Home Articles Assessment of Multi-Temporal Radar Imagery in Mapping Land System for Rainfed Lowland...

Assessment of Multi-Temporal Radar Imagery in Mapping Land System for Rainfed Lowland Rice in Northeast Thailand

S. Kaojarern
Asian Institute of Technology, P.O. Box 4, Khlong Luang
Pathumthani 12120, Thailand.
[email protected]

J.P. Delsol
Groupement pour le Developpement de la Teledetection Aerospatiale
8-10 rue Hermes, F 31526 Ramonville Saint-Agne Cedex, France

Thuy Le Toan
Centre d’Etudes Spatiales de la BIOshpere, 18 Avenue Edouard Belin
31401 Toulouse Cedex 4, France

S.P. Kam
International Rice Research Institute, DAPO Box. 7777
Metro Manila, Philippines

Thailand is the largest exporter of rice in the world market with a reputation of high grain quality. Eighty five percent of the rice area in Thailand is rainfed lowland rice ecosystem and one-half of the rice land is located in the Northeast region. This region is a heterogeneous rainfed lowland rice environment covering a broad area of complex natural system and therefore exhibits diverse conditions of local landscape, sparse tree stands in the rice fields, and light soil with low moisture retention. Mapping land system into various classes will lead to better management and specific strategy applied to each class. This subject has been little studied using remote sensing. The approach chosen have brought together elements that have been already studied elsewhere (in Indonesia by Le Toan et al., 1997; in Vietnam by Liew et al., 1997; in Thailand by Aschbacher et al., 1995; in Japan by Kurosu et al., 1993) but in terms of thematic results there is a new development for mapping land system for rice.

Test Site and Land System Units
The test site is situated in Ubon Ratchathani province, Northeastern Thailand between 14o 30′ and 15o 30′ north latitudes and 104o 30′ and 105o 30′ east longitudes, covering an area of 75×75 km2. Mountain ranges lie in the south as the border between Cambodia and in the east as the border between Loas. Land system classification for rainfed lowland rice environments according to local conditions in the Northeast Thailand (Junthotai et al., 1990), are using a system approach based on a combination of biophysical land components, i.e. vegetation cover conditions, terrain types, relief, surface materials and soils (Table 1).

Materials
Six ERS-2 SAR images [C-band, l=5.6 cm, 23o incidence angle and VV (vertical transmit and vertical receive) polarization, 35 days revisit capacity, and 100 km swath width] representing the rice growing season of 1997 were acquired from June to January of the years 1996 and 1997. The temporal series of these images, due to limited funding, was put on a monthly sequence for the growing season of 1997, starting by June and July 1997 and followed by September, October, and November 1996, and January 1997 images. All images were acquired at the Center of Remote Imaging, Sensing and Processing (CRISP) in Singapore.

Landsat-5 Thematic Mapper (TM) imagery acquired on January 21, 1991, was visually interpreted for identification of general terrain types and land use classes. The 1:50,000 topographic maps (1984) from the Royal Thai Survey department, the 1:100,000 soil map (1991) and the 1:100,000 land use map (1988) from the Land Development Department were used to identify and map land system units with GIS techniques.

Methodology
Two main steps for SAR image processing were pre-processing and classification. The pre-processing step included registration, calibration, conversion to 8 bit data, multitemporal and spatial filtering. Mutitemporal and spatial filtering, considered as the most critical step, involves reducing speckle to a level where the error in classification is acceptable. After pre-processing, the resulting filtered images have in principle 294 equivalent number of looks {(6*3)(7* 7/3) ENL}. According to Le Toan et al., 1997, the ENL ~100 looks, from two multidate images, are allowed to detect changes in radar intensity less than 1 dB with a confidence interval of 80%. The classification procedure comprises selection of training areas, signature analysis, interpretation, and classification. Training samples were selected based on the desired classes of land system units for rice. The resulting 16 signature profiles corresponding to non-rice (9 signature profiles) and rice (7 signature profiles) were delineated. Classification was done using supervised maximum likelihood classifier as this classifier is considered to be the most accurate, compared with others if a set of criteria is met. These criteria include: temporal changes of signatures allowing discrimination of classes, the appropriate numbers and dates of selected images, normal distribution of training samples, and a large number for each training sample. The classification was applied and the results are discussed in Section 6. 

SAR Image Analysis and Interpretation

Temporal Signature Analysis
The analysis of temporal backscatter was based on the understanding of the radar scattering mechanism of wave-rice-water interaction. Knowledge of rice plant morphology, cultivating practices and rice field environment at different growth stages are required. In the area of sufficient water supply, the so-called inundated rice field, the radar backscatter generally increases with time after planting rice. During the sowing or transplanting period, rice is detected by radar as water surface. This interaction gives very low radar response due to specular reflection. As the plants develop tillers during the early vegetative stage, they expand both horizontally and vertically and the radar backscatter increases due to direct scattering of rice plants and multiple reflections between the plants and water surface. At the late vegetative stage before full coverage of the rice canopy where penetration of radar waves to the water surface is still possible, the radar backscatter peaks due to volume scattering of rice biomass and multiple reflections as an interaction between vertical plant structures and horizontal water surface. During the reproductive stage, the plants develop heads, forming panicles and flowering. As there is no significant change in plant biomass, height and density, volume scattering dominates; however wave penetration to water underneath the rice plants decreases leading to a slight decrease, in radar responses. As the crop ripens the plants’ water content decreases which causes a further slight decrease in the radar backscatter. Theoretical simulation (Le Toan et al., 1997) shows that the scattering mechanism is dominated by double scattering between the water surface and the rice plants. The backsccattering coefficient is found to increase from -16 dB to about -8 dB at the saturation level in the case of irrigated rice. Temporal backscatter profiles of land system units; floodplain, lower low-terrace, upper low-terrace, middle terrace, and irrigated low terrace rice areas are presented in Figure 1 (a-e).


Temporal Signature Changes
The temporal signature changes are evaluated to examine the extent to which they can be used to establish a general rule or to modify the temporal signature profiles and perhaps give better results. The evaluation is made in two periods, within (from July to November) and beyond (June and January) the rice season (Table 2 and Figure 2).

Within the rice growing season from July to November, floodplain rice shows high temporal change from -16 to -6 dB (about 10 dB) which coincides with the theoretical curve of irrigated floodplain as in the study of Le Toan et al,.1997. The signature profiles of the lower low-terrace 1 (LL1) and the lower low-terrace 2 (LL2) have moderately high temporal changes from -12 to -6 dB (6 dB) whereas the upper low-terrace 1 (UL1) has moderate temporal change from -12 to -7 dB (5 dB) and the upper low-terrace 2 has moderately low temporal change from -11 to -7 dB (4 dB). All these temporal changes can be considered in general as medium changes (4-6 dB) and can be explained by the increase in the backscatter responses from the planting (flooded condition) to maturity stages. The irrigated low terrace signature profile has less temporal change from -11 to -8 dB (about 3 dB). The irrigated low terrace in the study area does not behave as the usual irrigated rice field because of shallow sandy soil (50 cm depth) that is ineffective in holding water. The rice plants do not benefit much from the irrigation system during the growing season and is largely dependent upon rain water. The middle terrace signature profile has a low temporal change from -9 to -8 dB (1 dB difference) due to lack of image at the planting period.

Beyond the rice growing season, after harvest (January), the floodplain and the irrigated low terrace show high backscatter values at -6 and -8 dB, respectively which indicate a wet soil condition. The land preparation in June showed high signature variability due to surface water and cultivation practices.

Image Selection
Classification scheme is decided based upon the temporal signature pattern of the most appropriate dates of SAR images. The July image can be used to distinguish floodplain and middle terrace from the other classes; January image; floodplain and irrigated low terrace; September image; the lower low-terrace 1 (the early season rice); October image; the lower low-terrace 2; June image; the backscatter values are quite separable and used for the upper low-terrace 2 and also to improve classification accuracy; November image does not give much information (Figure 2).

Classification Results and Discussion
From the prior discussion, floodplain and middle terrace were easily separable on the July image, lower low-terrace 1 on the September image, lower low-terrace 2 on the October image, irrigated low terrace on the January image, and upper low-terrace 2 on the June image. Upper low-terrace 1 did not show any distinct temporal signature, so it was a class that required separation using temporal pattern. From this point, it can be concluded that the separation into seven land system classes requires at least five SAR images; three images within the rice growing season in July, October and September, and two images, one after harvesting in January and the other at the land preparation period in June. The classified image was assessed using a set of ground truth information derived from the GIS land system map (Figure 4) which was created using the same criteria as defined in the characterization of the land system units (Section 2). The overall accuracy obtained from the error matrix showed 73.8%. The accuracy increased to 84.7% when applied with 6 images.

As for practical application to rice growing, it might be more appropriate to map the land system into five units by having two units of low terrace, namely, lower and upper low-terraces with the remaining classes. The accuracies obtained for five classes using five and six SAR images are 88.4% and 91.2% respectively. To generalize the mapping units into broad categories of four classes that focused on one combined unit of low terrace and the remaining classes, the accuracy was higher than 90% either applied with 4, 5 or 6 images.

In mapping detailed land system units (seven classes), accuracy for individual units based on four, five, and six images. The floodplain shows high producers’ accuracy but low users’ accuracy due to the main reason that ground truth information derived from GIS land system map at 1:100,000 scale may not be detailed enough to derive the narrow strip shape of floodplain in the area. Mapping accuracy for lower low-terrace 1 does not show much improvement either classified with four, five or six images. The accuracies for lower low-terrace 2 and upper low-terrace 1 improve significantly when classified with five or six images whereas the upper low-terrace 2 also improve significantly when classified with six images. The middle terrace gives similar accuracy either classified with 4, 5 or 6 images. The irrigated low terrace significantly improves when using five or six images.

Various classified images based on combination of broad and detailed categories of land system for rice and numbers of multi-temporal ERS-2 data (four, five, and six images) are presented in Figures 3-a, 3-b and 3-c.

Conclusions
This study demonstrates the use of multi-temporal ERS-2 data for mapping land system for rice. Even thought this temporal data set was not ideal, from two consecutive years, it was found that this data set was efficient in identifying and mapping land system units through the phenomenology of rice growth stages. The method developed employing supervised maximum likelihood classification gave good results with the necessity of applying pre-processing steps to reduce speckle, using both multitemporal and spatial filters, prior to classification.

In mapping detailed land system units for rice (7 identified classes i.e., floodplain, lower low-terrace 1, lower low-terrace 2, upper low-terrace 1, upper low-terrace 2, middle terrace and irrigated low terrace) by applying six images, floodplain, middle terrace and irrigated low terrace were found to be easily discriminated due to high temporal change in case of floodplain rice (10 dB), unique pattern of radar response at rice planting stage for middle terrace rice (delay in planting) and still moist/wet after harvesting for irrigated low terrace. The four sub-units (lower low-terrace 1 and 2, and upper low-terrace 1 and 2) of low terrace occupying majority of the study area were the focus in this study. The two profiles of lower low-terraces showed early season and normal season rice with 6 dB temporal changes in radar response (minimum in planting and maximum in maturity stage), but shifting in planting periods. The two profiles of upper low-terraces show similar radar responses within the rice season period therefore making it difficult to discriminate them. However when adding an image outside the rice growing period particularly the June image (prior to planting image), these two units were able to be separated easily (-3 dB for upper low-terrace 1and -9 dB for upper low-terrace 2). The overall accuracy of the classified image was 84.7%. In case that a number of images used were selectively decreased to five or four, the accuracy reduced to 73.8% and 55.7% respectively.

For individual mapping units, aiming at accuracy higher than 80%, the classes of floodplain, lower low-terrace 1 and middle terrace require four images, lower low-terrace 2 requires five images, and upper low-terrace 1 and 2 and irrigated low terrace require six images.

For broad categories of land system, when the number of classes reduced from 7 to 5 or 4 units, the accuracy improved significantly with the applying either 4, 5, or 6 images. The overall accuracy was higher than 90%. Selection of SAR images can be as few as four images, i.e. planting stage (July image), vegetative stage (September image), pre-harvesting stage (October) and post-harvesting stage (January image). Addition of the fifth image during land preparation stage (June image) helped to improve the classification result significantly but the sixth image at maturity stage (November image) had minor contribution.

Comments and Recommendations
The use of SAR multitemporal analysis concept for classifying land system for rainfed lowland rice seems to be promising. However a temporal data set for the same rice cycle should be tried with the same methodology and additional studies in other geographical locations should be carried out to validate the classification results. Real time ground truth will help in better understanding the backscatter behavior of rice fields.

References

  • Aschbacher J., Pongsrihadulchai A., Karnchanasutham S. Rodprom C. Paudyal D.R. and Le Toan, 1995. Assessment of ERS-1 Data for Rice Crop Mapping and Monitoring. Proceedings IGARSS, Florence, Italy, July 1995, pp. 2183-2185.
  • Junthotai K., and Mongkolsawat C., 1990. Land Ecosystem Mapping and Soil Fertility Evaluation Using Landsat TM Data: A Case Study in Upper Namphong Watershed Area. Proceeding of the Seminar on Remote Sensing and Water Management, Thailand.
  • Kurosu T., Suitz T., and Moriya T., 1993. Rice Crop Monitoring with ERS-1 SAR: A First Year Result, Proceeding Second ERS-1 Symposium, Germany, pp. 97-101, (ESA: SP-361).
  • Le Toan, T., Ribbes, F., Wang, L., Floury, N., Ding, K.H., Kong, J. A., Fujita, M., 1997. Rice Crop Mapping and Monitoring using ERS-1 Data based on Experiment and Modelling Results. IEEE Transcations on Geoscience and Remote Sensing, vol 35, no 1, January 1997, pp. 41-56.
  • Liew S. C., Kam S. P., Toung C. P., Minh V. Q., Balababa L. and Lim H., 1997.
  • Application of Multitemporal ERS Synthetic Aperture Radar in Delineating Rice Cropping Systems in the Mekong River Delta, Proceedings 3th ERS Symposium on Space at the Service of Our Environment, Italy, 1997.

Table 1: Land system units for rainfed lowland rice in the study area.

Land system Units  Vegetation conditions  Terrain types  Relief/Elevation  Surface material  Soils (Soil series)
1. Floodplain  Paddy field with bare cover of shrubs  Flood plain  Level(115-120m)  Quaternary deposit  Alluvial complex, Aeric paleaquults (Re)
2. Lower low-terrace1  Paddy field with bare/slight cover of scattered dipterocarp trees  Lower fluvial terrace  Nearly level (120-140m)  Alluvial/Eolian deposit  Vertic tropaquepts (Pm)
3. Lower low-terrace2  Aeric paleaquults, Aquic quartzipsament, (Re/Ub)
4. Irrigated low terrace  Aeric paleaquults, Oxic Plinthaquults, (Re and On)
5. Upper low-terrace1  Paddy field with moderate cover of scattered dipterocarp trees  Association of lower and upper fluvial terrace Undulating to nearly level(130-160m)  Alluvial/Eolian deposit  Oxic paleustults (Kt)
6. Upper low-terrace2  Oxic paleustults, Typicplinthustults (Kt/Pp)
7. Middle terrace  Paddy field with dense cover of scattered dipterocarp trees  Upper fluvial terrace  Gently rolling to undulating(150-170m)  Alluvial/Eolian deposit  Typicplinthustults (Pp)

Table 2: Temporal changes of signature profiles for rice classes based on within and outside rice growing season.

Rice Signature classes   Backscatter of SAR signature profile   Temporal stability
Within growing season June to November    
1) Floodplain   -16 to -6 dB (10 dB)   High temporal change (10dB)
2) Irrigated low-terrace    -11 to -8 dB ( 3 dB)   Low temporal change (3dB)
3) Lower low-terrace 1&2   -12 to -6 dB ( 6 dB)   Medium temporal change (6dB)
4) Upper low-terrace 1   -12 to -7 dB ( 5 dB)   Medium temporal change (5dB)
5) Upper low-terrace 2   -11 to -7 dB ( 4 dB)   Medium temporal change (4dB)
6) Middle terrace 3   -9 to -8 dB* ( 1 dB)   Low temporal change* (1dB)
*partial temporal curve, missed 
beginning of the season (Aug image)
Outside growing season In January     Comments
1) Floodplain   -6 dB    Wet soil condition
2) Irrigated low-terrace    -8 dB    Wet soil condition
3) The remaining 5 classes   -10 dB   Dry soil condition
In June    
1) all classes   -12 to -3 dB ( 9 dB) High variability due to water 
surface and cultivation practices