This study aims to increase the performance of the U-NET model through training, effective acquisition of deep learning study data, and accurate true values. To this end, the filtering method applying the step-by-step deep learning algorithm was used to minimize the unclear analysis basis of the deep learning model, and efficient analysis judgment was conducted, and the following conclusions were obtained.



A) Performance comparison of U-NET model according to training data
- For the U-NET model using the period analysis extraction, the true value was analyzed to be 7.6% higher than the U-NET using the existing provided true value in detecting and improving the performance, along with the detection of dead trees with pine wilt disease using the U-NET algorithm. The cause of the degradation of U-NET detection performance is determined as follows.
- Since trees are expressed in vertical image shapes in orthoimages, it is difficult to identify conifers and broad-leaved trees, dead trees, autumn leaves according to seasons, and browning of leaves, which can lead to incorrect judgment.
- Therefore, it is judged that using the accurate true value of pine wilt disease, deadwood can increase the detection level, and the performance can be improved through the ''period analysis extraction true value'' data collection method.


B) Performance of U-NET model using period analysis extraction true value
- The precision of the U-NET model was 86.6% and 80.2% as a result of analyzing the U-NET model using the orthoimages provided by the KOFPI(Korea Forestry Promotion Institute) using the test verification image and the orthoimage of the Daegu Arboretum taken for model transference test. yielding more than 80% performance. However, in this study, it is impossible to calculate the TN value, making it difficult to analyze the accuracy.
- To construct a high-precision true value of dead trees, the method of constructing and utilizing location data analyzing dead trees removed by the time of orthoimages is suitable. If the performance is improved through more learning in the future, it is judged that the drone surveillance method using drone orthoimages and deep learning models can be used in the pine wilt disease disaster prevention project.

C) Performance comparison of model transference test and deep learning models
- In this study, the performance of the U-NET deep learning algorithm trained using high-precision true values was evaluated through model transference test.
- Also, the Mask R-CNN model algorithm was trained on the same training data to perform analysis and compare the performance of the two models.
- For model transference test, an orthoimage was produced by shooting an area damaged by pine wilt disease near the Daegu shooting range using a drone, and the recall rate (0.966) as a result of analysis with the U-NET model algorithm using the GeoServer program and precision (0.802) were obtained. In addition, the Mask R-CNN model algorithm obtained the precision of recall (0.864) and precision (0.750), and it was found that the U-NET model algorithm performed better by 5.2%p.
- However, compared to the U-NET algorithm, Mask R-CNN has a feature of detecting the complete shape by judging the shape and color of dead trees, so it is judged that more learning data and training are required.
- Therefore, it is expected that the performance and precision of the U-NET algorithm can be improved beyond that of the U-NET algorithm by verifying the analysis result data through deep learning and using it as learning data.