Monitoring the damage of Armyworm in Summer 2 Corn by Unmanned Aerial Vehicle imaging

: Background: Monitoring armyworm ( Mythimna separata Walker) damage in crops requires timely, 12 rapid and accurate observations to avoid severe yield losses. 13 Results: The Random Forest (RF) classifier was more effective at automatically and accurately 14 monitoring armyworm damage compared with Support Vector Machine (SVM), Multilayer 15 Perceptron Classifier (MLPC) and Naive Bayes Classifier (NB) classifiers. Furthermore, the 16 incorporation of an Unmanned Aerial Vehicle (UAV) image-generated digital surface model 17 improved the performance of the RF classifier, increasing the F-score from 0.985 and 0.970 to 0.997 18 and 0.994, and increasing the Kappa coefficient from 0.955 to 0.990. In addition, we found that Band 19 3 (735 nm) of the UAV image and Band 6 (740 nm) of a coincident Sentinel-2 image were not 20 sensitive to an armyworm infestation in this study. 21 Conclusions : We developed an accurate algorithm for the automated identification of 22 armyworm-damaged corn plants using UAV images at the field scale. The study also indicated the 23 feasibility of the developed method for monitoring corn armyworm damage at regional scale when 24 combined with Sentinel-2 images.


101
[36] developed a disease and pest warning system for coffee to predict the incidence of various 102 diseases and insect pests by machine learning algorithm, such as the RF Regressor, Artificial Neural 103 Networks, etc.

104
In this study, we used the RF algorithm to identify the damage of armyworm in a summer corn 105 field by extracting the pest damage from UAV images of the study area. The spectral characteristics 106 of different incidences of the pest in Sentinel-2 images were also analyzed. The contributions of this 107 paper were: i) to evaluate the performance of RF for monitoring the damage of armyworm as a pest 108 of summer corn and compare it with other machine learning classifiers including SVM, Multilayer

124
The UAV images were collected on 18 August 2019 between 10:50 and 11:25 local time, almost 125 simultaneously with the acquisition time of a Sentinel-2 image. The meteorological conditions were 126 sunny and windless. The Parrot Disco-Pro AG fixed-wing UAV system was used to collect the UAV 127 images. The system, developed for agricultural applications, carried an automated multispectral 128 sensor -the Parrot Sequoia camera (Table 1). The camera was connected to an irradiance sensor 129 which had the same spectral bands as the multispectral sensor and recorded the light conditions    Table 2). Therefore, the SupReME algorithm [39] was used to generate a unified

Classification by machine learning algorithms
The damage of corn plants infected by armyworm shows with a decrease in leaves, which leads 168 to changes in physiological activities, morphological characteristics and canopy spectrum of corn.

169
The classifiers used image features to classify individual corn pixels into different categories.

189
The OA (Eq. 1) indicates the proportion of correctly predicted pixels including both healthy and 190 attacked by insect pests. The F-score (Eq. 2) is an index that combines precision (p, Eq. 4) and recall 191 (r, Eq. 5), and it is calculated for either healthy or infested corn. In Eq. 2, β is to control the weights of 192 p and r: when β = 1, the two weights are the same; when β > 1, the weight of p is significant; and 193 when β< 1, the weight of r is significant. In this study, β = 1.

196
The RF classifier can perform an implicit feature selection to process the high-dimensional data

212
Apart from RF, three classifiers were compared using these two types of datasets, including 213 SVM, MLPC and NB classifiers (Figure 4). For these three classifiers, we observed the following 214 three results. First, using the multispectral UAV images alone (Figure 4

243
The percentage of pest pixels in a 10 m × 10 m grid (generated from the Sentinel-2 image) was 244 calculated based on Figure 5 (a), and classified into five pest incidence levels of corn ( Figure 6). The 245 pest incidence is most serious in the southern portion of plot A, the top center of plot B and whole 246 of plots C and D, mostly 60%-100%. This situation is consistent with field observations. In terms of incidence, plots A and B, C and D are adjacent to each other, and the infested areas are connected.

248
Therefore, that could be the cause of the local armyworm outbreak in the cornfield.     Figure 8 shows the UAV image 291 representative average reflectance from the corn area damaged by pests and healthy summer corn 292 canopies. The reflectance of corn damaged by pests is higher than that of healthy corn in Band 1 -

293
Green and Band 2 -Red, and it is lower than that of healthy corn in Band 4 -NIR. However, in Band 294 3 -Red-edge, the reflectance is almost the same in both cases. Figure 9 shows the Sentinel-2 295 representative average reflectance from the armyworm infested (pest incidence level at 80%-100%) 296 and healthy (pest incidence level at 0%-20%) summer corn canopies. The two curves intersect at two

331
Third, as a low-cost, flexible and easy to operate field monitoring method, UAV remote sensing 332 is playing an increasingly important role in agricultural development. Using UAV images, we can 333 quickly understand the growth of crops in the field, and quickly identify the region and level of 334 infestation of armyworm [24]. As a large agricultural country, China represents a broad application 335 prospect for agricultural insurance. However, rapid and accurate claim settlements have always 336 been a problem hindering the development of the agricultural insurance industry. The application of 337 UAV in insect pest identification, damage assessment and yield estimation can provide effective 338 means for the rapid settlement of agricultural insurance claims. In addition, we can combine 339 multi-temporal UAV images and satellite images to dynamically detect insect pests in the field, 340 understand the source and diffusion patterns of pests, and manage the farmland more intelligently.

Conclusions 342
Armyworm is one of the most serious insect pests of corn. In this paper, we proposed a 343 method to successfully monitor the damage of armyworm as a pest in summer corn in the study 344 area based on UAV images and the RF algorithm.

345
The conclusions that can be drawn from this study are as follows: 1) RF could monitor the 346 damage of armyworm (Mythimna separata Walker) infestation on summer corn by UAV images 347 from the heading stage to the milk stage in the study area; 2) the addition of UAV image-generated 348 DSM will not only improve the classification accuracy of RF, but also improve its operation efficiency; 3) the order of importance of five features of UAV for the identification of armyworm is distributions based on multi-source satellite remote sensing imagery. Optik -International Journal for classifying insect defoliation levels. ISPRS Journal of Photogrammetry and Remote Sensing. classification using multispectral data acquired from an unmanned aerial vehicle. International Journal of