Automated mapping of early Fall armyworm (FAW) damage in sweet corn and maize using machine vision

Automated mapping of early Fall armyworm (FAW) damage in sweet corn and maize using machine vision

This document outlines UniSQ’s development of automated mapping of early fall armyworm (FAW) damage in sweet corn, maize and sorghum using machine vision. This report details the completion of the project as detailed below.

Activity 1: Refine algorithms to detect early FAW damage (4 April 2022 -1 June 2022)
Activity 2: Finalise algorithms and protocols to detect FAW damage (2 June 2022 – 1 February 2023)
Activity 3: Evaluate scouting methodology for FAW damage detection (2 February 2023 – 31 May 2023)

This project identified machine vision algorithms that could detect FAW damage in sweet corn, maize and sorghum from nine sites over three seasons, focussing on low level damage caused by 1st and 2nd instar larvae of FAW (or 1-3 on the Davis Scale). Models using the normalized green red difference index (NGRDI) from a standard UAV colour camera produced the highest accuracy of 73% of detecting FAW presence in unseen fields and 88% on unseen images from the same fields from 30 m altitude in maize. This was more accurate than NDVI from a multispectral near-infrared camera at both 10 m altitude (accuracy of 66% on unseen images from the same field) and 30 m altitude (accuracy of 65% on unseen images from the same field). Models incorporating leaf damage information did not increase accuracy in unseen fields. NGRDI-driven models also maintained their accuracy on downsampled images collected at the equivalent of 100 m altitude with 71% accuracy which would substantially reduce flight times. Models were developed to identify FAW in sorghum with accuracies in unseen fields of 77%.

FAW patches were also monitored both spatially and temporally at the Gatton research site in 2023 using NGRDI to help understand FAW progression. At this site, FAW moved into the field from the border, regardless of orientation, and moved parallel with the crop rows before spreading perpendicular to the rows and eventually becoming ubiquitous. This suggests that initial FAW scouting and potential control should occur near field borders and subsequently parallel with the crop rows. Given the small size of initial FAW patches, satellite imagery was assessed to be unsuitable for FAW detection because of low spatial resolution (0.8m²/pixel) which would lead to inaccuracies due to pixel mixing.

Files associated with this item cannot be displayed due to copyright restrictions.