应用卷积神经网络的遥感图像云层自主检测

Autonomous cloud detection for remote sensing images using convolutional neural network

摘要:

为实现遥感卫星对遥感图像的自主云层判别能力,提升目标自主识别的效率,避免云层覆盖面积较大的遥感图像丢失关键的目标信息而给后续算法处理带来不必要的计算资源浪费,提出一种基于卷积神经网络的云层自主检测方法,实现遥感图像云层的自主检测,达到了较高的检测精度.首先,根据遥感图像的特性建立卷积神经网络.然后,使用大量人工标识的遥感图像完成云层检测网络训练,使其达到预期检测精度.最后,在卫星在轨运行阶段,将所拍摄的遥感图像根据尺寸划分为若干个子图,并通过训练完成的卷积神经网络对子图是否被云层覆盖进行分类预测.综合所有子图的预测结果给出整幅遥感图像的云层覆盖占比.结果表明:以Landsat卫星遥感图像为测试对象,该方法可以实现有云层覆盖检测正确率为95.3%,无云层覆盖检测精度为97.8%,误判率为2.58%,漏判率为0.90%,综合精度为97.9%;由于使用了卷积神经网络和并行计算技术,该方法基本满足实时性需求,提高了算法的自主性与鲁棒性,为基于遥感图像的在轨实时应用奠定了基础.

Abstract:

Aiming at realizing the autonomous cloud detection of remote sensing satellites for remote sensing images, improving the efficiency of autonomous target recognition, and avoiding the loss of key target information in remote sensing images with large cloud coverage that may bring about unnecessary waste of computing resources to subsequent algorithm processing, a cloud detection method was proposed based on convolutional neural network, which can achieve autonomous cloud detection with high accuracy. Firstly, a convolutional neural network was built based on the requirements of the task and the characteristics of the remote sensing image. Secondly, a large number of manually labeled remote sensing images were used to train the network, which was adjusted according to the training result until it reached a certain classification accuracy for the test image set. Finally, the captured remote sensing image was divided into hundreds of sub-images based on size, and the pre-trained network was used to predict whether the sub-images were covered by clouds. By analyzing the prediction results of all sub-images, the cloud coverage information of the original image was obtained. Taking the Landsat remote sensing images as the test object, results show that the proposed method achieved 95.3% for cloud coverage detection accuracy, 97.8% for no-cloud coverage detection accuracy, misjudgment rate of 2.58%, missing rate of 0.90%, and overall accuracy of 97.9%. With the implementation of a simple structured network and parallel computing technology, the method can be used for real-time space missions with guaranteed autonomy and robustness. The achievement of this work will provide a foundation for real-time orbit applications based on remote sensing images.