Hyperspectral remote sensing to detect leafminer-induced stress in bok choy and spinach according to fertilizer regime and timing

Abstract

Background: Detection and diagnosis of emerging arthropod outbreaks in horticultural glasshouse crops, such as bok choy and spinach, is both important and challenging. A major challenge is to accurately detect and diagnose arthropod outbreaks in growing crops (changes in canopy size, structure, and composition), and when crops are grown under three fertilization regimes. Day-time remote sensing inside glasshouses is highly sensitive to inconsistent lighting, spectral scattering, and shadows casted by glasshouse structures. To avoid these issues, a unique feature of this study was that hyperspectral remote sensing data were acquired after sunset with an active light source. As part of this study, we describe a comprehensive approach to performance assessment of classification algorithms based on hyperspectral remote sensing data.

Results: Based on average hyperspectral remote sensing profiles from individual crop plants, none of the 31 individual spectral bands showed consistent significant response to leafminer infestation and non-significant response to fertilizer regime. Multi-band classification algorithms were subjected to a comprehensive performance assessment to quantify risks of model over-fitting and low repeatability of classification algorithms. The performance assessment of classification algorithms addresses the important 'bias-variance trade-off'. Truly independent validation (training and validation data sets being separated over time) revealed that leafminer infestation could be detected with >99% accuracy in both bok choy and spinach.

Conclusion: We conclude that detailed hyperspectral profiles (not single spectral bands) can accurately detect and diagnose leafminer infestation over time and across fertilizer regimes. Hyperspectral remote sensing data acquisition at night with an active light source has the potential to enable arthropod infestations in glasshouse-grown crops, such as, bok choy and spinach. In addition, we conclude that effective use and deployment of hyperspectral remote sensing requires thorough performance assessments of classification algorithms, and we propose an analytical performance method to address this important matter. © 2020 The Authors. Pest Management Science published by John Wiley & Sons Ltd on behalf of Society of Chemical Industry.

Keywords: IPM; bok choy; machine vision; reflectance profiling; spinach.

Figure 1

Damage caused by leafminers during different periods of experimental infestation. Leaf damage caused by egg‐laying and feeding leafminer adults during early infestation period in (a) bok choy and (b) spinach. Leaf damage caused by mining activities of newly emerged larvae during late infestation period in (c) bok choy and (d) spinach.

Figure 2

Night‐based hyperspectral remote sensing system. 1, Full‐spectrum light bulbs (PLT‐11088 250 W – PAR38 – Flood – Halogen – 3750 Lumens – 120 V); 2, infra‐red light bulbs (PAR‐138628 90 W – PAR38 – Narrow Flood – IR Halogen – 1550 Lumens); 3, BaySpec hyperspectral camera; 4, remotely controlled DJI Ronin‐M gimbal; 5, portable battery to power remotely controlled motor; 6, remotely controlled motor mounted with portable computer hard drive to store data and connect hyperspectral camera with remote control laptop.

Figure 3

Analyses of variance of average first derivatives in individual spectral bands. Only statistically significant (P < 0.05) effects of leafminer infestation (squares) and fertilizer regime (filled dots) of bok choy (a) and spinach (b) are presented. Separate analyses of variance were performed for three time periods [baseline (before infestation) are presented as red symbols, ‘early’ (2–4 days of infestation) are presented as blue symbols, and ‘late’ (7–10 days of infestation) are presented as green symbols]. In (a) spectral bands 862 and 924 nm are highlighted, because they showed significant responses to leafminer infestation without showing significant response to fertilizer regime.

Figure 4

Classification accuracies (minimum, maximum, and average) of leafminer infestation. Entire data sets acquired from bok choy (a) and spinach (b) were portioned into training and validation data with training data accounting for 50 to 95% of the data. For each training: validation data ratio, 1000 randomized partitions were performed, and minimum, maximum, and average classification accuracies were determined.