|BARNES, EDWARD - COTTON INC CARY NC
|LESCH, SCOTT - GEORGE E BROWN JR SAL LAB
Submitted to: Computers and Electronics in Agriculture
Publication Type: Peer Reviewed Journal
Publication Acceptance Date: 6/15/2006
Publication Date: 8/19/2006
Citation: Fitzgerald, G.J., Lesch, S.M., Barnes, E.M., Luckett, W.E. 2006. Directed sampling using remote sensing with a response surface sampling design technique for site-specific agriculture. Computers and Electronics in Agriculture. 53(2):98-112.
Interpretive Summary: One hindrance to the application of remote sensing data to precision farming is the cost of ground sampling needed to develop relationships between the imagery and ground-measured crop attributes. A directed sampling technique presented here utilized pixel value variations and their spatial relationships across images to select small sets of georeferenced locations to sample on the ground. These plant and soil measurements were then used to calibrate the imagery, converting it from brightness values to maps of the attribute of interest, for example plant height. Imagery was acquired on several dates over a two-year period. Either a raw band in the red part of the spectrum or imagery converted to a normalized difference vegetation index (NDVI) was used to direct ground sampling. The red and NDVI are known to relate strongly to measures of plant biomass, height, etc. These images were used as inputs to a software package called, ESAP-RSSD, which selected 12 georeferenced points from each image based on brightness variations and spatial relationships of the pixels. Using a differentially corrected global positioning system (DGPS), cotton height, width, and leaf area index (LAI) were measured at the selected sites. These points were then used to develop regression models, which were validated using a separate set of field-sampled locations. The resulting equations were used to convert the imagery to maps of cotton height, width, and LAI. Thus, by sampling from a small number of locations, plant attributes could be estimated within a large field resulting in a more cost-effective sampling strategy. This could be useful to farmers or consultants for variable rate application of chemicals or to researchers as inputs to crop growth and evaporation models.
Technical Abstract: Remotely sensed imagery can provide contiguous coverage of a field and be used as a surrogate to measure crop and soil attributes. Although an apriori relationship may exist between the imagery and ground attributes, determining the specific parameter estimates to convert imagery to attribute maps potentially requires a large number of expensive ground samples to develop the relationship. A directed sampling approach using the ECe Samping, Assessment, and Prediction - Response Surface Sampling Design (ESAP-RSSD) software was used to develop a ground sampling design from input imagery to predict crop attributes at all non-sampled field locations. It selects a minimum set of calibration samples based on the observed magnitudes and spatial locations of the data. It was originally developed to calibrate apparent soil electrical conductivity data to soil salinity; but the underlying response surface theory is broad enough to allow input of other types of geospatial data, including georeferenced imagery. In a two-year study, imagery was acquired on multiple dates from an aircraft for a cotton field in central Arizona. The ESAP-RSSD software selected 12 georeferenced sampling sites in each image from which crop height, width, and LAI were collected shortly after each flight and used to calibrate the images. Validation and estimation of regression parameters between co-located pixels and ground data showed that this technique can reasonably predict crop attributes across a field from a minimal set of input locations. Crop height was estimated more consistently across dates followed by crop width. LAI was poorly predicted. Maps of this type could be used directly by farmers or consultants in variable rate applicatoins of chemicals or as inputs to crop simulation models providing a spatial extension to their time-series nature.