Effect of N Availability on Vegetative Index of Cotton Canopy: A Spatial Regression Approach

Abstract

Remote sensing is known to indicate nitrogen stress in field crops such as cotton (Gossypium hirsutum L.). Multi-temporal images can indicate many yield-limiting factors in a production system as a change in canopy reflectance. The objective of this research was to test various spatial models to analyze the variations in GNDVI, a vegetative index derived from aerial remote sensing data in the visible and near-infrared (VNIR) region, in response to nitrogen treatments and petiole nitrate content. Four experiments were conducted in two irrigated cotton fields with five N rates (0, 34, 67, 101, and 134 kg N ha-1) during 2003-2005. Six regression models tested included an ordinary least square (OLS) model developed with N treatments as the fixed effect and blocks within a location as the random effect, and five spatial models. The five spatial models were developed by incorporating the spatial correlation structure in the regression model through various covariance models. The five covariance models used in this study were a spherical variogram model fitted to the mixed OLS model residual, a power (POW) or autoregressive model, a POW model with heterogeneous variance at different fields, an anisotropic power model, and an anisotropic power model with heterogeneous variance for fields. The spatial model with anisotropic power covariance structure was the most accurate model because it accounted for the difference in covariance structure along and across rows. With this model structure, 92% of the variation in GNDVI was explained by N rates and days after planting (DAP). Only 30% of the variability in GNDVI could be explained by petiole nitrates and DAP together. In general, all spatial models performed better than the OLS model, indicating that auto-correlated data tend to inflate R2 values. Therefore, spatial regression models are better suited to analyze autocorrelated data.