Detection of Typhoon Damaged Regions on Uav Captured Rice Field Images Using an Ensemble of Cnn and Artificial Neural Network

Date of Award

5-1-2023

Document Type

Thesis

Degree Name

Master of Science in Computer Science

First Advisor

Proceso L. Fernandez, Jr., PhD

Abstract

Typhoons can cause extreme damages to rice fields resulting in rev- enue loss for farmers especially if not addressed early. However, timely damage assessment is difficult due to the scale of rice fields. In this study, we utilized images captured by a commercially available unmanned aerial vehicle (UAV) to create a model that can identify rice plant damage, specif- ically lodging, caused by typhoons. Local officials helped gather and estab- lish ground truth data captured within a period of seven to ten days after a typhoon. The images were divided into uniform-sized tiles and served as the dataset for the baseline CNN model which yielded 87.67% accuracy and 7.28% F1 score. A corresponding 8-feature numeric dataset was de- rived from the kurtosis and skewness of the image histograms in four color channels (red, blue, green and greyscale) produced from each tile provided the dataset for the NN model which yielded 90.33% accuracy and 19.08% F1 score. Experiments with varying parameters ranging from annotation size, image split size, training dataset ratio, and different label threshold were performed to increase the model performance. By assigning weights to both CNN and NN, an ensemble method with improved results was created and post processing methods were also applied. The final ensemble model yielded a 91.81% accuracy and 42.36% F1 score. The improvement from the baseline CNN is shown to be statistically significant, and this shows that we can create a model to distinguish the damage regions on UAV rice field images caused by typhoons.

Link to Full Text

Share

COinS