By Arnan Araza (Philippines)
Water from watersheds or catch basins are being regulated by forest ecosystem and forest loss is affecting it negatively. Philippine government started accounting for Ecosystem Services (ES) where water regulation service is often quantified using data-intensive process-based models. The indicator of this ES is streamflow (l/sec) or river flows in watersheds often modeled using a single period land cover/land use input, with series of parameterization, resulting to a time series output.
This study performed an innovative Remote Sensing-dependent technique to predict streamflow at large scales using Random Forest (RF) regression within 6 major river basins in the Philip- pines. It integrated yearly forest loss pixels in predicting daily seasonal streamflow looking at evidences of forest loss effect in the context of water regulation service. A total of 58 physical and climatic covariates were assembled using 4 grouping-learning methods of valuetable, or covariates pool of information, to assess how good RF is in predicting at varying degree of watershed information. The best model was used to predict at full-scale time series from 2000 to 2016 and applied to 6 validation sites without observed data.
The RF models learned better when combination of subwatershed valuetable is applied. The RF model also captured and learned from mixed information. Results revealed which subwa- tersheds have artificial water regulating structures like dams (regulated subwatersheds) which were outperformed by unregulated subwatersheds based on accuracy measures. The latter also showed rainfall-responsiveness or being linearly reactive to rainfall (i.e. high rainfall-high streamflow) in a daily basis. Streamflow predictions were not correlated with the number of training data and subwatershed size. The 6 validation sites, which are all unregulated, showed rainfall-responsiveness and good potential for upscaling.
Forest loss and its associated covariates (FLAC) were valuable explanatory based on permuta- tion and removal measures. More importantly, the RF models were capable of learning forest loss effect to seasonal streamflow based on the following indicators: (1) decrease and increase in dry and wet predicted daily mean streamflow, respectively; (2) linear trends between forest loss and predicted seasonal streamflow according to regression and partial dependence plots; and (3) regulation of wet season streamflow by controlling outliers and reducing over-prediction. Moreover, an increase in forest loss rate resulted into decrease in streamflow during dry season and increase in peak flows during extreme rainy days for unregulated subwatersheds.
Keywords: Remote sensing; forest loss; time series; random forest; stream ow; ecosystem services; watershed