Development of Generalized Higher-Order Synaptic Neural–Based ETo Models for Different Agroecological Regions in India

This paper aims at developing generalized higher-order synaptic neural (GHSN), i.e., generalized quadratic synaptic neural (GQSN) and generalized cubic synaptic neural (GCSN), reference evapotranspiration (ETo) models corresponding to various methods. The GHSN models (GHSNs) were developed using pooled climate data of different locations under four agroecological regions (semiarid, arid, subhumid, and humid) in India. The inputs for the development of GHSNs include daily climate data of minimum and maximum air temperatures, minimum and maximum relative humidity, wind speed, solar radiation, and pan evaporation with different combinations, and the target consists of ETo estimated by one of the methods. Comparisons of developed GHSNs with the generalized first-order neural network, i.e., generalized linear synaptic neural (GLSN) models, were made to test the relative merits of one model over the other. Comparisons were also made between GHSNs and conventional methods. Based on the comparisons, it is concluded that the GHSNs along with GLSN models performed better than their conventional methods. Comparison of GHSNs and GLSN models among themselves reveals that the GQSN followed by GCSN models performed superior to the GLSN for almost all regions. The GHSNs corresponding to one of the methods ranked first and GHSNs corresponding to another method ranked second for all regions. For semiarid and arid regions, the GHSNs corresponding to one of the methods ranked third and for subhumid and humid regions the GHSNs corresponding to another method ranked third. Further, GHSNs were applied to model development and model testing locations to test the generalizing capability. The testing results suggest that the GQSN followed by GLSN models have a better generalizing capability than GCSN for almost all regions. (C) 2014 American Society of Civil Engineers.