Detection the Bi-directional Flow of the Tonle Sap River Using Multi-Sensor Imagery and Deep Learning Model

Abstract

The TSR of Cambodia has bi-directional flows or two reversal flow directions seasonally, which has great impact on the regional environment. It flows upward and downward during wet and dry season respectively. This issue has not been addressed before and the literature has also been silent on this phenomenon. Many researches focused on modeling, monitoring, and forecasting river flow velocity and water discharge of one-directional flow and utilized at the field. The methods used produced the best outcomes for those geographical study area conditions. However, those may not work well for the disorder flow condition like the TSR. Therefore, the author considered a novel approach to detect and monitor this bi-directional flow by utilizing remotely sensed data, S-1 SAR and S-2 optical data. The specific aim of this study is to detect and monitor the impact after the start date of the TSR’s inflow-outflow current seasonally for the period of May – November/2019. The author combined the remote sensing data with the ground data to analysis the regional environment. This new approach is simple but crucial for further research to monitor, forecast, and assess the impacts of the TSR’s flood pulse on LUCC in surrounding area which is driven by the mainstream MR, since it can replace the traditional methods which are limited due to insufficient number of monitoring stations.

The method used the combining outputs from weekly water extent variation extracted from time series SAR data and suspended sediment change direction from the optical data. Thus, the approximately date of each flow direction is determined. Firstly, the VH and VV polarization of S-1 time series images are compared and selected the one which is more sensitive with the water surface in order to extract the water extent variety within the wet season from May to November, 2019. VH polarization was selected. Median value of water surface in April, 2019, is also computed to perform as a reference water data since the river water is assumed to be stable in the dry season. Secondly, the differences between the wet season images and referenced median value are calculated in weekly interval. The outputs were applied terrain filter using SRTM-DEM of 30 m with slope angle of 3°. Then, water features were extracted with threshold of 1.25 and applied the global threshold filter to get a better weekly flood water extent from 01 May to 30 November, 2019. In order to be qualified, the flood extent map 2019 from NFFC cooperating with WFP was used to compare for assessing accuracy. Additionally, the observed water level from gauged stations fused with DEM data was used also to compare on the possible same date. To analyze the impact of flooding on LULC, three classifiers, RF, SVM, and DL, were used to compare and selected the best one for further process. DL classifier was chosen and used to overlap with flood extend in order to figure out how flooded LULC. The influent areas by flood water were computed based on ModelBuilder tool in ArcMap software. The pre-flood water in April was extracted from the LULC map production and the post-flood water in October was extracted from the SAR map production. Both data were computed by using Erase tool (Post-flood - Pre-flood) to get the flood water and then calculate it with LULC data by using Extract by Mask tool to get Flooded LULC map. Finally, Zonal statistics tool was applied to the Flooded LULC by overlapping with District boundary map data to achieve the last output is a District-wise zonal statistics tabular data.

The results showed that the TSR’s water started flowing upward to the TSL on the 3rd week (17th) of May, 2019 until October and keep the water level stable for 2 days possibly on 3rd and 4th October. Then, on 5th of October the river started flowing backward reversely to the main stream. The accuracy assessment results higher than 95%. Therefore, the proposed methods are performing well with this disorder flow condition. LULC classification with CNN-based DL shows the highest overall accuracy of 94.9% with Kappa coefficient of 0.93 followed by SVM, and RF overall accuracy of 91.8% and 85.4%, with Kappa coefficient 0.90 and 0.82, respectively. On the other hand, there was flooding in those LULC, but fortunately in 2019 there was not significantly damaged due to long period of dry season prolonged by La Niña event in 2019, and most of the flood water went into the wetlands, 19.4 km2 (48.8%) and shrublands 7.6 km2 (25%) but not much in agricultural land, 1.3 km2 (4.2%) only. Therefore, it’s assumed that the floods did not cause any harm. Compared to Flood Report in Cambodia in 2019, there was parallel result. In the future, the author will continue to monitor and compare the effects of this reversal flow on the LULC with time series remote sensing images.

-