Object-based image analysis (OBIA) is regarded as an effective technology for high-spatial resolution (HSR) image classification due to its clear and intuitive technical process. However, OBIA depends on the manual tuning of image classification features, which is a tricky job. Deep learning (DL) technology autolearns image features from massive images and obtains higher image classification accuracy than traditional techniques. In this article, a novel method called object-scale adaptive convolutional neural network (OSA-CNN), which combines OBIA with CNN, is proposed for HSR image classification. First, OSA-CNN collects the image patches along the main axes of the object primitives obtained through image segmentation; the size of the former is automatically determined through the axis widths of the latter. This step generates the input units required for the CNN classification. Second, the Squeeze-and-Excitation block is extracted from the SE network into the network structure of GoogleNet, which realizes the weighted fusion of the multiscale convolutional features, enhances useful features, and suppresses useless ones. In the classification stage, multiscale image segmentation and CNN classification are fused using an object-scale adaptive mechanism. Finally, object primitives are classified through majority voting on the image patches. The network structure modifications, multiscale classification fusion, and other improvements are verified by gradually incorporating these steps into the original GoogleNet. The experiments show that these improvements effectively enhance the image classification accuracy. This article presents an effective way of combining OBIA and DL techniques to utilize the advantages of both approaches and facilitate HSR image classification.

中文翻译：

---

用于高空间分辨率遥感图像分类的对象尺度自适应卷积神经网络

基于对象的图像分析（OBIA）因其清晰直观的技术流程而被认为是一种有效的高空间分辨率（HSR）图像分类技术。然而，OBIA 依赖于图像分类特征的手动调整，这是一项棘手的工作。深度学习（DL）技术从海量图像中自动学习图像特征，获得比传统技术更高的图像分类精度。在本文中，提出了一种称为对象尺度自适应卷积神经网络 (OSA-CNN) 的新方法，该方法将 OBIA 与 CNN 相结合，用于 HSR 图像分类。首先，OSA-CNN沿图像分割得到的对象图元的主轴收集图像块；前者的大小通过后者的轴宽自动确定。此步骤生成 CNN 分类所需的输入单元。其次，将SE网络中的Squeeze-and-Excitation block提取到GoogleNet的网络结构中，实现多尺度卷积特征的加权融合，增强有用特征，抑制无用特征。在分类阶段，多尺度图像分割和 CNN 分类使用对象尺度自适应机制进行融合。最后，通过对图像块的多数投票对对象图元进行分类。网络结构修改、多尺度分类融合和其他改进通过逐步将这些步骤合并到原始 GoogleNet 中得到验证。实验表明，这些改进有效地提高了图像分类精度。