Use of Earth Observation Data and Artificial Neural Networks for Drought Forecasting: Case Study of Narumoro Sub-Catchment.

Abstract

Droughts are a major problem in Kenya especially in the Arid and Semi-Arid Lands (ASALs) where they are frequent and causes a great deal of suffering and loss. Drought monitoring and forecasting requires extensive climate and meteorological data which is usually largely missing in developing countries or not available in the required spatial and temporal resolutions. Use of the readily available remotely sensed an alternative to observed data for drought monitoring, is faced by many challenges as to the utility and its applicability at the local sub-catchment level. This study examined the use Tropical Rainfall Measuring Mission (TRMM) precipitation estimates in meteorological drought monitoring alongside stream flow modelling for drought identification using Artificial Neural Networks (ANN). Monthly TRMM data was downscaled from its original 0.25° x 0.25° resolution to 1km x 1km resolution using NDVI (Normalized Differential Vegetation Index) from the SPOT VEGETATION program. Using TRMM and observed data meteorological droughts were identified by SPI (Standardized Precipitation Index), and hydrological drought using the threshold method. Both downscaled and original resolution TRMM were found to detect similar meteorological drought pattern and number of drought months as the observed data. TRMM data was found to be suitable for streamflow modeling since the identified hydrological drought pattern was similar to the precipitation pattern. ANN was found to be effective in modelling TRMM-streamflow relationship where the model was able to reproduce a similar flow pattern as the observed streamflow at a 6 month lead time. However ANN model was found to underestimate high flow peaks and overestimate low flows. The best TRMM-streamflow relationship model was found at 1 month lag with correlation coefficient of 0.79 and regression coefficient of 0.87, the least performance was at 6 month lead time where the correlation coefficient was 0.53 and R2 of 0.43. For medium forecasting e.g. 3 month lead time, time lagged TRMM was found to produce better results that using a combination of TRMM and flow. This is based on R2 of 0.72 and 0.54 for TRMM only input and TRMM + flow input respectively. It was concluded that TRMM whether downscaled or at the original resolution can be used to monitor meteorological drought in Narumoro sub-catchment, and that ANN can be used to simulate flow for hydrological drought forecasting using TRMM as input with similar skill to that of ground observed data sets.