Automated and Semi-automated methods for Mapping Wetlands using multi-sensor High-resolution satellite sensor data and SRTM data

Caption :Ruhuna Basin Sri lanka Tip: Click on Image for high resolution picture

Islam, A., Thenkabail, P. , De Silva, S., Finlayson, M., Alankara, R., Gunasinghe, R., Edussriya, C., Gunawardana, A., and Mariagrazia, B. The overarching goal of this study was to develop a comprehensive methodology for mapping and characterizing natural and human-made wetlands using high- and moderate resolution imagery and secondary data. The dry season satellite sensor data for the nominal year 2000 were used and consisted of orthorectified Landsat ETM+ images of 30-m, IRS LISS III images of 23.5-m, and the IRS Panchromatic images of 5-m spatial resolution.  The 90-m SRTM data and the ground truth data were also extensively used. The methods were tested for the Ruhuna river basin in Sri Lanka which has diverse landscape ranging from sea shore to hilly areas, low to very steep slope, arid to semi-arid zones, and rain fed to irrigated lands. The first and the foremost challenge in wetland mapping is in precise delineation of wetland boundaries from its uplands. Investigations were carried out for rapid delineation of wetlands through automated approaches. The methods investigated were: (a) threshold SRTM slopes, (b) threshold band 5 reflectivity, (c) Tasseled Cap wetness index (TCVI), and (d) thresholds of various two-band vegetation indices (TBVSs). The best of these indices, had an overall mapping accuracy of about 30 percent. If mapping accuracy increases above this, the errors of omissions and errors of commissions become too large. For example, using TCVI the best accuracy of 82 percent was achieved, but the errors of commission was 305 percent.  Algorithms written for automated delineation of drainage either over-estimated or under-estimated the drainage densities and frequencies. When 10 SRTM upstream pixels are chosen to constitute a stream, there is a 400 percent higher drainage densities than that generated from Landsat ETM+. However, more glaring was the inconsistency in the precise physical location of the drainage and often generation of non-existent drainages.

Semi-automated approaches involved image enhancements through various TBVIs to highlight the same by false color RGB display techniques to highlight the lowland wetlands and then use visual interpretation approaches to delineate the wetland boundaries through screen digitization. The most useful enhancement were RGB displays of ETM+ bands: (a) ETM+4/ETM+7, ETM+4/ETM+3, ETM+4/ETM+2, (b) ETM+4, ETM+3, ETM+5, and (c) ETM+3, ETM+2, ETM+1. In addition, the SRTM slope threshold on < 1 percent was very useful in delineating higher-order wetland boundaries. The ground truth data showed that the accuracy of wetland boundary delineation was 96 percent. The boundary delineation for a river basin of 607562 hectares took 2-weeks for one analyst, so it is fairly fast and accurate process. In Ruhuna basin 24 Percent (145733 hectares) of the total basin area was wetlands. Such a high percent was as a result of the high proportion of the human-made irrigated areas, mainly under rice cropping. In comparison, the Limpopo river basin in Southern Africa, which was also studied using similar methodology by this group, had only 12.5 percent of the basin area under wetlands as a result of poor irrigation development in Limpopo. But within the limpopo, the lowland flood plains of Mozambique had 24.7 percent of basin area (8.8 Mha) of Mozambique under wetlands.

The delineated wetlands were classified to determine land use\land cover (LULC) patterns by using a hierarchical classification system that lead to 13, 8, 6, and 4 class maps. For the 4, 6, and 8 classes the accuracies were about 92 percent (Khat=0.92) with low levels of errors of omission (about 5 percent) and errors of commission (1 percent).  Even for 13 classes the accuracy was high (87 percent) and errors of omission reasonable (12 percent) and errors of commission low (1 percent). The wetlands are dominated by irrigated lands (41.5 percent) and riparian vegetation and home gardens (36.4 percent). The other significant LULC classes were: seasonal wetlands (10.4 percent), fresh water bodies (8.8 percent), and lagoon (1.5 percent). Permanent marshes, magroves, and salt pans were the other classes. These classes were mapped with an overall accuracy of 94 percent. Time-series MODIS 500-m monthly NDVI maximum value composite (MVC) images were used to study the characteristics of each of the classes. 

The study clearly implies that automated methods are highly inaccurate to map wetlands. The accuracies may increase using time-series data and/or non optical data. But, large area mapping using such data is resource intensive. The semi-automated methods provide highly accurate results and large area mapping using, mainly freely available data sources, makes the method highly attractive.