Utilizing neural networks for image downscaling and water quality monitoring

Abstract

Remotely sensed images are becoming highly required for various applications, especially those related to natural resource management. The Moderate Resolution Imaging Spectroradiometer (MODIS) data has the advantages of its high spectral and temporal resolutions but remains inadequate in providing the required high spatial resolution. On the other hand, Sentinel-2 is more advantageous in spatial and temporal resolution but lacks a solid historical database. In this study, four MODIS bands in the visible and near-infrared spectral regions of the electromagnetic spectrum and their matching Sentinel-2 bands were used to monitor the turbidity in Lake Nasser, Egypt. The MODIS data were downscaled to Sentinel-2, which enhanced its spatial resolution from 250 and 500m to 10m.Furthermore, it provided a historical database that was used to monitor the changes in lake turbidity. Spatial approach based on neural networks was presented to downscale MODIS bands to the spatial resolution of the Sentinel-2 bands. The correlation coefficient between the predicted and actual images exceeded 0.70 for the four bands. Applying this approach, the downscaled MODIS images were developed and the neural networks were further employed to these images to develop a model for predicting the turbidity in the lake. The correlation coefficient between the predicted and actual measurements reached 0.83. The study suggests neural networks as a comparatively simplified and accurate method for image downscaling compared to other methods. It also demonstrated the possibility of utilizing neural networks to accurately predict lake water quality parameters such as turbidity from remote sensing data compared to statistical methods.