Skip to main content
ARS Home » Research » Publications at this Location » Publication #124569
Title: TESTING AND APPLICATION OF SPATIAL ANALYSIS NEURAL NETWORKS: SENSITIVITY TO STRUCTURAL PARAMETERS

Author
item MARTINEZ, ANA - COLORADO STATE UNIVESITY
item SALAS, JOSE - COLORADO STATE UNIVERSITY
item Green, Timothy

Submitted to: Annual Hydrology Days Conference Proceedings
Publication Type: Proceedings
Publication Acceptance Date: 3/2/2001
Publication Date: N/A
Citation: N/A

Interpretive Summary: A Spatial Analysis Neural Network (SANN) algorithm (Shin and Salas, 2000), was developed for the spatial analysis of geophysical data, based on the concepts of traditional Artificial Neural Networks. It consists of four layers in which the neurons or nodes between layers are interconnected successively by feed-forward direction. Each node has a transfer or activation function that only responds (or activates) when the input pattern falls within its receptive field, which is defined by its smoothing parameter or width. The activation widths are functions of the model parameters, including the number of the nearest neighbor points P, and a control factor F. There are two operation modes, namely, a training mode in which the model structure is constructed, and an interpolation mode. In this paper we discuss the effect of varying P and F upon the accuracy of the estimation in a two-dimensional domain for different input field sizes and for both training and interpolation modes. A data set of spatial wheat crop yield and several topographic attributes from eastern Colorado is used for testing and illustration. Crop yields are estimated as functions of the two-dimensional Cartesian coordinates (easting and northing)

Technical Abstract: A Spatial Analysis Neural Network (SANN) algorithm (Shin and Salas, 2000), was developed for the spatial analysis of geophysical data, based on the concepts of traditional Artificial Neural Networks. It consists of four layers in which the neurons or nodes between layers are interconnected successively by feed-forward direction. Each node has a transfer or activation function that only responds (or activates) when the input pattern falls within its receptive field, which is defined by its smoothing parameter or width. The activation widths are functions of the model parameters, including the number of the nearest neighbor points P, and a control factor F. There are two operation modes, namely, a training mode in which the model structure is constructed, and an interpolation mode. In this paper we discuss the effect of varying P and F upon the accuracy of the estimation in a two-dimensional domain for different input field sizes and for both training and interpolation modes. A data set of spatial wheat crop yield and several topographic attributes from eastern Colorado is used for testing and illustration. Crop yields are estimated as functions of the two-dimensional Cartesian coordinates (easting and northing)