GISdevelopment.net ---> AARS ---> ACRS 1999 ---> Land Use

Evaluation of Filtering and Classification Techniques for Floodplain Land Use/Cover Mapping using Fadarsat Sar Data

Quazi khalid Hassan, Timothey C.Martin, khaled Hasan, Ahamadul Hassan and Md.Shawkat Ali
EGIS( Environment and GIS Support project for Water Sector planning
House 49, Road 27, Banani, Dhaka-1213, Bangladesh
Tel:+880-2-881570-2;Fax:+880-2-883128 E-mail:http://www.gisdevelopment.net/aars/acrs/1999/ts5/qhassan@cegisbd.com

Abstract
The subtropical country like Bangladesh has perennial cloud cover in the monsoon season which creates difficulties in mapping floodplain land use/cover using optical remote sensing data. For its all weather imaging capabilities, RADARSAT SAR is a potential image data source to map the seasonal dynamics in the floodplain. Bangladesh occupies the lower portion of two of the largest rivers of the world, the Ganges and the Barahmaputra- Jamuna, and large amount of water and sediments flow through them in monsoon season. This flow causes recurrent flooding in 25%to35% of the country in an average year and it can nesting site as high as about 60%-65% of the total area of the country.

To support a reliable floodplain land use/cover mapping tool, image processing techniques including filtering an classification algorithms - were evaluated for extraction of information from RADARSAR SAR data. "Two adaptive filter, Gamma-map and Frost were tested on the data. Nine RADARSAT Fine Beam(F3)images was used to cover temporal variability of land use/cover during the monsoon of 1998.This molti-temporal data set was classified using a maximum likelihood classifier. The two filtering techniques influenced the accuracy level for land use/cover dynamics classification. The accuracy of frost filtered dataset (i.e.69%)was better than that of gamma-map filtered dataset(i.e. 64%)n . The advantages and disadvantages of the filtering techniques for floodplain land use/cover mapping are discussed and potential of satellite SAR as a floodplain land use/cover mapping tool in Bangladesh is demonstrated.

1. Introduction
Satellite radar remote sensing is being explored and increasingly utilized for land use/cover mapping during monsoon seasons of subtropical countries. To investigate the potential of RADARSAT SAR for study of floodplain dynamics, this project was launched in collaboration with RADARSAT SAR International under the Advanced Data Research Opportunity (ADRO Project 418). This work builds upon that reported in earlier studies (Hasan et al. 1998, Martin et. al. 1998). The main focus of this study was to review image processing techniques, including filtering and its impact on digital classification, as well as the utility of multi-temporal image dataset.

Filtering is essential to reduce the speckle of SAR data and different filtering techniques have been developed by researchers. Commonly used filters are adaptive filters, such as Lee (Lee 1980), Frost (Frost et. al.1982), Kuan (Kuan et al., 1985) Gamma- MAP (kuan et al., 1987). Gamma -Map filter has been reported suitable for flood monitoring using SAR imagery in Bangladesh (Fap 19,1995;EGIS 1997). On the other hand the Frost has been reported better for crop type and land cover discrimination (Connery et al.,1996) using SAR imagery . In this study, Gamma- Map and Forest filter were evaluated to determine their effectiveness in land use/cover discrimination .

The simple method of radar data classification is density slicing (Richards 1986). FAP 19 (1995) used this method to classify the SAR images into four classes:

(1) urban/homestead areas;(2) non- flooded cropland; (3) folded crop; and (4) river, fold water and bells. FAP 19 found overall accuracy of more than 80% Dlorio et al. (1995) evaluated RADARSAT performance in identifying land use using an unsupervised classification. Minimum distance and parallelepiped classification algorithms, which required multiple images of the same area Wu (1984) used a maximum likelihood classifier to classify the land cover into eight classes and have found overall accuracy of 59% using SAR imagery. Nieuwenhuis and Schotten (1992) used maximum like hood classifier for land cover monitoring in the Netherlands. They reported that classification accuracy varied from 60 to 90% .

Many researchers (Hoogeboom1983, Bush and Ulbay 1978, Brisco et al., 1984 Foody et al. 1989) have classified crop types from multi-temporal SAR data and obtained significant results, whereas single data SAR can map the extent of open water flooding (FAP 19, 1995;EGIS 1997). Pope et al. (1994) concluded that multi-temporal data dramatically improved detection and delineation of tropical vegetation in flooded and non-flooded environments.

2. Study Area and Data used
The study area is a floodplain formed by three distributaries of the Jamuna river: the Elanjane the Pungle and the Lohojong. The central part of the floodplain is protected by embankments and has controlled flooding during the monsoon season. The protected area is locally called the Lohojang floodplain. The entire study area lies within 240 06/ 56//N to 240 25/02//N altitude and 890 46/27// E to 900 00/ 50// longitude types of flooding are common in this area: controlled and natural.

RADARAST F3 images were the primary imagery used in this research. IRS-ID PAN image was used as base map. In addition, aerial photos of 1990 were used for identifying the field monitoring plots. In addition SPOT XSS hardcopy at 1:50,000 scale were used to georeferenced the the RADARSAT SAR and IRS-ID PAN images . On the acquisition dates of the SAR images (Table1) extensive field information described conditions of land cover degree of flooding, crop or vegetation canopy , crop height etc. Field data were arranged in a GIS and the main categories included settlements, permanent water bodies, and seasonal flooding , inundated or non-inundated crops.

Table 1. Description of the satellite data used in this study
Satellite, Sensor and Beam Mode Acquisition date Season
IRS ID PAN Feb.12,1998 Dry
RADARSAT SAR F3 MAY 28, 1998 Pre-mansoon
RADARSAT SAR F3 June 21,1998 Pre-mansoon
RADARSAT SAR F3 July 15,1998 mansoon
RADARSAT SAR F3 Aug 08,1998 mansoon
RADARSAT SAR F3 Sep01,1998 mansoon
RADARSAT SAR F3 Sep 25,1998 mansoon
RADARSAT SAR F3 Oct 19,1998 Post-mansoon
RADARSAT SAR F3 Dec 06,1998 Post-mansoon
RADARSAT SAR F3 Dec 30,1998 Post-mansoon

Methodology
The raw SAR data were converted to obtain SAR backscatter (beta nought-b0) in Db using the formula derived by ALTRIX System (1998). Then this log-scaled (b) image was converted into a power-scaled image plane. The power- scaled SAR data were filtered using frost and gamma-map adaptive filters. All the filtered images were co registered with reference to July 15,1998 image. The July 15,1998 image was a transitional one in between pre flooding and flooding condition. The July 15,1998 SAR data was georeferenced using 1:50,000 scale SPOT hardcopy satellite image. The Georeferncing was perfumed using a first order transformation with cubic convolution resampling at 6.25m resolution.

The processing steps for IRS ID PAN data include co registration with SAR data and was georeferenced using a first order transformation with nearest neighborhood resampling. The purposes of processing of the SAR was to produce a multi-temporal dataset. The IRS ID PAN was processed to produce a base map and to obtain an understanding of land use/cover in the study area.

The images were filtered using 3*3,5*5 7*7 9*9 and 11*11 kernel sizes for the gamma map and frost filters. The results were checked visually and the 11*11 kernel sizes was found better than the other kernel. An unsupervised classification was performed on the both filtered datasets into 140 classes. The signature derived from this operation were analyzed and modified. These modified signature were used to classify the multi temporal dataset using the maximum likelihood classifier. Six broad classes of land use/cover depicting the pattern of dynamics of landuse and water (table 2) were produced from this operation. In the floodplain, the seasonal open water flood extent and change in crop growth stage were of interest. The landuse/cover types were considered according to the dominant crops in Kharif 11 season.

Table 2. Land use/cover classes used in this study
Class no Class name
1 Settlements
2 Permanent water
3 Seaconal flooding
4 Rice (i.e. Aman)
5 Jute
6 Sugarcane

4. Result and Discussions
The supervised maximum likelihood classification on a Gamma-map filtered data for land use/cover type is presented in Table 3 and the same over a Frost filtered dataset is presented in Table 4. Comparison of these two filtered datasets shows differences in settlement, permanent water, seasonal water, aman and sugarcane classes were 87%, 67%,68% 27% and 90%, respectively from the Gamma-map seasonal water , aman and sugarcane classes were 76% 73% 94%75% and 53%, respectively. The overall accuracy of the entire classification from Gamma-map and Frost filtered data ser was 64% and 69% respectively.

Table 3: Classification result over Gamma-map filtered dataset
SAR data
Gro-Und Obs   Settle-Ments (ha) Water (ha) Seasonal water (ha) Rice (ha) Jute(ha) SugarCane(ha) Total (ha) %Accuracy %Omission
Settlements(ha) 83 0 0 0 1 11 95 87 13
Water(ha) 2 17 5 0 0 0 26 67 33
Seasonal Water(ha) 4 2 15 0 0 0 22 68 32
Rice(ha) 8 0 2 6 2 4 23 27 23
Jute (ha) 1 0 0 0 2 2 5 45 55
Sugarcane(ha) 1 0 0 0 1 16 18 90 10
Total(ha) 99 20 23 7 6 34 189    
%Commision 16 15 35 6 33 52      
Overall accuracy 64                

Table 4. Classification results over Frost filtered dataset
SAR data
Ground Obs   Settlements (ha) Water (ha) Seasonal water (ha) Rice (ha) Jute(ha) SugarCane(ha) Total %Accuracy %Omission
Settlements(ha) 73 1 2 6 3 12 95 76 24
Water(ha) 1 16 4 0 0 1 22 73 27
Seasonal Water(ha) 0 1 24 1 0 0 26 94 6
Rice(ha) 1 0 4 17 0 1 23 75 25
Jute (ha) 0 0 1 1 2 1 5 44 56
Sugarcane(ha) 2 0 1 6 1 10 18 53 47
Total(ha) 76 17 34 30 7 25 189    
%Commision 5 2 11 26 12 19      
Overall accuracy 69                

The jute class had similar for the differently filtered dataset. However, in the classification process, the jute class had a different problem Normally, Jute is planted in early March and is harvested in July- August . In classifying jute, the deficiency eal in the available time series of SAR data. In higher land, generally rice (Aman) is planted the same plots immediately after jute is harvested. But due to excessive flooding in 1998. the farmers were unable to plat rice tn these plots.

In Bangladesh, only 5% areas of the settlement area are urban settlements with concrete buildings and other structures. The remaining 95% are in rural settlements and most of the houses are constructed of straw, tin shade bamboo, wood, etc,. Due to these complex cover types, it was difficult to identify these rural settlements from the SAR data. In the analysis of the signature, some limitations were found, especially in case of smooth hard -surface roads. These feature have a signature similar to that of permanent water. Another limitation was found in the SAR imagery the presence of heavy rainfall, where backscattering was higher by around 4 dB over water areas.

5.Conclusions
Classification of land use/cover in the Bangladesh floodplain using RADARSAT F3 data was addressed in this study. For reducing speckle in SAR imagery, two filters, Gamma-map and Frost were tested and Frost filter was found better than Gamma-map filter for both crop and flood monitoring. The disadvantage of using the Frost filter was that it takes longer exaction time than the Gamma-map filter ( about 4 times more ). Therefore the Gamma-map filter would be more efficient for near real time flood monitoring.

The ground truth data, acquired at the time of satellite overpass, were used to evaluate the classification of the radar data. By comparison of the SAR classification and ground observation data, the overall accuracy of gamma-map and frost filtered dataset were 64% and 69% .

The study area has a complex heterogeneous land use/cover. As it was possible to classify this complex land use/cover, it is expected that other parts of the country, where less complexity in land use/cover found could also be classified using similar SAR images. Hence, this study has demonstrated the potential of using radar data in monsoon season as a floodplain land use/cover mapping tool.

The study period of 1998 was an excessive flooding year and a similar study to assess the potential of SAR data in an average flooding year could be useful. Such a study, combined with the result of this project, would provide firmer conclusions about the potential of using SAR to manage flooding dynamics in Bangladesh.

Acknowledgements
The F3 images from May'98 to August'98 were by RADARSAT International under the Application Development Research Opportunity (ADRO) Project NO 418. This study was a collaborative project of the Tan Gail Compartmentalization Pilot project (CPP), with EGIS. Both CPP and EGIS are project under Water Resources Planning Organization of Ministry of Water Resource of Bangladesh Government. EGIS is executed with funding from the Royal Netherlands and Bangladesh Government. The support from Ir. Rob Koudaastal, Team Leader, EGIS is highly appreciated RADARSAT data are projected under copyright by the Canadian Space Agency, received by Canada Centre for Sensing and pre-processed by RADARSAT International Inc.

References
  • Altrix System, 1998, Extraction of Beta Nought and Sigma from RADARSAT CDPF products, Report No AS97-5001,Rev.2
  • Brisco, B., F.T. Ulaby and R. Protz ,1984, Improving crop classification through Attention to the Timing of Airborne Eadar Acquisitions, Photogrammetric Engineering , Vol. 50, precursors 739-745
  • Bush, T .F. and F.T. Ulaby, 1978, An Evaluation of crop classifier, Remote Sensing of the Environment, Vol. 7, precursors. 15-36
  • Lee, J.S., 1980, Digital Image Enhancement and Boise Filtering, By Use of Local Statistics, IEEE Trans. Pattern Analysis and Machine Intelligence, Vol 2 (2) , precursors. 165-168
  • Diorio , M.S., Vivulsresth. D.Dowrenag, C. Soalpathong, S. Polngam and H. Gordon, 1995, Identification of Agriculture and Land Use Practices in Southern Thailand from SAR Data , Canadian Journal of Remote Sensing, vol 21,precursors.165-176
  • EGIS 1997 , Mapping and Monitoring Floods with RADARSAT Images, Environmental and GIS Support Project for water Sector Planning/ Delft Hydraulics', Water Resources Planning Organization ( WARPO), Ministry of Water Recourses, Dhaka, Bangladesh.
  • Flood Action Plan ( fap 19,GIS ) 1995, Satellite Based Radar for Floods Monitoring in Bangladesh, ISPAN/USAID, Flood plan Coordination Organization (FPCO)m Ministry of Water Resources ,Dhaka, Bangladesh.
  • Goody, G.M. and P.J Curran , G.B Groom and D.C. Munro, 1989, Multi- temporal
  • Air bone Synthetic Aperture Radar Data for crop Classification, Geocarto International, Vol. 3, precursors.19-29
  • Frost, V.S., J.A. Stiles, K.S. Shanmugan and J.C. Holtman, 1982, A Model for Radar Images and Its Application to Adaptive Digital filtering of Multiplicative Noise, IEEE. Trans. Pattern Analysis and Machine Intelligence, Vol 4(2) , pp. 157-166
  • Hasan, K., T.C. Martin, S.A. Shireen and A.M Choudhary, 1998, Dynamics of monsoonal folding in Bangladesh using RADARSAT imagery, In the Proceedings of Symposium RADARSAT ADRO, Canada, 13-15 October 1998, p.1-13 (on CD)
  • Hoogeboom, P, 1983 Classification agricultural crops in Radar Images IEEE Transactions on Geosciences and Remote sensing , Vol. GE-21, pp. 329-336
  • Kuan, D.T., A.A.Sawchuk, T.C. Strand and p. Chavel, 1987,Adaptive Restoration of Images with Speckle ,IEEE Trans, ASSP., vol 35(3), pp. 373-383
  • Martin, T.C., K. Hasan, A Rahman and D. Werle, 1998, Satellite- Based SAR for Bangladesh land and water resource application, In the Proceedings of symposium RADARSAT ADRO, Canada, 13-15 October 1998, p.1-14 (CD)
  • Nieuwenhuis, G.J.A., C.G.J. Schotten, 1992 Land Cover Monitoring with ERS-1 SAR Observation in the Netherlands, In the proceeding of First ERS-1 Symposim-Space at the Service of Our Environment, Cannes, France, 4-6 November 1992, pp.527-531
  • Pope ,K.O., J.M.Rey- Benayas and J.F. Paris 1994,Radar Remote Sensing of the Enviornment , Vol. 48(2),pp205-219
  • Richards , J.A., 1986, Remote Sensing Digital Image Analysis: Introduction, Spring-Verlag, Berlin
  • Wu, S.T., 1984 Analysis of Synthetic Aperture Radar Data Acquired over aVariety of Land Cover , IEEE Transaction on Geoscience and Remote Sensing, GE-22:550_557