Relating crop yield to topographic attributes using Spatial Analysis Neural Networks and regression

Land-surface topographic attributes can be useful for estimating stable spatial patterns of crop yield caused by spatial variability in soils and water availability. We present spatial analyses of grain yield for three fields of dryland winter wheat in northeastern Colorado using topographic attributes as explanatory variables. Topographic attributes including elevation, slope, aspect, curvature, specific contributing area, and wetness index were computed from a 10-m digital elevation model. A Spatial Analysis Neural Network (SANN) algorithm was used for joint spatial interpolation and yield prediction from the topographic attributes. SANN prediction errors were compared with the results of Multiple Linear Regression (MLR). SANN and MLR were assessed in terms of bias and relative root mean squared error (rRMSE) using validation data. SANN out-performed MLR in multivariate estimation, but not for the univariate cases. The greatest advantage of SANN was seen using four or more topographic attributes, whereas MLR showed diminishing efficiency with more than three explanatory variables. Prediction/interpolation errors within a given field were reduced substantially by using the spatial coordinates (latitude and longitude) in tandem with topographic attributes. The rRMSE value reached a minimum of 0.44 (model efficiency, E = 0.80) for interpolation with SANN on the West field. Using only topographic attributes as input, the minimum rRMSE values were 0.59 (E = 0.65) for SANN with 5 variables and 0.72 (E = 0.48) for MLR with 4 or 5 explanatory variables. Thus, this study demonstrated the utility of SANN with topographic attributes that contain implicit soil and water information for estimating spatial patterns of dryland crop yield.