Machine learning paradigms for weed mapping via unmanned aerial vehicles

Abstract

This paper presents a novel strategy for weed monitoring, using images taken with unmanned aerial vehicles (UAVs) and concepts of image analysis and machine learning. Weed control in precision agriculture designs site-specific treatments based on the coverage of weeds, where the key is to provide precise weed maps timely. Most traditional remote platforms, e.g. piloted planes or satellites, are, however, not suitable for early weed monitoring, given their low temporal and spatial resolutions, as opposed to he ultra-high spatial resolution of UAVs. The system here proposed makes use of UAV-imagery and is based on: 1) Divide the image, 2) compute and binarise the vegetation indexes, 3) detect crop rows, 4) optimise the parameters and 4) learn a classification model. Since crops are usually organised in rows, the use of a crop row detection algorithm helps to separate properly weed and crop pixels, which is a common handicap given the spectral similitude of both. Several artificial intelligence paradigms are compared in this paper to identify the most suitable strategy for this topic (i.e. unsupervised, supervised and semi-supervised approaches). Our experiments also study the effect of different parameteres: the flight altitude, the sensor and the use of previously trained models at a different height. Our results show that 1) very promising performance can be obtained, even when using very few labelled data and 2) the classification model can be learnt in a subplot of the experimental field at low altitude and then applied to the whole field at a higher height, which simplifies the whole process. These results motivate the use of this strategy to design weed monitoring strategies for early post-emergence weed control.