• Application of an artificial neural network to estimate groundwater level fluctuation Azhar K. Affandi, Kunio Watanabe and  Haryadi Tirtomihardjo (06-0340)

   ABSTRACT This paper examines and compares the capability of an artificial neural network (ANN) with five different backpropagation (BP) algorithms, namely Gradient descent with momentum (GDM), Gradient descent with adaptive learning rate and momentum (GDX), The Fletcher-Reeves Conjugate gradient (CGF), Quasi-Newton (BGF) and Levenberg-Marquardt (LM), and a radial basis function (RBF) architecture for estimating groundwater level fluctuation (GLF). MATLAB was used to develop the ANN programming. Five-daily measurements of GLF in an observation well provided the data for analyzing the model. An input model using six time lags to estimate actual GLF and 10 hidden nodes gave an optimum result. In general, the work showed that an ANN could be used to estimate GLF even with relatively few data samples. The Levenberg-Marquardt (LM) algorithm was not only the best algorithm in the BP class but also delivered better results than RBF. This result may be very useful in helping developing countries develop groundwater monitoring and management systems. Such countries typically have very few observation wells and lack long-period time-series data due to budget limitations and government policy.

   Keywords: groundwater level fluctuation, estimating, artificial neural network, backpropagation algorithms, radial basis function, MATLAB.

• Spatial Variability and Prediction Modeling of Groundwater Arsenic Distributions in the Shallowest Alluvial Aquifers in Bangladesh M. Shamsudduha 06-0346

Abstract Elevated arsenic in groundwater is the greatest environmental problem in Bangladesh. Spatial variability of arsenic in groundwater has been examined by semivariogram analysis that revealed high degree of small-scale spatial variability in alluvial aquifers. Small-scale variability of arsenic concentrations, indicated by high “nugget” values in semivariograms, is associated with heterogeneity in local-scale geology and geochemical processes. In unsampled locations, arsenic concentrations have been predicted using both deterministic and stochastic prediction methods. Natural neighbor (NN) method predicted better than inverse distance to power (IDP) method, and small-scale variations of arsenic concentrations are preserved. Ordinary kriging (OK) method on the untransformed arsenic data and their residual values performed considerably in predicting spatial arsenic distributions on regional-scale. Predicted results are evaluated by cross-validation, mean prediction error, and root mean square methods. Results show that approximately 25% area of Bangladesh, excluding Chittagong Hill Tracts and southern coastal parts, is below the concentration of 10 µg L^-1 of arsenic. Approximately, 43% area in Bangladesh has arsenic concentrations of 10-50 µg L^-1 at shallow depth (< 25 m). More than 17% area has arsenic concentrations between 50 µg L^-1 and 100 µg L^-1. High density dataset and small-scale modeling would perform better in prediction of spatial distributions of groundwater arsenic. Sequential simulation and co-kriging methods can be applied to evaluate the spatial distributions of arsenic in groundwater in Bangladesh.

Keywords: Arsenic, Bangladesh, distribution, spatial variability, semivariogram, prediction models.

• Evaluation of the Wetland Mapping Methods using Landsat ETM+ and SRTM Data Kulawardhana, R. W., Thenkabail, P. S., Vithanage, J., Biradar, C., Islam Md. A., Gunasinghe, S., Alankara, R. 03-0350

  Abstract Overarching goal of this paper was to evaluate automated and semi-automated methods of mapping wetlands using Landsat ETM+ and SRTM data.

  Automated methods consisted of: (a) slope derived from SRTM, (b) Tasseled cap Wetness Index (TCWI), (c) Normalized Difference Water Index (NDWI), (d) multi-band vegetation indices (MBVIs), (e) two band vegetation indices (TBVIs), (f) normalized difference vegetation index (NDVI), and (g) data fusion involving ETM+ and SRTM and then classifying the same. The best of these indices or methods provide an accuracy of less than 30 percent with high errors of omissions and\or commissions.

  Semi-automated methods consisted of 3 key techniques: (a) image enhancements to highlight wetlands, (b) image display to discern precise boundaries of wetlands, and (b) digitizing directly off screen to separate wetlands from their neighboring landscape. The most useful displays of ETM+ image enhancements (e.g., ratios) and band combinations, displayed as false color composite (FCCs) of RGBs were: (a) NIR/SWIR2, NIR/red, NIR/green; (b) NIR, Red, SWIR1; and (c) red, green, blue. The near-infrared (NIR) is centered at 0.825 μm and the short-wave infrared bands 1 and 2 (SWIR1 and SWIR2) are centered at 1.650 μm and 2.22 μm. The SRTM slope threshold of less than 1 percent was also very useful in delineating higher-order floodplain wetland boundaries.

  The wetlands were delineated with an accuracy of 86.4 percent using the semi-automated methods. The total wetland area in the Limpopo river basin was 12.5 percent of the total basin area of 41.5 million hectares. The overall accuracy of the 4 aggregated wetland classes in the basin was 82 percent with reasonable errors of omissions (20 percent) and low errors of commissions (12 percent).

  Keywords:     wetlands, remote sensing, mapping, delineation, automated methods, semi-automated methods, Limpopo river basin.