |
| RADARSAT | 5-Jun-97 | Estimating paddy field in the paddy field candidates |
| RADARSAT | 17-May-98 | transforming digital number of path image into backscatter coefficient |
| RADARSAT | 27-Jun-98 | |
| RADARSAT | 28-Jul-98 | |
| Landsat TM | 21-May-87 | Verifying land use |
| Landsat TM | 24-Jul-93 |
3. Method of Transforming the Digital Number into Backscatter Coefficient
The following procedures are used to extract the value of the radar brightness for pixels in processed image for detected product of path image. If DNj is the digital number, which represents the magnitude of the jth pixel from the start of a range line in the detected image data, then the corresponding value of radar brightness for the pixel is given by;
b0j = 10*log10 [(DNj2 + A3) / A2j] db (1)
Where A2j is the scaling gain value for the j_th pixel, and A3 is the fixed offset. These values are obtained from Radiometric Data Records.
Radar brightness data may be converted to radar backscatter coefficient using the following equation;
s0j = b0j + 10*log10(sin Ij) db (2)
Where Ij is the incidence angle at the j_th range pixel. This formula assumes that the earth is a smooth elllipsoid at sea level. Following procedures were then conducted to complete the analysis.
- Geocoding: RADARSAT data were projected onto 1: 50,000 scale topographical maps. Used map projection of these images was UTM and map system was Bessel. The resample images were prepared using nearest neighbor method. Resampled pixels were set to 12.5m.
- Selection of training fields: Using 1:50,000 map and Landsat TM, training fields of land use were selected.
- Transforming into backscatter coefficient: Using (1) and (2) equation, digital number were transformed into backscatter coefficient.
- Relationship between backscatter coefficient of rice and rice growth: Rice growth index of fresh weight, dry weight, LAI (Leaf Area Index) and height were measured and compared to backscatter coefficient.
The verification of the accuracy of estimation of paddy field areas was conducted as follows.
- Geocoding: Landsat TM data and RADARSAT data were projected onto 1: 50,000 scale topographical maps. Used map projection of these images was UTM and map system was Bessel. Nearest neighbor method was used for resampling. Resample pixels were set to 25m in Landsat TM data and 12.5m in RADARST data.
- Land cover classification: Respective Landsat TM data with bands 2, 3, 4, 5 and 7 were classified by an unsupervised method (using ISODATA method), in which 60 initial classes and 95 per cent convergence limit were used as parameters. After processing the data, 60 classes were given to grand truth data and re-coded to 8 classes.
- Combining two lands cover classified images: Using two classified images of Landsat TM, new classes were set. The new classes were defined as in Table 1.
- Extracting open water: Pixels labeled as open water were extracted from the land cover classification.
- Reduction of classes: Pixels isolated in the forest area and some classes of forest were re-coded to forest.
- Selection of paddy field candidates: Paddy field candidates were selected from the image processed by (3), (4) and (5). Paddy field candidates were selected previously.
- Extracting paddy field candidates: Water was extracted from the Paddy field candidates using processed image as given in (6).
- Processing district identification file: District boundaries were read from 1: 50,000 maps and its coverage file are created by Arc/Info. Each districts was identified by ID codes. This vector file was transformed in to raster file.
- Speckle noise suppression: Three methods of speckle noise suppression were tested and evaluated. These methods are Lee filter, Frost filter and MAP refined filter.
- Decision of threshold vale in paddy field: After processing speckle noise suppression, histograms of land use were calculated and threshold vale of paddy field was decided from them.
- Selecting paddy field areas from RADARSAT: Using threshold vale of paddy field, paddy field candidates were selected from RADARSAT. By overlaid two paddy field candidates that were processed from RADARSAT and Landsat TM, paddy field areas were selected.
- Sum up the paddy field areas for each district: Paddy field areas were summed up for each district using district file.
- Selecting paddy field areas from Landsat TM and comparing: Land cover classification obtained by processing Landsat TM data acquired on 17 June, 1997 was used for this purpose.
5.1 Monitoring for Paddy Field
Characteristics of backscatter coefficient were shown in Fig. 2. In general, urban areas had high coefficient throughout. High backscatter urban areas were urban1 and urban2 and in descending data set (July 28) they were high (about 5 dB) while in ascending data sets (May 17 and June 27) they were low (about -5dB). On the other hand, urban3 and urban4 had high backscatter in ascending data sets and while low backscater in descending data. This might have occurred due to the direction of houses or buildings.
Fig. 2 Backscatter coefficient of land use in 1998. Paddy1 and urban4 show standard deviation extents.
Although, backscatter coefficients of paddy field were increased from May 17 to June 27 it remain almost unchanged from June 27 to July 28. River and pond had low backscatter coefficients in all three data sets. The C band of SAR data showed that it is easy to distinguish paddy field from other land covers, when we use data obtained during flooded seasons and growing periods.
Fig. 3 Relationship between Backscatter and rice growth
Backscatter coefficient and rice growth index is related, especially backscatter and LAI is observed to lineal relationship (Fig3).
The study fields had complex land usage due to a mixture corps and growing at various stages. Backscatter coefficient of field area were decreased from May to July as the crop grows, ploughed field would be covered with crop covers.
5.2 Estimating for Paddy Field Area
Fig. 4 Comparison of filter processing to reduce speckle noise.
Paddy field and water area are were distinguished from other land covers using TM data set of 12 May, 1995 as it was the flooded season,. Then paddy field were distinguished form water area using TM data set of 8 July 1993. Combing these two classified data of Landsat TM, image of paddy field candidates were created.
Fig. 5 Histogram of main land use
| A | B | A/B | A-B | |
| District | Stat. Area | Est. Area | Rate (%) | Diff. |
| Fukagawa C. | 6,950 | 6,959 | 100.1 | -9 |
| Numata T. | 2,550 | 2,508 | 98.3 | 42 |
| Asahikawa C. | 7,440 | 6,952 | 93.4 | 488 |
| Hokuryu T. | 2,180 | 2,062 | 94.6 | 118 |
| Tippubetsu T. | 2,320 | 2,340 | 100.9 | -20 |
| Moseushi T. | 2,790 | 2,901 | 104.0 | -111 |
| Uryu T. | 2,470 | 2,312 | 93.6 | 158 |
| Shintotsukawa T | .3,660 | 3,730 | 101.9 | -70 |
| Takikawa C. | 2,770 | 2,993 | 108.1 | -223 |
| Ashibetsu C. | 1,670 | 1,446 | 86.6 | 224 |
| Akabira C. | 464 | 469 | 101.0 | -5 |
| Sunagawa C. | 609 | 754 | 123.9 | -145 |
| Urausu T. | 1,830 | 1,699 | 92.8 | 131 |
| Nakafurano T. | 1,860 | 1,879 | 101.0 | -19 |
| Nie T. | 1,550 | 1,484 | 95.7 | 66 |
| Tsukigata T. | 1,800 | 1,646 | 91.4 | 154 |
| Bibai T. | 6,440 | 6,065 | 94.2 | 375 |
| Furano C. | 1,260 | 1,232 | 97.8 | 28 |
| Total | 50,613 | 49,429 | 97.7 | 1,184 |
Table 3 Comparision statistical paddy field area and the estimated value. <
Digital number of radar brightness for pixels contains speckle noise. Fig. 4 shows that digital number of original pixel was varied in same land use. Among three filtering methods that were tested MAP refined filter yielded the best result. Its result shows that homogeneous area was averaged and that edges of different areas were detected sharply. This method was adopted in this analysis.
Histogram of land use after filter processing, are shown in Fig. 5. Digital number of paddy field and water area were lowest in them. Only riverbed arranged and covered with grass was not distinguished from paddy field. Threshold was decided to be 34.
After overlaying paddy field candidates of RADARSAT and paddy field candidates of Landsat TM, paddy field was determined. Paddy field area of 18 districts was totaled from the number of pixels. The result was shown in table 3 and Fig.6. The total difference between statistical area and estimated area was -2.3 %. As overlaying RADARSAT data and Landsat TM data, center of Ishikari plain were correspond to each other, but paddy field that is distributed in high elevation area was in a different position because of fore shortening phenomena. As a result, total paddy field area was under estimated. This could have been avoided if it was geocoded using DTM data.
Fig. 6 Comparision statistical paddy field area and estimated value.
6.Conclusions
Using seasonal RADARSAT data, characteristics of backscatter coefficient in agricultural area were investigated. As the result, paddy field has low backscatter in flooded season. Backscatter of paddy field increases as rice grows. However, it is impossible to monitor rice fields precisely due to speckle noise of SAR data. Combining geographical information system with SAR data might help solve this problem.
For an accurate estimation of paddy field area under any weather conditions, use of RADARSAT data in combination of paddy field candidates made from Landsat TM data are very useful. Merely using RADARSAT data, which contains speckle noise results in low accuracy of estimation of paddy field area. There are no created paddy fields from forest or grassland in Japan. It is possible to use paddy field candidates made from Landsat TM every year. However in this study it became possible for us to estimate paddy field area with 97.7% accuracy. Moreover, this method can be used to estimate crop area very early on the growing season. One problem is to correct the fore shortening phenomena. Further studies are being carried out to improve the problem caused by fore shortening.
Acknowledgments
This study was performed as the part of collaboration project of Agriculture Remote Sensing by National Space Development Agency and Ministry of Agriculture, Forestry and Fishery in Japan. The Crops Statistics Section, Statistics and Information Department, Economics Affairs Bureau and Sapporo Statistics and Information Office of the Ministry of Agriculture, Forestry and Fisheries cooperated in collecting paddy field area data and in conducting field studies for this work. Dr. Shoji Takeuchi, Masahiko Honzawa and Miss Chinatsu Yonezawa, researcher at Research Department, Remote Sensing Technology Center of Japan, cooperated in analyzing the RADARSAT data. Mr. Lukman Thalib suggested statistical analysis. We would like to express our deep gratitude to all the members.
References
- Kurosu T., Fujita M. and Chiba K., 1997, The identification of rice fields using multi-temporal ERS-1 C band SAR data. International Journal of Remote Sensing, 18(14), pp.2953-2965.
- Okamoto, K. and Fukuhara, M., 1996, Estimation of paddy field area using the area ratio of categories in each mixel of Landsat TM. International Journal of Remote Sensing, 17(9), pp.1735-1749.
- Panigrahy, M., Chakraborty, M., Sharma, S. A., 1997, Early estimation of rice area using temporal ERS-1 synthetic aperture radar data - a case study for the Howrah and Hughly districts of West Bengal, India. International Journal of Remote Sensing, 18(8), pp.1827-1833.