TY - GEN
T1 - Automatic detection of thistle-weeds in cereal crops from aerial RGB images
AU - Franco, Camilo
AU - Guada, Carely
AU - Rodríguez, J. Tinguaro
AU - Nielsen, Jon
AU - Rasmussen, Jesper
AU - Gómez, Daniel
AU - Montero, Javier
PY - 2018
Y1 - 2018
N2 - Capturing aerial images by Unmanned Aerial Vehicles (UAV) allows gathering a general view of an agricultural site together with a detailed exploration of its relevant aspects for operational actions. Here we explore the challenging task of detecting cirsium arvense, a thistle-weed species, from aerial images of barley-cereal crops taken from 50 m above the ground, with the purpose of applying herbicide for site-specific weed treatment. The methods for automatic detection are based on object-based annotations, pointing out the RGB attributes of the Weed or Cereal classes for an entire group of pixels, referring to a crop area which will have to be treated if it is classified as being of the Weed class. In this way, an annotation belongs to the Weed class if more than half of its area is known to be covered by thistle weeds. Hence, based on object and pixel-level analysis, we compare the use of k-Nearest Neighbours (k-NN) and (feed-forward, one-hidden layer) neural networks, obtaining the best results for weed detection based on pixel-level analysis, based on a soft measure given by the proportion of predicted weed pixels per object, with a global accuracy of over 98%.
AB - Capturing aerial images by Unmanned Aerial Vehicles (UAV) allows gathering a general view of an agricultural site together with a detailed exploration of its relevant aspects for operational actions. Here we explore the challenging task of detecting cirsium arvense, a thistle-weed species, from aerial images of barley-cereal crops taken from 50 m above the ground, with the purpose of applying herbicide for site-specific weed treatment. The methods for automatic detection are based on object-based annotations, pointing out the RGB attributes of the Weed or Cereal classes for an entire group of pixels, referring to a crop area which will have to be treated if it is classified as being of the Weed class. In this way, an annotation belongs to the Weed class if more than half of its area is known to be covered by thistle weeds. Hence, based on object and pixel-level analysis, we compare the use of k-Nearest Neighbours (k-NN) and (feed-forward, one-hidden layer) neural networks, obtaining the best results for weed detection based on pixel-level analysis, based on a soft measure given by the proportion of predicted weed pixels per object, with a global accuracy of over 98%.
KW - Image analysis
KW - k-Nearest neighbours
KW - Neural networks
KW - Precision agriculture
KW - Soft measures
KW - Weed detection
U2 - 10.1007/978-3-319-91479-4_37
DO - 10.1007/978-3-319-91479-4_37
M3 - Article in proceedings
AN - SCOPUS:85048054602
SN - 9783319914787
T3 - Communications in Computer and Information Science
SP - 441
EP - 452
BT - Information Processing and Management of Uncertainty in Knowledge-Based Systems. Applications - 17th International Conference, IPMU 2018, Proceedings
PB - Springer
T2 - 17th International Conference on Information Processing and Management of Uncertainty in Knowledge-Based Systems, IPMU 2018
Y2 - 11 June 2018 through 15 June 2018
ER -