ESTIMATION OF RAINFALL AND TEMPERATURE FROM REMOTE SENSING DATA USING ARTIFICIAL NEURAL NETWORKS

Abstract

Remotely sensed information from satellites, having multi temporal, high spectral and spatial resolutions, can a play a major role in estimating rainfall and temperature in hill regions, flood prone areas and other extreme weather events where ground truth is limited. Such information is of particular importance in areas with sparse hydro-meteorological station networks, from which reliable and daily assessments of rainfall and temperature can be obtained. The present work attempted the potential use of artificial neural networks (ANNs) to estimate the maximum temperature and rainfall at three stations in Uttarakhand using MODIS data. MODIS data was collected from Snow and Avalanche Study Establishment, DRDO, Chandigarh. ERDAS Imagine and ARCGIS softwares were used for image processing and extraction of MODIS selected band values from the selected point locations. U-PROBE data and MODIS data from Oct, 2007 to Feb, 2009 was used in the study. The Observed data was collected from U-Probe reports and from the Department of Hydrology. To test various types of data schemes, the different time series were prepared. A three layer feed forward neural network with incremental back propagation algorithm was used for the selected stations namely DOH-IIT Roorkee, GIC New Tehri and GIC Askot. The use of neural networks offered good to moderate correlation, between observed and calculated maximum temperature values. The correlation coefficients for the estimation of maximum temperature at DOH-IIT Roorkee, GIC Askot and GIC New Tehri are 0.814, 0.627, and 0.807 respectively. Similarly RMSE values of estimated temperature at these stations are 1.902°C, 5.22°C and 2.309°C respectively. The selected scheme and ANN have shown good correlation co-efficient for both DOH-IIT Roorkee and GIC New Tehri stations but not for GIC Askot. The reason for the same is insufficient data for training at GIC Askot. The correlation between observed and calculated rainfall at three stations considered in the study area vary from 0.2 to 0.868 and RMSE values from 5.4 to 15.5 rim. Analyses shows good correlation co-efficient for GIC New Tehri stations and poor results for DOH-IIT Roorkee and GIC Askot. The rainfall that is available from U- PROBE data is cumulative rainfall for the last 24 hours. In order to estimate total rainfall, the best scheme was to consider the previous day MODIS data of either satellites and same day data of Terra satellites. However with this scheme, it became difficult to generate the time series of remote sensing as well as observed data. Therefore, the main reason for unsatisfactory results of rainfall estimation is the lack of desired length of required data. The present study demonstrates that the application of remote sensing data in estimation of meteorological parameters using ANN is possible. However the good length observed data are required for the training of artificial neural networks. iv