Multi-objective Wavelength Selection for Snap-bean Yield Assessment Using Remote Sensing: A Field Study

Abstract:

There is a renewed focus on ensuring a sustainable food production system, given global population growth. This mandates smarter and more efficient ways to implement agricultural management practices. Farmers typically rely on expensive, labor-intensive, and relatively unreliable approaches to assess crop growth trajectories. However, precision agriculture can facilitate this process by offering site-specific solutions that are inherently designed to maximize profit and minimize the use of resources. The objective of this study was to assess yield early in the growing season, based on remote sensing. Our study focused on snap bean, as proxy crop, for a field located in Geneva, NY to model yield of snap bean over two different harvest timings, namely early and late. Six different cultivars, each replicated four times, were studied in a field of size 9×110 meters. The preprocessing was approached via three different vegetation detection algorithms, including the Red Edge Normalized Difference Vegetation Index (RENDVI), Excess Green Index (ExGI), and a spectral library. Multi-objective feature selection, NSGA-II, then was used to minimize the number of identified spectral features, while maximizing yield regression carried out via Multi-Layer Perceptron (MLP). The produced pareto front and identified wavelengths exhibited high coefficients of determination (R²=0.82-0.89) and low root mean squared errors (RMSE=0.49-0.69 tons/acre). The identified wavelengths reside in reflective blue, green, and red, as well as the red-edge spectral regions, with nine wavelengths as an optimal number of features. Our promising results suggest that one could use the down-sampled identified spectral features and design an affordable multispectral sensor, which mounted on an unmanned aerial systems (UAS), could offer a rapid and low-cost alternative to not only the hyperspectral approach, but also classic yield assessment approaches.