Automated spider mite damage detection on tomato leaves in greenhouses

A.T. Nieuwenhuizen*, J. Kool, H.K. Suh, J. Hemming

*Corresponding author for this work

Research output: Contribution to journalArticleAcademicpeer-review

Abstract

Spider mites are one of the most challenging pests in the greenhouse, and their management is more and more challenging due to resistance to pesticides. Spider mites prefer young leaves at the outer canopy, stay at the bottom of leaves, and can crawl to nearby plants. Typical symptoms are small yellow and white spots on the upper side of the leaf due to chlorophyll depletion, leading to bronzing of the leaves. The challenge is to detect the spider mite damage in an early stage, as it is hardly possible to detect the actual spider mites underneath the leaves. Based on a structured requirement list, the following camera systems were identified as detection system candidates. Three multispectral cameras and one RGB camera were further investigated in the research. A lab study with hyperspectral imaging was performed that identified discriminating wavebands between healthy and diseased leaves. The wavebands 700, 716 and 747 nm were selected and used in the multispectral systems. Using Fisher LDA, accuracies up to 0.88 could be reached at pixel level with these multispectral cameras. However, none of the setups could meet the spatial resolution to detect spider mite damage at early stage on the leaves. Second step was to perform color image analysis. An identical accuracy of 0.88 with sufficient spatial resolution was reached. Based on these results, a high resolution RGB camera was mounted on a measurement cart. Experiments were performed in a greenhouse compartment with an ongoing spider mite damage infestation in a tomato crop. The results show the automated pipeline for spider mite damage detection in a time series of four recordings over three months, having an accuracy of 0.90. Heat maps of the infestation rate are presented to the grower and are used for input in models for integrated pest management and to support the decision on the release of beneficial insects.

Original languageEnglish
Pages (from-to)165-172
Number of pages8
JournalActa Horticulturae
Volume1268
DOIs
Publication statusPublished - 1 Jan 2020

    Fingerprint

Keywords

  • Deep learning
  • Greenhouse
  • Machine vision
  • Pest detection
  • Spider mite

Cite this