An automatic method for weed mapping in oat fields based on UAV imagery

• Published in: Computers and electronics in agriculture Vol. 173; p. 105385
• Main Authors: Gašparović, Mateo, Zrinjski, Mladen, Barković, Đuro, Radočaj, Dorijan
• Format: Journal Article
• Language: English
• Published: Amsterdam Elsevier B.V 01.06.2020; Elsevier BV
• Subjects: affordability; Agricultural equipment; agricultural land; Agricultural management; Algorithms; artificial intelligence; Biological properties; Cameras; Classification; Crop yield; data collection; forestry equipment; Herbicides; Imagery classification; Machine learning; Mapping; Oats; pollution; Precision agriculture; soil; Soils; Spatial data; Temporal resolution; UAV; Unmanned aerial vehicles; weed control; Weed mapping; Weeds; Precision agriculture; Imagery classification; UAV; Weed mapping; Oats
• ISSN: 0168-1699; 1872-7107
• DOI: 10.1016/j.compag.2020.105385

The accurate detection and treatment of weeds in agricultural fields is a necessary procedure for managing crop yield and avoiding herbicide pollution. With the emergence of unmanned aerial vehicles (UAV), the ability to acquire spatial data at the desired spatial and temporal resolution became available, and the resulting input data met high standards for weed management. In this paper, we tested four independent classification algorithms for the creation of weed maps, combining automatic and manual methods, as well as object-based and pixel-based classification approaches, which were used separately on two subsets. Input UAV data were collected using a low-cost RGB camera due to its affordability compared to multispectral cameras. Classification algorithms were based on the random forest machine learning algorithm for weed and bare soil extraction, following an unsupervised classification with the K-means algorithm for further estimation of weeds and bare soil presence in non-weed and non-soil areas. Of the four classification algorithms tested, the automatic object-based classification method achieved the highest classification accuracy, resulting in an overall accuracy of 89.0% for subset A and 87.1% for subset B. Automatic classification methods were robustly developed, using at least 0.25% of the scene size as the training data set in all circumstances anticipated for the random forest classification algorithm to operate. The use of the algorithm resulted in weed maps consisting of zoned classes and covering areas with similar biological properties, making them ready for use as inputs in weed treatments that use agricultural machinery.