Weed and crop discrimination using hyperspectral image data and reduced bandsets

Accurate and efficient weed detection in crop fields is a key requirement for directed herbicide application in real-time Site-Specific Weed Management (SSWM). Using very high spatial resolution (1.25 mm) hyperspectral (HS) image data (61 bands, 400-1000 nm at 10 nm spectral resolution), this study determined that reduced HS bandsets are feasible for discriminating weeds (wild oats, redroot pigweed) from crops (field pea, spring wheat, canola) using Artificial Neural Network (ANN) classification. A 7-band set identified through principal component analysis and stepwise discriminant analysis yielded ANN classification accuracies (88% to 94%) that were nearly equivalent to the full 61-band HS results (89% to 95%) at replicate field plots in southern Alberta, Canada. Therefore, low dimensional narrowband sensors or similar bandsets derived from HS data warrant consideration for SSWM. The computational savings possible from this substantial level of data reduction are potentially critical for enabling optimal use of HS data in real-time ground-based SSWM systems. Recommendations made based on these results have potentially broader implications to SSWM with respect to on-board processing efficiency, weed-crop discrimination method, and sensor and algorithm design.