A Rotating Object Detection Method for Greenhouses in Remote Sensing Images

Object detection in remote sensing images faces numerous challenges, to address the unique characteristics of greenhouses, this paper designs a rotational object detection model aimed at improving the accuracy and robustness of greenhouse extraction in remote sensing images. The characteristics of greenhouses are summarized as follows: (1) significant scale differences between targets; (2) extreme aspect ratios of targets; (3) dense distribution of targets with orientation information; (4) lack of corresponding public datasets. The proposed method is based on the rotational object detection approach Oriented RCNN and includes the use of PKINet as the backbone network, which is equipped with multi-scale convolution kernels. Additionally, a novel aspect ratio weighted strategy in the loss function is introduced, giving greater weight to targets with larger aspect ratios when calculating the loss. Furthermore, a greenhouse dataset containing 10,931 greenhouse instances was constructed. Comprehensive experiments on the custom dataset demonstrate that the proposed method outperforms current state-of-the-art methods.