USING MULTISPECTRAL DRONE IMAGING AND MACHINE LEARNING TO MONITOR SOYBEAN CYST NEMATODES

Author

Joseph Moses Kalinzi, Southern Illinois University CarbondaleFollow

Date of Award

8-1-2023

Degree Name

Master of Science

Department

Geography and Environmental Resources

First Advisor

Li, Ruopu

Abstract

Soybean Cyst Nematode (SCN) poses a significant threat to soybean production in North America and the world at large. Early and accurate detection of SCN infestations is crucial for implementing effective management strategies and minimizing yield losses. The conventional method of SCN detection involves uprooting plants to examine the roots and collecting soil samples. Drone-based multispectral imaging has been used as a viable alternative for crop monitoring due to its detailed spatial and spectral information and scheduling flexibility. This thesis aims to examine the potential of using multispectral drone images for SCN detection in a soybean production field and develop a non-destructive approach to support improved precision agricultural management practices. Using the DJI Matrice 210 drone and a MicaSense Altum sensor, at a height of 50 meters above ground level and top speed of 6 meters/second, a total of 2,550 multispectral images per flight were collected for a total of fourteen flights beginning in June 2022 up to September 2022 from a production field with variable SCN infestation levels located in Carmi, IL. These images were postprocessed with geometric and radiometric correction to produce orthomosaic photos.   Ten vegetation indices namely, NDRE, NDVI, EVI, GNDVI, BNDVI, SIPI, R-EDGE/G, NIR/G, R-EDGE/R and MSR, were computed for each flight date and study plot. The count of SCN eggs was appended to each study plot to find the correlation between the vegetation indices and the field parameters. The VIs having the highest correlation with the eggs and also having the highest number of correlation coefficients significantly different from zero were NDRE, NDVI and GNDVI. I computed the mean values of these VIs for each study plot and flight date which resulted into a time-series trend analysis. To identify study plots with similar trends, an agglomerative hierarchical clustering was performed which resulted into two clusters for each VI. After conducting the ANOVA test, NDVI returned statistically significant results for all the field parameters, GNDVI returned one while NDRE returned three outcomes that were not statistically significant. The study plots belonging to Cluster 1 had a higher mean of SCN count while those in Cluster 2 portrayed little or no SCN. I found NDVI to be the optimal VI because the results from statistical tests and modeling techniques conducted were significant for all SCN parameters, such as cyst and egg count for the plots clustered based on the NDVI trend. Therefore, I used the plots clustered based on the NDVI trend to train and test six ML classification models (Support Vector Classifier, Naïve Bayes, K-Nearest Neighbors, Linear Discriminant Analysis, MLP-Neural Network and Gradient Boost) such that when presented with information in a format like that used in training, it becomes possible to identify plots with high or no SCN. Gradient Boost, MLP-NN and LDA performed with 89%, 82% and 80% accuracy respectively.