**1. Introduction**

Pinewood nematode (*Bursaphelenchus xylophilus*) is a microscopic worm-like creature that causes pine wilt disease (PWD), which poses a serious threat to pine forests, as infected trees die within a few months [**?** ]. Pinewood nematodes can quickly pass from sick to healthy trees via biologic vectors or human activities. The disease is responsible for substantial environmental and economic losses in the pine forests of Europe, the Americas, and Asia [**?** ].

Remote sensing technology is very powerful and widely used to monitor criminal activity or geographical changes, forecast weather, scan using airborne lasers, and plan urban developments. Unmanned Aerial Vehicles (UAVs) are capable of capturing highquality aerial photographs or videos through various high-precision sensors and automated GPS (the global positioning system) navigation. Researchers [**????** ] have recently used UAVs for tree species classification. In early 2005, ref. [**?** ] successfully utilized remote-piloted vehicles to collect viable spores of Gibberella Zeae (anamorph Fusarium graminearum) and evaluate the impact of their transport. Some studies [**???** ] have

**Citation:** You, J.; Zhang, R.; Lee, J. A Deep Learning-Based Generalized System for Detecting Pine Wilt Disease Using RGB-Based UAV Images. *Remote Sens.* **2022**, *14*, 150. https://doi.org/10.3390/rs14010150

Academic Editor: Karem Chokmani

Received: 19 December 2021 Accepted: 23 December 2021 Published: 30 December 2021

**Publisher's Note:** MDPI stays neutral with regard to jurisdictional claims in published maps and institutional affiliations.

**Copyright:** © 2021 by the authors. Licensee MDPI, Basel, Switzerland. This article is an open access article distributed under the terms and conditions of the Creative Commons Attribution (CC BY) license (https:// creativecommons.org/licenses/by/ 4.0/).

5

focused on plant disease identification based on spectral and texture features captured by aerial images.

Despite these achievements, it is still difficult to accurately detect PWD in high accuracy using UAV images for the following reasons: (1) PWD data collection is timeconsuming and costly. Data must be collected from August through September. As PWDinfected trees typically start to die and appear red in late August, it is best to collect images after August, once these symptoms have appeared. However, after October, broad-leaved trees (such as maple trees) change color and show a similar appearance to PWD-infected trees. (2) It is difficult to obtain high-quality orthophotographs, because they require the careful selection of proper settings in terms of image resolution, shooting perspectives, exposure time, and weather. Typically, captured patch images can be overlapped, and orthophotograph can be created based on these overlapped patch images. However, such orthophotographs often suffer from poor image registration due to the elevation differences in forest areas. (3) PWD symptoms vary between stages. In the early stage, PWD-infected tree has a similar appearance to a healthy tree. In the late stage, infected trees show visual symptoms with features that appear close to those of yellow land, bare branches, or maple trees. Color-based algorithms typically show poor detection performance for this issue. (4) Annotation of these data are a challenge and time-consuming task; mis-annotation often occurs due to the poor image resolution and the background similarity.

A common method of locating PWD is based on the handcrafted features of texture and color; specifically, identifying their corresponding relationship to find infected trees [**???** ]. Their results have been based on a limited number of samples (Table **??**), and it is more desirable to analyze PWD using a large number of data samples. Deep learning technology has a powerful ability to process complicated GIS (geographic information system) data. While the encoder of a deep convolutional network automatically extracts inherent feature information from a given input data *X*, the decoder tries to approximate the desired outputs *Y* as closely as possible to solve complex classification and regression problems. Previous studies have used deep learning methods [**?** ] to detect interesting objects. In this paper, we propose a deep learning-based PWD detection system which is verified to be effective and can be generalized for various object detection models using RGB-based images.

To summarize, our major contributions include:


out of 730 PWD-infected trees. The predicted model is made freely available as an open-source module for further field investigation.

**Table 1.** Information about the data sets in various approaches. Symbol "-" denote the missing data in reference paper.

