Directed sampling using remote sensing with a response surface sampling design for site-specific agriculture

Abstract

Remotely sensed imagery provides contiguous spatial coverage of a field and can be used as a surrogate to measure crop and soil attributes. Empirical regression models are often used to convert imagery to attribute maps, when an a priori linear relationship can be assumed to exist between the imagery and ground attributes. In this study, we used the response surface approach incorporated in the EC e Sampling, Assessment, and Prediction (ESAP) software to create ground sampling designs from input imagery in order to develop regression equations for predicting crop height and width attributes in a 3.4-ha cotton field. We examined both the reliability of this model-based sampling approach as well as the validity of the assumed linear models using multiple-date imagery and sample data collected from a 3-year remote sensing experiment. Predictions of height and width from regressions between the imagery and ground sampling at the calibration locations gave coefficients of determination for height ranging from 0.34 to 0.90 and for width, 0.30 to 0.94. All regression models but one were statistically significant at the α = 0.01 level. To test the reliability of the sampling approach, the regression models developed during the first year were used to predict additional crop height and width attributes at a randomly chosen set of validation sites. Multiple statistical tests indicated that these predictions were both unbiased and within the specified precision of the estimated regression equations. This regression-based directed sampling and estimation method requires fewer points than co-kriging to develop reliable imagery-crop attribute relationships, and thus is potentially less expensive. We hypothesize that other variables such as crop nitrogen might also be accurately predicted using this approach as long as the crop attribute and spectral index meet the model assumptions. Maps of crop attributes and/or soil properties could be used by farmer consultants to schedule variable-rate applications of chemicals or as inputs to crop simulation models providing a spatial extension to their time-series nature.