Click on “Download PDF” for the PDF version or on the title for the HTML version.
If you are not an ASABE member or if your employer has not arranged for access to the full-text, Click here for options.
Automatic pest detection utilizing machine vision and artificial intelligence
Published by the American Society of Agricultural and Biological Engineers, St. Joseph, Michigan www.asabe.org
Citation: 2022 ASABE Annual International Meeting 2200412.(doi:10.13031/aim.202200412)
Authors: Vitor Andrade Gontijo da Cunha, Lucas Costa, Yiannis Ampatzidis, Dylan Pullock, Christopher Weldon, Kerstin Kruger, Aruna Manrakhan
Keywords: Faster R-CNN, Huanglongbing, Machine learning, Object Detection, Trioza erytreae.
Abstract. The South African citrus industry is currently faced with the devastating disease of citrus called Huanglongbing (HLB). Diasphorina citri is the main vector of HBL, but it can also be vectored by an indigenous pest in South Africa, the African citrus triozid (ACT), Trioza erytreae. Yellow sticky traps are being used in citrus orchards in South Africa for continued monitoring of ACT. Identification of ACT on traps requires skilled personnel, which might become problematic when trapping is carried out on a larger scale. Due to this, to render monitoring of this HLB vector more precise and efficient, an automatic pest detection system was developed utilizing artificial intelligence. Three machine learning models were developed using different neural network architectures and the Faster R-CNN algorithm was developed to detect and distinguish the ACP from other species on the traps. The network architectures evaluated were R50 FPN, R101-DC5, and R101 FPN. The precision of each model was estimated based on the F1 score, accuracy, and recall. The machine vision system using the R50 FPN network achieved a precision of 81%, a recall of 81%, and an F1 score of 0.81, while the one using R101-DC5 achieved a precision of 84%, a recall of 69%, and an F1 score of 0.7, and the last one using R101 FPN achieved a precision of 85.7% and a recall of 78%, and an F1 score of 0.82.
(Download PDF) (Export to EndNotes)