Effect of spectral band selection and bandwidth on weed detection in agricultural fields using hyperspectral remote sensing

Abstract

Presence of weeds in agricultural fields affects farmers' economic returns by increasing herbicide input. Application of herbicides traditionally consists of uniform application across fields, even though weed locations can be spatially variable within a field. The concept of spot spraying seeks to reduce farmers' costs and chemical inputs to the environment by only applying herbicides to infested areas. Current spot spraying technology relies on broad spectral bands with limited ability to differentiate weed species from crops. Hyperspectral remote sensing (many narrow, contiguous spectral bands) has been shown in previous research to successfully distinguish weeds from other vegetation. Hyperspectral sensor technology, however, might not currently be practical for on-tractor applications. The research objectives were to determine (1) the utility of using a limited number of narrow spectral bands as compared to a full set of hyperspectral bands and (2) the relative accuracy of narrow spectral bands compared to wider spectral bands. Answers to these objectives have the potential for improving on-tractor weed detection sensors. Reference data was provided by field observations of 224 weed infested and 304 uninfested locations within two winter wheat fields in Gallatin County, Montana, USA. Airborne hyperspectral data collected concurrently with the reference data provided 6-nm spectral bands that were used in varying combinations and artificially widened to address the research objectives. Band selection was compared using Euclidean, divergence, transformed divergence, and Jefferies-Matusita signature separability measures. Certain three and four narrow band combinations produced accuracies with no statistical difference from the full set of hyperspectral bands (based on kappa statistic analysis, alpha = 0.05). Bands that were artificially widened to 96 nm also showed no statistically significant difference from the use of 6-nm bands for both all bands and select band combinations. Results indicate the potential for bands that can differentiate weed species from crops and that the narrowest spectral bands available might not be necessary for accurate classification. Further research is needed to determine the robustness of this analysis, including whether a single set of spectral bands can be used effectively across multiple crop/weed systems, or whether band selection is site or system specific.