Objective To enhance the detection accuracy of infested individual trees in Unmanned Aerial Vehicle (UAV) imagery, this study proposes an improved detection method termed ECenterNet, built upon the foundational CenterNet architecture.

Method (1) According to the characteristics that most boundary boxes of the infested trees in UAV forest images are rectangular rather than square, a 2D elliptical Gaussian rather than the 2D circular Gaussian is used to generate the penalty area for pixels near the center of infested trees because 2D elliptical Gaussian can better adapt to the shape of the infested trees’s boundary box. (2) A ResNet101 pre-training model is trained on a large number of unlabeled UAV forest image data sets by a self-supervised learning method. As a pre-training model of the infested trees detection network, this model provides rich prior information for the detection model, which can at last improve the accuracy of the infested trees detection. (3) Besides the existing loss functions of the original CenterNet method, a contrastive learning loss function is also designed based on the idea of a self-supervised learning method. This function can make the infested trees characteristics extracted from the model gather within the class, and separate between classes as much as possible, which improves the classification accuracy of the model. (4) In the branch of the category layer in the network, the central difference convolution is used to replace the original ordinary convolution, so that the category branch layer of infested trees detection model can not only extract the semantic information of the image but also extract the difference information, which can provide more information for the classification of models.

Result After the above improvements, the mAP@0.5, 0.95 accuracy of the proposed model is improved from 0.498 to 0.543, with a total increase of 4.5% when compared to the original CenterNet model, while the inference time only increase a little.

Conclusion The improved method can effectively improve the detection accuracy of the model and provide technical support for accurate monitoring and control of forest pests and diseases.