Prediction of insect-herbivory-damage and insect-type attack in maize plants using hyperspectral data

Abstract

Accurately detecting the insect damage caused in plants might reduce losses in crop yields. Hyperspectral data is a well-accepted data source to attend this issue. However, due to their high dimensional, both robust and intelligent methods are required to extract information from these datasets. Therefore, we explore the processing of hyperspectral data with artificial intelligence methods joined with clustering techniques to detect insect herbivory damage in maize plants. We measured the leaf spectral response from three different groups of maize plants: control (undamaged plants); damaged by Spodoptera frugiperda herbivory, and damaged by Dichelops meiacanthus. Data were collected with a FieldSpec 3.0 Spectroradiometer from 350 to 2500 nm for eight consecutive days. We adjusted eight machine learning methods. We also determined the most contributive wavelengths to differentiate undamaged from damaged plants by insect herbivore attack using clustering strategy. For that, we applied the clusterization method based on a self-organizing map (SOM). The Random Forest (RF) model is the overall best learner, and up to the 5th day of analysis represents the most adequate day to segregate maize undamaged from damaged maize. RF was able to separate the three groups of treatments with an F1-measure of up to 96.7% (Recall of 96.7% and Precision of 96.7%). Additionally, we found out that the most representative spectral regions are located in the near-infrared range. Our approach consists of an original contribution to early differentiate the undamaged plant from the damaged one due to insect-attack, highlighting the most contributive wavelengths to map this occurrence.

Keywords

• Proximal hyperspectral sensing
• Precision agriculture
• Random forest