| GISdevelopment.net ---> AARS ---> ACRS 1999 ---> Agriculture |
Radarsat Data for Operational Rice Monitoring and its Potential for Yield Estimation
Yun Shao1 Hao Liu1 Xiangtao Fan1,2 Jianhua Xiao1 Qing Dong1 Xingyuan Wang1,2 S. Ross3 B. Brisco3 R. Brown4 G. Staples5
1.Institute of Remote Sensing Applications,
Chinese Academy of Sciences,
P.O. Box 9718, Datun Road, Beijing 100101, China
Tel: 86-10-64876313, Fax: 86-10-64889786
E-Mail:yunshao@public.bta.net.cn
2. Department of Earth Sciences, Nanjing University, Nanjing, China
3. Noetix Research Inc., Ottawa, Ontario, Canada.
4. Canada Centre for Remote Sensing, Ottawa, Ontario, Canada
5. RADARSAT International, Vancouver, British Columbia, Canada
Yun Shao1 Hao Liu1 Xiangtao Fan1,2 Jianhua Xiao1 Qing Dong1 Xingyuan Wang1,2 S. Ross3 B. Brisco3 R. Brown4 G. Staples5
1.Institute of Remote Sensing Applications,
Chinese Academy of Sciences,
P.O. Box 9718, Datun Road, Beijing 100101, China
Tel: 86-10-64876313, Fax: 86-10-64889786
E-Mail:yunshao@public.bta.net.cn
2. Department of Earth Sciences, Nanjing University, Nanjing, China
3. Noetix Research Inc., Ottawa, Ontario, Canada.
4. Canada Centre for Remote Sensing, Ottawa, Ontario, Canada
5. RADARSAT International, Vancouver, British Columbia, Canada
Abstract
Rice monitoring and yield estimation has special significance to China, as rice is the staple grain and accounts for 42% of the crop yield for this country. Radar remote sensing is appropriate for monitoring rice, as cultivated areas are most often cloudy and rainy. For this reason, SAR is anticipated to be the dominant data source in tropic and sub-tropical regions and also provide re-visit schedules suitable for agricultural monitoring. This paper presents the results of a study examining the backscatter behavior of rice using multi-temporal RADARSAT dataset acquired in 1996 and 1997. A rice-type distribution map was produced, showing 4 types of rice with different life spans ranging from 80 days, to 120-125 days. The life span of a rice crop has significant impact on yield, as well as the taste and quality of the rice. The yield of three counties and two administrative regions, totaling 5000 square kilometers, are estimated in this study. The accuracy was found to be 91%, providing confidence that multi-temporal RADARSAT data is capable of rice monitoring and yield estimation. Based on the studies carried out in the Zhaoqing test site, it is suggested that rice yield estimations require three radar data acquisitions taken at 3 stages of crop growth circle. That is at the end of the seedling development period, in the ear differentiation period, and at the beginning of the harvest period. Alternatively, if multi-parameter radar data is available, only two data acquisitions are required: at the end of the seedling period, and at the beginning of the harvest period. In 1999, we started a pilot experiment on operational rice monitoring and potential yield estimation using multi-temporal RADARSAT data. Finally, this paper proposes a pilot scenario for operational rice monitoring and yield estimation.
1. Introduction
Rice is a heat and water favorite crop. Most paddy rice in the world grows in warm, humid environment with heavy cloud cover and rainfall. It is hard to acquire optical remote sensing data in rice growing regions. Synthetic Aperture Radar (SAR), with all weather, independent of illumination imaging capability and frequent revisit schedule, is anticipated to be the dominant data source for agriculture monitoring in tropic and sub-tropic regions. Rice monitoring and yield estimation has special significance to China, as rice is the staple grain and accounts for 42% of the crop yield for this country. The estimation of crop yield is a topic of global interests (McDonald and Hall, 1980), and the efficient management of agricultural land resources is strongly related to social and economic sustainable development, especially in China. It is well known that China is the largest country in population. However, as the population increases, and economy and industry develop, the quantity and quality of cultivated land is decreasing rapidly. The food supply to the current 1.2 billion people is a serious concern facing China, and will intensify as the population continues to grow. Therefore, it is important to find an efficient way to face this dilemma. Remote sensing technology will provide the needed information on crop distribution, acreage and potential yield.
The use of microwave remote sensing technology to study ecological systems is becoming more and more popular among the world's scientific community (Dobson, 1992; Kasischke and Christensen, 1990). Scientists have carried out extensive field measurements (Ulaby et al 1986), airborne flight mission (Zebker et al, 1991),
Table 1. System Parameters of SAR Data
(Campbell et al, 1995, Guo et al, 1995, 1997, Shao et al, 1995 a, b), spaceborne mission (Stofen et al, 1995). Many microwave backscatter models were developed to study the backscatter behavior of vegetation (Chauhan, et al, 1994; Karam and Fung, 1988; Matzler, 1994; Ulaby, et al, 1990). There are certainly much more literatures available on researches on forest applications, rather than agriculture (Engheta and Elachi, 1982; Freeman et al, 1992; Ulaby, et al, 1990, Brisco et al, 1997). However there are many successful examples in radar remote sensing for agricultural applications (Shao et al, 1994, 1996; Le Toan et al, 1989; Ulaby et al, 1982; Soares et al, 1987, 1997; Schotten et al, 1995; Anys and He, 1995), and a few papers specifically on rice monitoring (Shao, et al, 1997 a, b, c, d; Liu et al, 1997; Kurosu et al, 1995, 1997; Le Toan et al, 1997; Aschbacher et al, 1995 ). The results on rice monitoring using SAR technology were very promising.
This paper presents the results of a study examining the backscatter behavior of rice using multi-temporal RADARSAT dataset. A rice-type distribution map was produced, showing 4 types of rice with different life spans ranging from 80 days, to 120-125 days. The life span of a rice crop has significant impact on yield, as well as the taste and quality of the rice. The yield of three counties and two administrative regions, totaling 5000 square kilometers, are estimated in this study. The accuracy was found to be 91%, providing confidence that multi-temporal RADARSAT data is capable of rice monitoring and yield estimation. Based on previous studies in the Zhaoqing test site, it is suggested that rice yield estimations require three radar data acquisitionstaken at 3 stages of crop growth: at the end of the seedling development period, in the ear differentiation period, and at the beginning of the harvest period. Alternatively, if multi-parameter radar data is available, only two data acquisitions are required: at the end of the seedling period, and at the beginning of the harvest period. Finally, this paper proposes a pilot scenario for operational rice monitoring and yield estimation.
2. Test Site and Data Source
The Zhaoqing test site is located in Guangdong Province, south of China center at latitude 22.30, longitude 112.30. It is sited at the northwestern end of Pearl River Delta. Airborne SAR firstly imaged the test site in 1993 under the GlobeSAR program (Shao, 1995, 1996; Guo 1997). The Shuttle Imaging Radar C-band (SIR-C) and X-band SAR (X-SAR) also flew over the area on April 18, 1994. In addition, there were multiple RADARSAT images acquired from March to December in1996, and from April to July in 1997. The system parameters, imaging modes, and acquisition dates of images used in this study are listed in table 1. For Zhaoqing test site, there was a three years continues RADARSAT data acquisition for rice monitoring. The first two years data acquisition focus on studying the backscatter behavior of rice and the potential of RADARSAT for rice yield estimation. The third year's data acquisition is for operational rice monitoring and yield estimation.
3. Rice Calendar
In the Zhaoqing test site, there are two crops per year; early season rice and late season rice. There are five major growth periods in the life cycle of rice. 1) Transplanting period: rice plant seedlings are transplanted from the seedbed to the paddy field. The transplanting date depends on the weather, especially temperature; 2) Seedling developing period: the seedling splits up and begins to develop a root system; 3) Ear differentiation period; 4) Heading period: headings begin to form; 5) Mature period: the rice plants mature and are ready to be harvested. Temporally, these five periods for early season rice are March 25 ~ April 5, April 15 ~ 25,
May 10 ~ 30, June 10 ~ 25, July 5 ~ 31. For late season rice, the growth stages occur as follows: July 20 ~ August 5, August 10 ~ 20, September 1 ~ 30, October 1 ~ 20, November 1 ~ 25 respectively. The exact dates of each period vary with the weather condition during the whole life of rice and are totally rice life span dependent. Plant maturity rates and life span are species dependent.
This rice calendar only illustrates a conventional law of rice growing behavior.
With progress in agricultural technology, it is known that a long growing season is favorable to high photosynthetic rates and high rice yield. For this reason, the transplanting date of rice is getting earlier and the life span is getting longer. For instance, the local farmers grow three rice crops a year in past, but currently, they only grow two crops a year. In USA, there is only one rice crops a year. It starts in early may then harvests in end of October. The rice life circle is around 170 ~ 180 days. The yield reaches 6.2 ton/ha, which is higher than rice yield of 5.9 ton/ha in China. Improved varieties and cultural practices have contributed significantly to yield in last decades too.
4. Temporal Backscatter Behavior of Rice
This study aimed at understanding the backscatter behavior of rice over its whole life circle, and the relationship between the rice structure parameters and its backscatter coefficients. Based on a good understanding of backscatter characteristics of rice, we can then suggest the best date for radar data acquisition to monitor the rice growth and estimate the yield. Figures 1 and 2 show the backscatter coefficients of rice extracted from calibrated RADARSAT images acquired from summer to autumn in 1996 and from spring to summer of 1997 respectively. These are produced by applying the calibration procedure to the digital number of Radarsat image to produce for a training area (20 pixel by 20 pixel) of the multi-temporal RADARSAT dataset.
In figure 1, there are five types of rice: 1) medium mature rice. 2) late mature rice. The backscatter coefficient of late mature rice is slightly higher than medium mature rice on August 4. 3) Late transplanted rice, which has been transplanted about 25 days later than the normal spring rice, due to a cold spring and low temperatures in 1996. The RADARSAT image acquired on April 25 shows the backscatter coefficients of these fields as very low, responding similar to a still water body. The backscatter coefficients from the RADARSAT images taken after April 25 increase gradually, until September when the rice is harvested. Its life span is relatively short, and the backscatter coefficients are lower. 4) single rice and 5) autumn rice. These two rice types demonstrate how farming activities can change within a year, switch between fishponds and rice fields. Spring rice was a rice field in spring then switched a fishpond in autumn. Conversely, Autumn rice was a fishpond in spring then switched to rice field in autumn. This practice is very common in Zhaoqing area, as well as many other parts of China. In general, this swap of fishpond and rice implies that the quality of the field is less than excellent, and that the rice yield is normally lower. Based on the classification results of the 1996 multi-temporal fine mode RADARSAT dataset, A land cover map was produced clearly showing the five types of rice, as well as other vegetation covers and targets.
Figure 1. Backscatter coefficients of rice as a function of time (1996)
In figure 2, there are four major types of rice with different growth cycles. With the knowledge gained from the rice backscatter behavior of 1996, we can distinguish between rice types quite easily. They are early mature rice, medium mature rice, medium-late mature rice, and late mature rice. Their life span is about 80 days, 100 -115 days, 110-120 days and 120-125 days respectively. A false-color fusion image is produced based on the multi-temporal Standard mode RADARSAT images acquired in the spring of 1997. A rice distribution map is produced using the 7 RADARSAT scenes. If a type of rice has higher backscatter value in April, at the early stage or seedling development period, as found with late mature rice, then it was transplanted earlier than the other types of rice. If a type of rice has higher backscatter coefficients in July, the late stage or the harvest period, the rice is harvested later than other type of rice, as found with medium-late mature rice. From figure 2 we can conclude that late mature rice has the longest lifetime. The medium-late mature rice was transplanted a few days later than medium mature rice and late mature rice, and harvested later. The medium mature rice was harvested before July 22. From figure 2, we can see that early mature rice was harvested before June 28. It has very lower backscatter coefficients since early harvest period. It is concluded that from the early stage, 25 days after transplanting, and late stage, 80 days after transplanting, we can distinguish the rice life span easily, which is an important information for rice yield estimation. The life span is directly related to the yield and quality of rice. This information is provided by local agronomists. It has been proved in real life. The farmer used to grow three rice crops a year in past, nowadays, they only grow two rice crops a year and they make more money from rice farming.
Figure 2. Backscatter coefficients of rice as a function of time (1997)
Based on the knowledge we accumulated in our researches in Zhaoqing test site started in 1993, we propose a theoretic backscatter model for rice. It is mainly suit for rice monitoring using RADARSAT data. The backscatter coefficients from flooded rice field before transplanting, in another words, the water surface recorded by RADARSAT is around -25 dB. After transplanting, in the first 20 days, the backscatter coefficients from paddy rice field are -15 dB. Then entering the seedling-developing period, from 20 days to 45 days, the backscatter coefficients from paddy rice field increase to -12 dB. From 45 days to 75 days, within the ear differentiation period, the backscatter coefficients from paddy rice field increase to -9 dB. In the middle of ear differentiation period, about 65 days, the backscatter coefficients reach the highest point -6 dB. Then gradually drops to -8 dB and keep this value during the mature period. The theoretic backscatter model is expressed in figure 3. It is produced by cubic polynomial that simulated the backscatter coefficients of rice as function of time. It mainly represents the backscatter behavior of medium-late mature rice. Medium-late mature rice is transplanted in April 1 and harvested in July 30. The late mature rice and medium mature rice are transplanted earlier than medium-late mature rice. Medium mature rice is harvested earlier than medium-late mature rice. This model can be applied to other type of rice and modified in accordance with rice life span, transplant date and harvest date. Then we can use this model to calculate the backscatter coefficients of rice at certain growth stage. During the mature period of spring rice, it is flooding season (Guo et al, 1999), so the backscatter coefficients of rice can be much higher than normal value since the high moisture or dew effect.
The results from the studies carried out in 1996 and 1997 imply that the most important radar data acquisition time for rice monitoring is at the end of the seedling period and beginning of ear differentiation period, which is late April for spring rice and late August for autumn rice. For calculating the acreage of planted rice, two radar data acquisitions are required: one near the end of April, and another in the middle of May during the ear differentiation period. The first data acquisition tells the biggest difference between rice and other targets. The two data sets acquired in two dates is used for automatic classification. For rice yield estimation, three radar data acquisitions are required: one at the end of April, one in the middle of May, and one at the end of June. For autumn rice or late season rice, the corresponding date can be found by checking the rice growth calendar.
5. A Proposed Scenario for Operational Rice Monitoring and Yield Estimation
The value of a given data set and research project is measured in its practicality, and applicability to the real world. Researchers at the Institute of Remote Sensing Applications (IRSA), through a program entitled "SAR Technology for Rice Growth Status Monitoring and Land Cover Mapping" under the support of National High Technology Program, have strive to provide information on rice growth status and yield to both national and local government agencies, as well as to commercial users. The final goal of this project is to establish an operational rice monitoring and yield estimation system. This paper presents the research achievements of this program. Figure 4 outlines the operating procedures for rice monitoring and yield estimation as followed for this study. The entire working procedure can be divided into four parts:
5.1. Data Input:
5.1.1. Multi-Temporal and Multi-Parameter Radar Data
The radar data provides the latest information on rice growing status and the accurate acreage of planted rice. For single frequency and polarization radar data like RADARSAT, two or three data acquisitions are required: one about 20 days after the transplanting period, and one at the end of seedling development period. For the multi-parameter radar data (i.e. multi-frequency or multi-polarization), one data acquisition is required during the early stage of rice growth. The another critical radar data acquisition is at the end of the heading period.
5.1.2.Static Supporting Data
A digital elevation model (DEM) is required to understand the rice distribution limitations and to aid in image processing. In addition, soil maps help to know the soil quality and its effect on rice production. Knowledge of the crop calendar helps to decide the optimum radar data acquisition date, and explain the rice growth stage at the acquisition time. The crop calendar may shift a few days at the transplanting period, in accordance with the change of temperature.
5.1.3.Dynamic Supporting Data
Meteorological data, including temperature, total sun illumination, rainfall etc. are important in deciding the conventional yield of rice on a regional scale. Knowledge of tillage activities provides information on the seeding of the rice and transplanting methods, which is directly related to the yield per unit area.
In the early stages of the study, data sources B and C determine the rice yield model. Data source C will be a dominant factor in determining the annual yield.
5.2. Supporting Database
There are a series of database and numerical models to support the image processing and the yield estimation work.
A.
Statistical results of the acreage of rice and other targets, based on county and provincial administration boundaries:
Table 2 shows the statistical results for Shihui County, Gaoyao County, Shanshui County and Dinghu administration region, and Duanzhou administration region, covering total area of 4998.1 square km. The yield is 404227 mu ´ 0.45 ton = 181, 902.15 ton for medium mature rice, 379179 mu ´ 0.55 ton = 208,548.45 for late mature rice, 349272 mu ´ 0.5 ton = 174,636.0 ton for medium-late mature rice, and 51366 mu ´ 0.35 ton = 17,978.1 tons for early mature rice. The rice yield of per unit, in this case it is mu, is provided by local agronomist. The average accuracy of the classification is 90.1%, and the overall accuracy of the classification is 91.49%. The classification accuracy is examined by PCI software package. The classification correctness is checked by ground truth verification.
Table 2: Statistical Acreage Results of Land Cover Classes in Zhaoqing
B. Maps:
According to the meteorological data, 1997 was a good year for agriculture. For this reason, the yield should be at the 'normal' level. In order to classify the rice, it is important to separate it from other targets before considering the different rice types (Shao Yun et al, 1997 a,b,c,d). In 1997, the output of this project includes four map products (Guo, ed. 1999 a, b): 1) the multi-temporal RADARSAT image map of Zhaoqing test site; 2) the land cover map; 3) the crop distribution map; 4) the rice distribution map. In 1999, the output of this project includes two map products (Guo, ed. 1999 a, b): 1) the multi-temporal RADARSAT image map of Pearl River Delta test site; 2) the rice distribution map.
The multi-temporal RADARSAT image map was produced using the following procedures: 1) co-registration of the seven scenes by selecting GCPs. They are stored in the database for future use. 2) Generate a false-color image by combining the July 22 (red), the mean intensity of the June 4, 12, 28 images (green), and the mean Intensity of images acquired on April 25 and May 19 (blue) (Figure 7). 3) Based on ground truth data accumulated since the GlobeSAR mission in 1993, the backscatter behavior of rice and other targets at certain times in this region is known well. Defining training areas on the multi-temporal RADARSAT images, then running a neural network classifier produces the other maps. The training areas should be defined as precisely as possible in order to obtain an accurate classification output. 5) Filter and enhance the classified results to remove speckle and within field variability and produce a land covers map. 6) A rice distribution map is created by combining targets so as to display only the different rice crops, water bodies, forest land, urban areas, and non rice targets. After the filtering and enhancement, the overall accuracy of the products is 97%. The quality of these products created from multi-temporal RADARSAT data demonstrate the feasibility and effectiveness of RADARSAT data for agricultural applications such as rice crop monitoring.
6. Concluding Remarks
Multi-frequency and multi-polarization SAR imagery is very useful for separating agricultural vegetation types. Multi-temporal RADARSAT images, with single frequency and polarization, can also provide useful monitoring capabilities. The fine mode imagery has shown almost the same capabilities for discriminating the vegetation as the airborne SAR data collected under the GlobeSAR program. However, the multiple RADARSAT data can not only separate the different targets, but can also provide information on the growth status of rice. There still exists a need for better image processing to extract linear targets, and for classification of the image. Additional work is needed to understand the interaction mechanism between the plant parameters and the radar signal as a function of frequency and polarization. It is anticipated that modeling research will continue to provide a clearer, and better explanation and understanding. In addition, recently developed polarimetric (Ulaby et al, 1991) and interferometric research using multi-parameter SAR systems will provide even more information on all land cover targets.
Acknowledgments
This research work is founded by the 863 National High Technology Program, sponsored by Ministry of Science and Technology. The authors wish to take this opportunity to express their sincere acknowledgment to the members of the 863-308 Expert Group: Prof. H. D. Guo, J. M. Xu, Z. M. Yang, Z. Q. Wei, X. G. Ling, X. T. Zhou, G. Q. Ni, and Dr. .Y. J. Wu of MST, Mr. J.L. Miao of the 863-308 office. Special appreciation goes to Dr. Fred Campbell of the Canada Center for Remote Sensing for his kind support to this research.
References
(In Chinese)
Rice monitoring and yield estimation has special significance to China, as rice is the staple grain and accounts for 42% of the crop yield for this country. Radar remote sensing is appropriate for monitoring rice, as cultivated areas are most often cloudy and rainy. For this reason, SAR is anticipated to be the dominant data source in tropic and sub-tropical regions and also provide re-visit schedules suitable for agricultural monitoring. This paper presents the results of a study examining the backscatter behavior of rice using multi-temporal RADARSAT dataset acquired in 1996 and 1997. A rice-type distribution map was produced, showing 4 types of rice with different life spans ranging from 80 days, to 120-125 days. The life span of a rice crop has significant impact on yield, as well as the taste and quality of the rice. The yield of three counties and two administrative regions, totaling 5000 square kilometers, are estimated in this study. The accuracy was found to be 91%, providing confidence that multi-temporal RADARSAT data is capable of rice monitoring and yield estimation. Based on the studies carried out in the Zhaoqing test site, it is suggested that rice yield estimations require three radar data acquisitions taken at 3 stages of crop growth circle. That is at the end of the seedling development period, in the ear differentiation period, and at the beginning of the harvest period. Alternatively, if multi-parameter radar data is available, only two data acquisitions are required: at the end of the seedling period, and at the beginning of the harvest period. In 1999, we started a pilot experiment on operational rice monitoring and potential yield estimation using multi-temporal RADARSAT data. Finally, this paper proposes a pilot scenario for operational rice monitoring and yield estimation.
1. Introduction
Rice is a heat and water favorite crop. Most paddy rice in the world grows in warm, humid environment with heavy cloud cover and rainfall. It is hard to acquire optical remote sensing data in rice growing regions. Synthetic Aperture Radar (SAR), with all weather, independent of illumination imaging capability and frequent revisit schedule, is anticipated to be the dominant data source for agriculture monitoring in tropic and sub-tropic regions. Rice monitoring and yield estimation has special significance to China, as rice is the staple grain and accounts for 42% of the crop yield for this country. The estimation of crop yield is a topic of global interests (McDonald and Hall, 1980), and the efficient management of agricultural land resources is strongly related to social and economic sustainable development, especially in China. It is well known that China is the largest country in population. However, as the population increases, and economy and industry develop, the quantity and quality of cultivated land is decreasing rapidly. The food supply to the current 1.2 billion people is a serious concern facing China, and will intensify as the population continues to grow. Therefore, it is important to find an efficient way to face this dilemma. Remote sensing technology will provide the needed information on crop distribution, acreage and potential yield.
The use of microwave remote sensing technology to study ecological systems is becoming more and more popular among the world's scientific community (Dobson, 1992; Kasischke and Christensen, 1990). Scientists have carried out extensive field measurements (Ulaby et al 1986), airborne flight mission (Zebker et al, 1991),
Data Source Parameters |
GlobeSAR | SIR-C/X-SAR | RADARSAT (Fine) |
RADARSAT (Standard) |
| Frequency (GHz) | C X 5.30 9.25 |
L C X 1.24 5.3 9.6 |
C 5.3 |
C 5.3 |
| Polarization | HH, HV HH VV, VH VV |
HH HH VV HV HV |
HH | HH |
| Incidence Angle (°) | 14-45 | 34.1 | 43-46 (F4) | 36 -42 (S5), 41 - 46 (S6) |
| Nominal Resolution (m) | 6*6 | 25*25 12.5*12.5 |
10*10 | 30*30 |
| Pixel Spacing (m) | 6*6 | 25*25 | 6.25*6.25 | 12.5*12.5 |
| Swath Width (km) | 18 | 37.8 | 50 | 100 |
| Imaging Date | Nov. 20, 21, 93 | April 18, 94 | In 1996: June 17, Aug. 4, Sept.21, Oct. 15 Nov. 8, Dec. 2 |
In 1996: Mar. 26, Apr. 25, June 10, Aug. 23, Aug. 28, Sept. 16 Nov. 27 In 1997: Apr. 25, May 19, June4,June 12, June 28, July 6, July 22 In 1999: Apr. 17, Apr. 22, Apr. 24,May 11, May 16, May 18, |
Table 1. System Parameters of SAR Data
(Campbell et al, 1995, Guo et al, 1995, 1997, Shao et al, 1995 a, b), spaceborne mission (Stofen et al, 1995). Many microwave backscatter models were developed to study the backscatter behavior of vegetation (Chauhan, et al, 1994; Karam and Fung, 1988; Matzler, 1994; Ulaby, et al, 1990). There are certainly much more literatures available on researches on forest applications, rather than agriculture (Engheta and Elachi, 1982; Freeman et al, 1992; Ulaby, et al, 1990, Brisco et al, 1997). However there are many successful examples in radar remote sensing for agricultural applications (Shao et al, 1994, 1996; Le Toan et al, 1989; Ulaby et al, 1982; Soares et al, 1987, 1997; Schotten et al, 1995; Anys and He, 1995), and a few papers specifically on rice monitoring (Shao, et al, 1997 a, b, c, d; Liu et al, 1997; Kurosu et al, 1995, 1997; Le Toan et al, 1997; Aschbacher et al, 1995 ). The results on rice monitoring using SAR technology were very promising.
This paper presents the results of a study examining the backscatter behavior of rice using multi-temporal RADARSAT dataset. A rice-type distribution map was produced, showing 4 types of rice with different life spans ranging from 80 days, to 120-125 days. The life span of a rice crop has significant impact on yield, as well as the taste and quality of the rice. The yield of three counties and two administrative regions, totaling 5000 square kilometers, are estimated in this study. The accuracy was found to be 91%, providing confidence that multi-temporal RADARSAT data is capable of rice monitoring and yield estimation. Based on previous studies in the Zhaoqing test site, it is suggested that rice yield estimations require three radar data acquisitionstaken at 3 stages of crop growth: at the end of the seedling development period, in the ear differentiation period, and at the beginning of the harvest period. Alternatively, if multi-parameter radar data is available, only two data acquisitions are required: at the end of the seedling period, and at the beginning of the harvest period. Finally, this paper proposes a pilot scenario for operational rice monitoring and yield estimation.
2. Test Site and Data Source
The Zhaoqing test site is located in Guangdong Province, south of China center at latitude 22.30, longitude 112.30. It is sited at the northwestern end of Pearl River Delta. Airborne SAR firstly imaged the test site in 1993 under the GlobeSAR program (Shao, 1995, 1996; Guo 1997). The Shuttle Imaging Radar C-band (SIR-C) and X-band SAR (X-SAR) also flew over the area on April 18, 1994. In addition, there were multiple RADARSAT images acquired from March to December in1996, and from April to July in 1997. The system parameters, imaging modes, and acquisition dates of images used in this study are listed in table 1. For Zhaoqing test site, there was a three years continues RADARSAT data acquisition for rice monitoring. The first two years data acquisition focus on studying the backscatter behavior of rice and the potential of RADARSAT for rice yield estimation. The third year's data acquisition is for operational rice monitoring and yield estimation.
3. Rice Calendar
In the Zhaoqing test site, there are two crops per year; early season rice and late season rice. There are five major growth periods in the life cycle of rice. 1) Transplanting period: rice plant seedlings are transplanted from the seedbed to the paddy field. The transplanting date depends on the weather, especially temperature; 2) Seedling developing period: the seedling splits up and begins to develop a root system; 3) Ear differentiation period; 4) Heading period: headings begin to form; 5) Mature period: the rice plants mature and are ready to be harvested. Temporally, these five periods for early season rice are March 25 ~ April 5, April 15 ~ 25,
May 10 ~ 30, June 10 ~ 25, July 5 ~ 31. For late season rice, the growth stages occur as follows: July 20 ~ August 5, August 10 ~ 20, September 1 ~ 30, October 1 ~ 20, November 1 ~ 25 respectively. The exact dates of each period vary with the weather condition during the whole life of rice and are totally rice life span dependent. Plant maturity rates and life span are species dependent.
This rice calendar only illustrates a conventional law of rice growing behavior.
With progress in agricultural technology, it is known that a long growing season is favorable to high photosynthetic rates and high rice yield. For this reason, the transplanting date of rice is getting earlier and the life span is getting longer. For instance, the local farmers grow three rice crops a year in past, but currently, they only grow two crops a year. In USA, there is only one rice crops a year. It starts in early may then harvests in end of October. The rice life circle is around 170 ~ 180 days. The yield reaches 6.2 ton/ha, which is higher than rice yield of 5.9 ton/ha in China. Improved varieties and cultural practices have contributed significantly to yield in last decades too.
4. Temporal Backscatter Behavior of Rice
This study aimed at understanding the backscatter behavior of rice over its whole life circle, and the relationship between the rice structure parameters and its backscatter coefficients. Based on a good understanding of backscatter characteristics of rice, we can then suggest the best date for radar data acquisition to monitor the rice growth and estimate the yield. Figures 1 and 2 show the backscatter coefficients of rice extracted from calibrated RADARSAT images acquired from summer to autumn in 1996 and from spring to summer of 1997 respectively. These are produced by applying the calibration procedure to the digital number of Radarsat image to produce for a training area (20 pixel by 20 pixel) of the multi-temporal RADARSAT dataset.
In figure 1, there are five types of rice: 1) medium mature rice. 2) late mature rice. The backscatter coefficient of late mature rice is slightly higher than medium mature rice on August 4. 3) Late transplanted rice, which has been transplanted about 25 days later than the normal spring rice, due to a cold spring and low temperatures in 1996. The RADARSAT image acquired on April 25 shows the backscatter coefficients of these fields as very low, responding similar to a still water body. The backscatter coefficients from the RADARSAT images taken after April 25 increase gradually, until September when the rice is harvested. Its life span is relatively short, and the backscatter coefficients are lower. 4) single rice and 5) autumn rice. These two rice types demonstrate how farming activities can change within a year, switch between fishponds and rice fields. Spring rice was a rice field in spring then switched a fishpond in autumn. Conversely, Autumn rice was a fishpond in spring then switched to rice field in autumn. This practice is very common in Zhaoqing area, as well as many other parts of China. In general, this swap of fishpond and rice implies that the quality of the field is less than excellent, and that the rice yield is normally lower. Based on the classification results of the 1996 multi-temporal fine mode RADARSAT dataset, A land cover map was produced clearly showing the five types of rice, as well as other vegetation covers and targets.
Figure 1. Backscatter coefficients of rice as a function of time (1996)
In figure 2, there are four major types of rice with different growth cycles. With the knowledge gained from the rice backscatter behavior of 1996, we can distinguish between rice types quite easily. They are early mature rice, medium mature rice, medium-late mature rice, and late mature rice. Their life span is about 80 days, 100 -115 days, 110-120 days and 120-125 days respectively. A false-color fusion image is produced based on the multi-temporal Standard mode RADARSAT images acquired in the spring of 1997. A rice distribution map is produced using the 7 RADARSAT scenes. If a type of rice has higher backscatter value in April, at the early stage or seedling development period, as found with late mature rice, then it was transplanted earlier than the other types of rice. If a type of rice has higher backscatter coefficients in July, the late stage or the harvest period, the rice is harvested later than other type of rice, as found with medium-late mature rice. From figure 2 we can conclude that late mature rice has the longest lifetime. The medium-late mature rice was transplanted a few days later than medium mature rice and late mature rice, and harvested later. The medium mature rice was harvested before July 22. From figure 2, we can see that early mature rice was harvested before June 28. It has very lower backscatter coefficients since early harvest period. It is concluded that from the early stage, 25 days after transplanting, and late stage, 80 days after transplanting, we can distinguish the rice life span easily, which is an important information for rice yield estimation. The life span is directly related to the yield and quality of rice. This information is provided by local agronomists. It has been proved in real life. The farmer used to grow three rice crops a year in past, nowadays, they only grow two rice crops a year and they make more money from rice farming.
Figure 2. Backscatter coefficients of rice as a function of time (1997)
Based on the knowledge we accumulated in our researches in Zhaoqing test site started in 1993, we propose a theoretic backscatter model for rice. It is mainly suit for rice monitoring using RADARSAT data. The backscatter coefficients from flooded rice field before transplanting, in another words, the water surface recorded by RADARSAT is around -25 dB. After transplanting, in the first 20 days, the backscatter coefficients from paddy rice field are -15 dB. Then entering the seedling-developing period, from 20 days to 45 days, the backscatter coefficients from paddy rice field increase to -12 dB. From 45 days to 75 days, within the ear differentiation period, the backscatter coefficients from paddy rice field increase to -9 dB. In the middle of ear differentiation period, about 65 days, the backscatter coefficients reach the highest point -6 dB. Then gradually drops to -8 dB and keep this value during the mature period. The theoretic backscatter model is expressed in figure 3. It is produced by cubic polynomial that simulated the backscatter coefficients of rice as function of time. It mainly represents the backscatter behavior of medium-late mature rice. Medium-late mature rice is transplanted in April 1 and harvested in July 30. The late mature rice and medium mature rice are transplanted earlier than medium-late mature rice. Medium mature rice is harvested earlier than medium-late mature rice. This model can be applied to other type of rice and modified in accordance with rice life span, transplant date and harvest date. Then we can use this model to calculate the backscatter coefficients of rice at certain growth stage. During the mature period of spring rice, it is flooding season (Guo et al, 1999), so the backscatter coefficients of rice can be much higher than normal value since the high moisture or dew effect.
The results from the studies carried out in 1996 and 1997 imply that the most important radar data acquisition time for rice monitoring is at the end of the seedling period and beginning of ear differentiation period, which is late April for spring rice and late August for autumn rice. For calculating the acreage of planted rice, two radar data acquisitions are required: one near the end of April, and another in the middle of May during the ear differentiation period. The first data acquisition tells the biggest difference between rice and other targets. The two data sets acquired in two dates is used for automatic classification. For rice yield estimation, three radar data acquisitions are required: one at the end of April, one in the middle of May, and one at the end of June. For autumn rice or late season rice, the corresponding date can be found by checking the rice growth calendar.
5. A Proposed Scenario for Operational Rice Monitoring and Yield Estimation
The value of a given data set and research project is measured in its practicality, and applicability to the real world. Researchers at the Institute of Remote Sensing Applications (IRSA), through a program entitled "SAR Technology for Rice Growth Status Monitoring and Land Cover Mapping" under the support of National High Technology Program, have strive to provide information on rice growth status and yield to both national and local government agencies, as well as to commercial users. The final goal of this project is to establish an operational rice monitoring and yield estimation system. This paper presents the research achievements of this program. Figure 4 outlines the operating procedures for rice monitoring and yield estimation as followed for this study. The entire working procedure can be divided into four parts:
5.1. Data Input:
5.1.1. Multi-Temporal and Multi-Parameter Radar Data
The radar data provides the latest information on rice growing status and the accurate acreage of planted rice. For single frequency and polarization radar data like RADARSAT, two or three data acquisitions are required: one about 20 days after the transplanting period, and one at the end of seedling development period. For the multi-parameter radar data (i.e. multi-frequency or multi-polarization), one data acquisition is required during the early stage of rice growth. The another critical radar data acquisition is at the end of the heading period.
5.1.2.Static Supporting Data
A digital elevation model (DEM) is required to understand the rice distribution limitations and to aid in image processing. In addition, soil maps help to know the soil quality and its effect on rice production. Knowledge of the crop calendar helps to decide the optimum radar data acquisition date, and explain the rice growth stage at the acquisition time. The crop calendar may shift a few days at the transplanting period, in accordance with the change of temperature.
5.1.3.Dynamic Supporting Data
Meteorological data, including temperature, total sun illumination, rainfall etc. are important in deciding the conventional yield of rice on a regional scale. Knowledge of tillage activities provides information on the seeding of the rice and transplanting methods, which is directly related to the yield per unit area.
In the early stages of the study, data sources B and C determine the rice yield model. Data source C will be a dominant factor in determining the annual yield.
5.2. Supporting Database
There are a series of database and numerical models to support the image processing and the yield estimation work.
- The rice yield model is established based on the input source of static and dynamic data, which directly go to join the rice yield estimation procedure.
- Database ground control points (GCPS) are collected and used to register the acquired imagery to the stored DEM, as well as to other images from different dates or sources.
- Representative training areas for classification are selected for both rice and other targets. In general, the sites should be regular and stable.
- A rice theoretic backscatter model is a numerical model developed based on radiation transfer theory, which is used to explain rice backscatter behavior. A rice experiment backscatter model is established by studying the multi-temporal RADARSAT data of 1996 and 1997. These two models are used to classify the radar data.
- Registration of the imagery to existing data and maps using a DEM and GCPs.
- Filtering to remove the speckle noise and small pitches before and after classification.
- Extraction of linear features such as ditches between fields. The width of these linear features is enlarged by the corner reflection effect, allowing the linear features to be easily seen on an image. The area of these linear features should be calculated and considered in the estimation of crop acreage. Linear feature extraction method and the amount of area that should be considered in acreage estimations is a topic in need of further research.
- Classification procedure to separate different types of rice and assess the separability of rice from other targets. In this study, a neural network classifier (Liu and Shao, 1996) from the PCI software package was used to classify the multi-temporal RADARSAT data.
- Based on the classification and linear feature extraction, a precise calculation of the acreage of planted rice can be made and combined with the rice yield model and rice experiment backscatter model to make the first yield estimate and to forecast the acreage of planted rice. Meteorological data can help in the estimation and forecasting of the final yield after the harvest. During the harvest season, there may be some hazardous weather such as summer storms and flooding.
- Creation of new maps, and updating of existing databases with new results and knowledge.
A.
Statistical results of the acreage of rice and other targets, based on county and provincial administration boundaries:
Table 2 shows the statistical results for Shihui County, Gaoyao County, Shanshui County and Dinghu administration region, and Duanzhou administration region, covering total area of 4998.1 square km. The yield is 404227 mu ´ 0.45 ton = 181, 902.15 ton for medium mature rice, 379179 mu ´ 0.55 ton = 208,548.45 for late mature rice, 349272 mu ´ 0.5 ton = 174,636.0 ton for medium-late mature rice, and 51366 mu ´ 0.35 ton = 17,978.1 tons for early mature rice. The rice yield of per unit, in this case it is mu, is provided by local agronomist. The average accuracy of the classification is 90.1%, and the overall accuracy of the classification is 91.49%. The classification accuracy is examined by PCI software package. The classification correctness is checked by ground truth verification.
| Class | Acreage (m2) | Percentage of Total Area (%) |
| Euryale Ferox 1 | 87,012,656.0 | 1.74 |
| Euryale Ferox 2 | 54,568,124.0 | 1.09 |
| Medium Mature Rice | 269,454,368.0 | 5.39 |
| Late Mature Rice | 252,760,230.0 | 5.06 |
| Medium-late Mature Rice | 232,824,528.0 | 4.66 |
| Rice 4 | 34,240,312.0 | 0.69 |
| Grassland | 430,799,520.0 | 8.62 |
| Banana Field | 35,338,124.0 | 0.71 |
| Sugar Cane Field | 43,732,656.0 | 0.87 |
| Residence Areas | 59,249,844.0 | 1.19 |
| Water Body | 320,362,016.0 | 6.41 |
| Forest | 3,166,174,976.0 | 63.34 |
| Beach Land | 11,598,906.0 | 0.23 |
Table 2: Statistical Acreage Results of Land Cover Classes in Zhaoqing
B. Maps:
According to the meteorological data, 1997 was a good year for agriculture. For this reason, the yield should be at the 'normal' level. In order to classify the rice, it is important to separate it from other targets before considering the different rice types (Shao Yun et al, 1997 a,b,c,d). In 1997, the output of this project includes four map products (Guo, ed. 1999 a, b): 1) the multi-temporal RADARSAT image map of Zhaoqing test site; 2) the land cover map; 3) the crop distribution map; 4) the rice distribution map. In 1999, the output of this project includes two map products (Guo, ed. 1999 a, b): 1) the multi-temporal RADARSAT image map of Pearl River Delta test site; 2) the rice distribution map.
The multi-temporal RADARSAT image map was produced using the following procedures: 1) co-registration of the seven scenes by selecting GCPs. They are stored in the database for future use. 2) Generate a false-color image by combining the July 22 (red), the mean intensity of the June 4, 12, 28 images (green), and the mean Intensity of images acquired on April 25 and May 19 (blue) (Figure 7). 3) Based on ground truth data accumulated since the GlobeSAR mission in 1993, the backscatter behavior of rice and other targets at certain times in this region is known well. Defining training areas on the multi-temporal RADARSAT images, then running a neural network classifier produces the other maps. The training areas should be defined as precisely as possible in order to obtain an accurate classification output. 5) Filter and enhance the classified results to remove speckle and within field variability and produce a land covers map. 6) A rice distribution map is created by combining targets so as to display only the different rice crops, water bodies, forest land, urban areas, and non rice targets. After the filtering and enhancement, the overall accuracy of the products is 97%. The quality of these products created from multi-temporal RADARSAT data demonstrate the feasibility and effectiveness of RADARSAT data for agricultural applications such as rice crop monitoring.
6. Concluding Remarks
Multi-frequency and multi-polarization SAR imagery is very useful for separating agricultural vegetation types. Multi-temporal RADARSAT images, with single frequency and polarization, can also provide useful monitoring capabilities. The fine mode imagery has shown almost the same capabilities for discriminating the vegetation as the airborne SAR data collected under the GlobeSAR program. However, the multiple RADARSAT data can not only separate the different targets, but can also provide information on the growth status of rice. There still exists a need for better image processing to extract linear targets, and for classification of the image. Additional work is needed to understand the interaction mechanism between the plant parameters and the radar signal as a function of frequency and polarization. It is anticipated that modeling research will continue to provide a clearer, and better explanation and understanding. In addition, recently developed polarimetric (Ulaby et al, 1991) and interferometric research using multi-parameter SAR systems will provide even more information on all land cover targets.
Acknowledgments
This research work is founded by the 863 National High Technology Program, sponsored by Ministry of Science and Technology. The authors wish to take this opportunity to express their sincere acknowledgment to the members of the 863-308 Expert Group: Prof. H. D. Guo, J. M. Xu, Z. M. Yang, Z. Q. Wei, X. G. Ling, X. T. Zhou, G. Q. Ni, and Dr. .Y. J. Wu of MST, Mr. J.L. Miao of the 863-308 office. Special appreciation goes to Dr. Fred Campbell of the Canada Center for Remote Sensing for his kind support to this research.
References
(In Chinese)
- Guo Huadong, Shao Yun (1997), Airborne dual-band, full polarimetric SAR information analysis, Journal of Remote Sensing, 1(2), 94-100. Science Press.
- Liu Hao, Shao Yun, Wang Cuizhen (1997), Multi-temporal RADARSAT data for Rice classification in Zhaoqing, Guangdong Province, Land Resource Remote Sensing, 4:1-6.
- Shao Yun (1999 a), In Analysis of Radar Remote Sensing Imagery in China (Guo Huadong, ed.), Science Press, Section of Agriculture.
- Shao Yun (1999 b), In Radar for Earth Observation - Principle and applications (Guo Huadong, ed.), Science Press, Chap.9.
- Anys, H. and He, D. C. (1995), Evaluation of textural and multi-polarization radar features for crop classification, IEEE Trans. Geosci. Remote Sensing, 33(5): 1170-1181.
- Aschbacher, J., Pongsrihadulchai, A., Karnchanasutham, S., et al. (1995), Assessment of ERS-1 data for rice crop mapping and monitoring, in Proc. IGARSS, Florence, Italy, pp. 2183-2185.
- Brisco, B. (1998), In Manual of Remote Sensing (3rd Edition, F. Henderson & A. Lewis ed.), Chap. 4
- Campbell, F., Ryerson, R. and Brown, R. (1995), GlobeSAR: a Canadian radar remote sensing program, GEOCARTO International, Vol. 10, No.3, 3-8
- Chauhan, N. S., LeVine, D. M., Lang, R. H. (1994), Discrete scatter model for microwave radar and radiometer response to cone: comparison of theory and data, IEEE Trans. on. Geoscience and Remote Sensing, 32, 416- 426.
- Dobson, M.C., Pierce, L., Sarabandi, K., Ulaby, F. T., Sharik, T. (1992), Preliminary analysis of ERS-1 SAR for Forest Ecosystem Studies, IEEE Trans. Geosci. Remote Sens. 30:203-211
- El-Rayes, M. A., and Ulaby, F. T., (1987), Microwave dielectric spectrum of vegetation, part I: experimental observation, IEEE Trans. Geosci. Remote Sens. 25, 541-549.
- Engheta, N. and Elachi, C. (1982), Radar scattering from a diffuse vegetation layer over a smooth surface, IEEE Trans. on Geosci. and Remote Sensing, Vol. GE-20, No. 2,. 212-216.
- Freeman, A., Durden, S. and Zimmerman, R., 1992, Mapping sub-tropical vegetation using multi-frequency multi-polarization SAR data, in Proc. IGARSS'92, Houston, TX, pp. 1686-1689.
- Guo Huadong, Wang Chao, Liao Jingjuan, Shao Yun, Wei Chengjie (1995), Dual-frequency and Quad-polarization SAR Observations in Zhaoqing
- Karam, M. A. and Fung, A. K. (1988), Electromagnetic scattering from a layer of finite length, randomly oriented, dielectric, circular cylinders over a rough interface with application to vegetation, Int. J. Remote Sens., Vol. 9, No. 6. 1109-1134.
- Kasischke, E.S. and Christensen, N.L. Jr. (1990), Connecting Forest Ecosystem and Microwave backscatter Models. Int. J. Remote Sensing. 11:1277-1298
- Kurosu, T., Fujita, M. and Chiba, K. (1997), The identification of rice fields using multi-temporal ERS-1 C band SAR data, Int. J. Remote Sensing, 18(14): 2953-2965.
- Kurosu, T., Fujita, M., and Chiba, K. (1995), Monitoring of rice crop growth from space using the ERS-1 C-band SAR, IEEE Trans. Geosci. Remote Sens. 33(4): 1092-1096.
- Le Toan, T., H. Laur, E. Mougin, and A. Lopes (1989), Multi-temporal and dual-polarization observations of agricultural vegetation covers by X-band SAR images, IEEE Trans. Geosci. Remote Sens. 27:709-717.
- Le Toan, T., Ribbes, F. , Wang, L. F., et al. (1997), Rice crop mapping and monitoring using ERS-1 data based on experiment and modeling results, IEEE Trans. Geosci. Remote Sens., 35(1):41-56.
- Liu Hao, Shao Yun (1996), The neural net classifier applied to the GlobeSAR Data for land use mapping, Proc. Second Asia Regional GlobeSAR Workshop, Science Press, 73-76
- Matzler, C. (1994), Microwave (1-100GHz) dielectric model of leaves, IEEE Trans. Geosci. Remote Sens. 32, 947-949.
- McDonald, R. B., and Hall, F. C., (1980), Global crop forecasting, Science, 208:670-679.
- Region, China, GEOCARTO International, Vol. 10, No.3, 79-86.
- Schotten, C. G. J., van Rooy, W. W. L., and Janssen, L. L. F. (1995), Assessment of the capabilities of multi-temporal ERS-1 SAR data to discriminate between agricultural crops, Int. J. Remote Sens. 16, 2619-2637.
- Shao Yun, et al, (1995b), The GlobeSAR Data for Vegetation Discrimination, Microwave Remote Sensing for Earth Observation, Science Press, 195-201
- Shao Yun, F. Verjee, S. Staples (1997a), Cutting through cloud, GIS Asia Pacific, October/November, Singapore.
- Shao Yun, Fan Xiangtao, Wang Cuizhen, Liu Hao (1997b), Estimation rice growth stage using RADARSAT data, Proc. of Geos. Remote Sensing,IEEE, vol.4, Singapore.
- Shao Yun, Guo Huadong, Liu Hao, Brian Brisco (1994), Multi-frequency, Multi-polarization GlobeSAR Data for Land Use Mapping, Proc. First Regional GlobeSAR Workshop, Thailand, 32-45.
- Shao Yun, Guo Huadong, Liu Hao, Li Junfei, Lu Xinqiao (1995a), Effect of Polarization of GlobeSAR Data on Vegetation Discrimination, GEOCARTO International, Vol. 10, No.3, 71-76.
- Shao Yun, Guo Huadong, Liu Hao, Lu Xinqiao (1996), GlobeSAR Data for Agriculture Applications-Potentials and Limitations, Proc.Second Asia Regional GlobeSAR Workshop, Science Press, 79-83.
- Shao Yun, Liu Hao, Lu Xinqiao, Guo Huadong (1997d), Effect of Frequency and Polarization on Target detection, IEAS'97 & IWGIS'97, Beijing.
- Shao Yun, Wang Cuizhen, Fan Xiangtao, Liu Hao (1997c) Evaluation of SAR Image for Rice Monitoring and Land Cover Mapping, Presented at Geomatics in Era of RADARSAT, Ottawa, Canada
- Soares, J. V., Bernard, R. D. and Vidal-Madjar, D. (1987), Spatial and Temporal behavior of a large agricultural area as observed from airborne C-band scatterometer and thermal infrared radiometer, Int. J. Remote Sensing, 8(7):981-996.
- Soares, J. V., Renn, C. D., Formaggio, A. R., Yanasse, C. F. and Frery, A. C. (1997), An investigation of the
- selection of texture features for crop discrimination using SAR imagery, Remote Sensing of Environment, 59:234-247.
- Stofan, E. R., Evans, D. L. and Schmullius, C., et al. (1995), Overview of results of spaceborne imaging radar-C, X -band synthetic aperture radar (SIR-C/X-SAR), IEEE Trans. Geosci. Remote Sens., 33:817-828.
- Ulaby, F. T., Li, R. Y. and Shanmugan, K. S. (1982), Crop classification using airborne radar and Landsat data, IEEE Trans. Geosci. and Remote Sens., Vol. GE-20, No. 1, 42-50.
- Ulaby, F. T., Moore, R. E. and Fung, A. K. (1981-1986), Microwave remote sensing: active and passive, Vols. 1, 2 and 3, Addison-Wesley Publishing Company.
- Ulaby, F. T., Pierce, L., Oh, Y., Sarabandi, K., and Dobson, M. C. (1991), Scene classification and radar polarimetry for mapping of soil and vegetation parameters, in 6th Int. Symp. Physical Measurements and Signatures in Remote Sensing, Val D'Isere, France, Jan. 17-21.
- Ulaby, F. T., Sarabandi, K., McDonald, K. C., Whitt, M., and Dobson, M. C. (1990), Michigan microwave canopy scattering model (MIMICS), Int. J. Remote Sens., 11: 1223-1253.
- Zebker, H.A., Van Zyl, J.J., et al (1991), Calibrated imaging radar polarimetry: Techniques example and applications, IEEE Trans. Geosci. Remote Sens., 29:942-961.