Improving Wetland Mapping Techniques Using the Integration of Image Fusion Techniques and Artificial Neural Network (ANN)

Abstract

Wetlands provide a huge number of goods and services, albeit they have been converted into other land use classes and experienced degradation of habitat quality. Consequently, to conserve and restore the wetlands, high-precision wetland mapping and monitoring using remote sensing technology has become essential. Therefore, the present study aims to analyze the performance of image fusion techniques in wetland modeling and towards precise mapping of wetlands by integrating the image fusion techniques and an artificial intelligence model in the Sunamganj district of Bangladesh, also known as a haor basin because of the presence of a number of small and large wetlands (which constitute about 60–70% of the total area of the district), of which the Tanguar haor is the largest wetland. The image fusion was done using multispectral bands (5, 4, and 3) of Landsat 8 Operational Land Imager (OLI) and panchromatic band-8. The fused image was evaluated by the visual interpretation and edge detection techniques. The artificial neural network (ANN) model was employed to classify the wetlands from the high-resolution images generated through image fusion techniques. The wetland maps were validated using the Kappa coefficient and root mean square error (RMSE). All statistical analysis was performed using open source software, such as R studio software (version 3.5.3). Recently, the utilization of open source software has become popular because of its free accessibility and the availability of a huge number of tutorials, which help environmental scientists to work smoothly. The calculated wetland area was maximum in the fused image generated by Gram–Schmidt (2546.42 km2), followed by modified intensity hue saturation (Modified IHS) (1974.74 km2), and multi-spectral image (1945.24 km2), while the wetland area was minimum in the fused image generated by Wavelet IHS (1874.93 km2). The Brovey image fusion techniques yielded high overall kappa accuracy (89.23%) and lower RMSE (0.12), followed by the high pass filter–ANN and modified IHS–ANN. Furthermore, the visual interpretation and edge detection technique showed that the Brovey image fusion model preserved the spectral quality much better than the other image fusion techniques in the wetland modeling in this study.