Pest identification helps researchers improve the quality and yield of agricultural products. Earlier pest identification models are mainly based on machine learning techniques. For example, Le-Qing and Zhen (2012) utilizes local average color features and SVM to diagnose 10 insect pests based on a dataset containing 579 samples. Fina et al. (2013) combined K-mean clustering algorithm with adaptive filter for crop pest identification. Zhang et al. (2013) designed a field pest identification system and their dataset comprises approximately 270 training samples. Ebrahimi et al. (2017) used a differential kernel function SVM method for classification and detection, but the evaluated dataset is small, containing just 100 samples. Wang et al. (2018) uses digital morphological features and K-means to segment pest images.The above traditional pest identification algorithms have been studied with good results, but all of them have limitations, and their detection performance depends on the performance of the pre-designed manual feature extractor and the selected classifier.

Convolutional neural network (CNN) has strong image feature learning capability, such as ResNet (He et al., 2016) and GoogleNet (Szegedy et al., 2015) can learn deep higher-order features from images and can automatically learn shape, color, and texture of complex images and other multi-level features, overcoming the traditional manually designed feature extractors' limitations and subjectivity. It has obvious advantages in target detection, segmentation, classification of complex images, etc.

Liu and Wang (2020) constructed a tomato diseases and pests dataset and improved the YOLOV3 algorithm to detect tomato pests. Wang et al. (2020) introduced an attention mechanism in residual networks for improving the recognition accuracy of small targets. A two-stage aphid detector named Coarse-to-Fine Network (CFN) is proposed by Li et al. (2019) to detect aphids with different distributions. Zhu J. et al. (2021) uses super-resolution image enhancement technology and an improved YOLOv3 algorithm to detect black rot on grape leaves.

In general, CNN-based pest identification work can well avoid the limitations of traditional methods and improve the performance of pest identification. However, most target detection models have applied many hand-crafted components.To some extent, the parameters of the manual components increase the workload. To eliminate the impact of manual components on the model, researchers have considered using the versatile and powerful relational modeling capabilities of the transformer to replace the hand-crafted components. Carion et al. (2020) put forward the end-to-end object detection with transformers (DETR) by combining the convolutional neural network and the transformer, which built the first complete end-to-end target detection model and achieved highly competitive performance.

At present, deep learning-based agricultural pest identification and classification is maturing. The research scope includes a variety of cash crops such as crops, vegetables, and fruits, and relevant datasets have also been constructed.

Wu et al. (2019) constructed the IP102 pest dataset, which covers more than 70,000 images of 102 common crop pests. Wang et al. (2021) constructed the Agripest field pest dataset, which includes more than 49,700 images of pests in 14 categories. Hong et al. (2020) constructed a moth dataset by pheromone traps, which were labeled with four classes, including three moth classes and an unknown class of non-target insects. As a result of data collection and labeling, a total of 1,142 images were obtained. Liu Z. et al. (2016) constructed a rice pest dataset. The data were collected from image search databases of Google, Naver, and FreshEye, including 12 typical species of paddy field pest insects with a total of over 5,000 images. He et al. (2019) designed an oilseed rape pest image database, including a total of 3,022 images with 12 typical oilseed rape pests. Lim et al. (2018) build an insects dataset by specimens and Internet. The dataset consists of about 29,000 image files for 30 classes. Baidu constructed a forestry pest dataset that includes over 2,000 images for 7 classes through the specimen and traps. Chen et al. (2019) build a garden pests datasets. The dataset consists of about 9,070 image files for 38 classes. Liu et al. (2022) constructed a representative dataset of forest pests classification, including 67 categories and 67,953 original images. However, so far, only the dataset of Liu et al. (2022) is available for the detection of forest pests.

In conclusion, the research on crop diseases and insect pests based on deep learning covers a wide range, but in forestry, the detection and control of forest diseases and insect pests is still a challenge.

We collect and annotate the dataset with following four stages: 1) taxonomic system establishment, 2) image collection, 3) preliminary data filtering, 4) Data Augmentation, and 5) professional data annotation.

Our Forestry Pest Dataset contains 7,163 images and 31 pest species. To ensure the training results, we randomly divide according to the following ratio: (Train: Val=9: 1): Test=9: 1. Specifically, the Forestry Pest Dataset is split into 5,801 training, 645 validation, and 717 testing images for the object detection task.

In Table 2, we compare our dataset with some existing datasets related to forestry pest identification tasks. Sun et al. (2018) and Hong et al. (2021) created related datasets using pheromone trap collection, but their datasets only deal with specific species of pests. The forestry pest dataset proposed by Baidu is processed and collected in a controlled laboratory environment. Due to these limitations, these related datasets are difficult to apply to practical applications. Chen et al. (2019) and Liu et al. (2022) focus on the classification of forest pests. Their dataset is rich in pest species and has a sufficient number of samples, which has played a huge role in practical applications. However, they have not made relevant attempts on pest detection tasks, and the relevant datasets have not been published.

Pests with different life cycles have different degrees of damage to forestry, so we retained images of these different morphological pests during data collection and annotation. However, due to the small differences between classes (similar features) and large differences within classes (there are many stages in the life cycle) of pests, accurate classification of their features is a difficult task in detection tasks. In addition, the imbalanced data distribution brings challenges to the feature learning of the model, and the imbalanced data will cause the learning results of the model to be biased toward a relatively large number of classes.

The framework used for this experiment is python3.8, torch1.9, cuda11.1. The experimental hardware is shown in Table 3.

After the accumulation of R-CNN and Fast RCNN, Faster RCNN integrates feature extraction (feature extraction), proposal extraction, bounding box regression (rect refine), and classification into one network in structure, which greatly improves the comprehensive performance., especially in terms of detection speed. SSD is a single-stage target detection algorithm, which uses convolutional neural network for feature extraction, and takes different feature layers for detection output. SSD is a multi-scale detection method. Based on the original YOLO target detection architecture, the YOLOV4 algorithm adopts the best optimization strategy in the CNN field in recent years, and has different degrees of optimization in terms of data processing, backbone network, network training, activation function, loss function, etc., achieving the perfect balance of speed and precision. Based on DETR, Deformable DETR improves the calculation method of the attention mechanism through sparse sampling, reduces the amount of calculation, and greatly reduces the training time of the model while ensuring accuracy.

SSD, Faster RCNN, YOLOV4, and Deformable DETR initial model parameter settings are shown in Tables 4, 5. To take into account the accuracy and training time, in the previous Deformable DETR model training process, the model reached convergence around 150 epoch, therefore, we chose 150 epoch, and Deformable DETR performed a learning rate decay every 40 epoch, so we chose 80 epoch as the intermediate result, Compare the performance of the four models on the dataset. At the same time, to maintain the consistency of the training cycle, we set the same epoch as Deformable DETR for the other three models.

We use mAP and Recall as evalution metrics which are two widely used metrics in target detection. mAP and Recall are calculated as follows:

Where, TP is a positive sample predicted by the model as a positive class, FP is a negative sample predicted as a positive class by the model, FN is a negative class predicted by the model positive sample. Each class can calculate its Precision and Recall, and each class can get a PR curve, and the area under the curve is AP. mAP_α and mAP_{multi−scale} are the average of all classes AP at different confidence levels α and different scales value.

In the MS COCO dataset, objects with an area less than 32*32 are considered small objects, while objects with an area greater than 32*32 and less than 96*96 are considered medium objects.

Average precision performance of object detection methods under different IoU thresholds. The results are shown in Table 6.

From the experimental results in Table 6, it can be seen that the dataset in this paper has good accuracy on mainstream target detection models under short-time training. The recently proposed Deformable DETR can also be used on the dataset in this paper. Achieve roughly the same performance as SSD, Faster RCNN, and YOLOV4. An example of the detection of the model is shown in Figure 4.

From the above results, Deformable DETR based on Transformer architecture does not perform as well as YOLOV4 or even Faster RCNN in some cases. Based on our analysis, there are the following reasons.

The confusion matrix in target detection is very similar to that in classification, but the difference is that the object of the classification task is a picture, while the detection task is different. It includes two tasks of positioning and classification, and the object is each target in the picture. Therefore, to be able to draw positive and negative examples in the confusion matrix, it is necessary to distinguish which results are correct and which are wrong in the detection results. At the same time, the detection of errors also needs to be classified into different error categories. How to judge whether a detection result is correct, the most common way at present is to calculate the IOU of the detection frame and the real frame, and then judge whether the two frames match according to the IOU. For some targets below the threshold or not detected, they will be considered as the background class. The confusion matrix results of the model on the test set are shown in Figure 5.

Small targets have always been a difficult task in target detection due to their small size and lack of feature information. In the field of forest pest detection, the detection of small targets is also a difficult task due to the real complexity. Our dataset contains small objects such as larvae and eggs. We also consider the model's ability to detect small objects in our dataset. The results are shown in Tables 7, 8. The detection example of each model on small targets is shown in Figure 6.

As can be seen from the above table, YOLOV4 significantly leads the rest of the models in the detection of small targets, thanks to its powerful network structure and feature fusion, Deformable DETR is based on the attention mechanism of pixel-level computing, and the detection of small targets is not very friendly.

In this work, we collect a dataset, for forest insect pest recognition, including over 7,100 images of 31 classes. Compared with previous datasets, our dataset focuses on a variety of forestry pests, meets the detection needs of both real and experimental environments, and also includes pest forms in different periods, which some previous forestry pest datasets neglected. Meanwhile, we also evaluate some state-of-the-art recognition methods on our dataset. Exceptionally, this dataset has received good feedback on some mainstream object detection algorithms. However, in the detection of small objects, the existing deep learning methods cannot achieve the desired accuracy. Inspired by the success of the application in computer vision of the Transformer model, we also introduced the Transformer model to solve the forestry pest identification problem. We hope this work will help advance future research on related fundamental issues as well as forestry pests identification tasks.

To better promote the development of forestry pest identification, we will continue to collect forestry pest data and expand the dataset to 99 categories. For pests that have occurred or diseases caused by pests, there is a lack of relevant data sets and research support. In response, we will collect images of diseases caused by insect pests.

Although the existing deep learning models have achieved good results in forest pest identification, small target recognition is still a challenge. We will optimize and improve the model in the follow-up to further improve the model's ability to detect small targets.