A practical guide to UAV-based weed identification in soybean: Comparing RGB and multispectral sensor performance

Abstract

Precision agriculture relies heavily on accurate, efficient, and economical methods to distinguish between crops and weeds of various types. The advancement of unmanned aerial vehicle (UAV) technologies provides practical approaches for generating land-cover maps that are essential for monitoring and managing crop fields affected by various weeds. Although the overall cost of scouting crop fields with UAVs may be low and practical, it varies depending on the sensors used; and the existing studies have mainly focused on weed detection methods but not compared the sensors' performance. Therefore, to address this knowledge gap, this research aims to compare a UAV-mounted visual Red-Green-Blue (RGB) sensor and a multispectral sensor in differentiating between crops and weeds in soybean fields, with a particular focus on broadleaf and grass weeds. In this research, a field study was conducted using a support vector machine classification algorithm and object-based image analysis in ArcGIS Pro to examine the impact of sensor choice on weed type differentiation. The analysis with ground truths highlights nuanced discrepancies between the sensors, namely (i) DJI Phantom 4 Pro (RGBd), and (ii) DJI Phantom 4 Multispectral. Overall, with the RGB sensor, an accuracy of 93.8 % was achieved in identifying the land cover types, and the multispectral sensor also had an accuracy of 93.4 % in discriminating the various land cover types. These results show that both sensor's performances were similar, but the less expensive RGB sensor may be better suited precision agriculture at all scales.