EVAPOTRANSPIRATION ESTIMATED IN COLOMBIA USING NDVI DATA AND NEURAL NETWORKS

In this work are estimated the fields of real, and monthly evapotranspiration for Colombia from the precipitation reanalysis, NDVI fields extracted from satellite images and average monthly evapotranspiration data extracted from national hydrometeorological network. Artificial intelligence technique known as artificial neural networks for estimating the spatial and temporal evapotranspiration distribution over Colombian territory is used for the period 1981-2000. The methodology consists in the calibration of a neural network with sigmoid functions, which allows the nonlinear interaction between input and output variables. The input variables involved are on one side Normalized Difference Vegetation Index (NDVI) obtained from the Advanced Very High Resolution Radiometer (AVHRR) of the National Oceanic and Atmospheric Administration of the United States (NOAA) with a resolution of 8km; this variable has been shown to have a relatively high correlation with evapotranspiration in agricultural crops and natural ecosystems (r(2) = 0.81). The correlation between the evapotranspiration from the neural network and the real evapotranspiration from the evaporation tank, converted through Budyko equation, was r = 0.81. An estimate of the monthly evapotranspiration is then obtained for a period of about 19 years, with a spatial resolution of 9.3 km. The results correspond with the expectations, and the regions of the Amazon and Choco jungle, have the highest real evapotranspiration, while the regions of the Guajira and the highest peaks of the mountain range have the lowest evapotranspiration present, due to the low rainfall in the Guajira's region, and the low temperature in the peaks of the Andes Mountain Range.