On-the-go forecasting of crop reflectance indices for controlling aerial prescription applications

Abstract

Whole-paddock, fertiliser prescription maps derived from high spatial-frequency variations in canopy reflectance data requires spatial statistical procedures such as kriging which, in turn requires the complete 2-0 dataset. Recent successful trials of active crop reflectance sensors in low-level aircraft provide a realistic opportunity for real time sensing and control of fertiliser application rates at a single pass. However in-flight sensing and actuation requires a predictive map of crop reflectance variability, preferentially in zones, that can be created from a 1-dimensional data stream. Forecasting the required prescription zones ahead of the aircraft avoids actuation delays and mechanical loading on components associated with responding to high spatial frequency noise. A dynamic aerial survey algorithm (OAS) has been devised that utilises each transect flown to create a full-field predictive map. OAS uses a radial, basis function, kernel - support, vector machine regressor, which progressively updates its field estimate using the cumulative data from all previous transects. A fixed-cut, contouring algorithm then segments this current prediction into a predefined number of zones (prescriptions). A recent field trial comparing a full field NDVI map with the successive prediction maps from an aircraft demonstrates the error in the forecast map reduces linearly with increasing transect number.