多尺度轻量化的光学遥感影像云检测方法

Multi-scale lightweight cloud detection method for optical remote sensing images

针对目前对光学遥感图像云检测精度低、误判和漏判高的问题,提出了一种多尺度轻量化卷积神经网络,通过本文提出的轻量化LM-DeeplabV3+框架,并结合资源三号卫星影像数据,以更大的感受野和多个空间尺度来提取特征,从而实现云检测任务。提出了一种改进的轻量化L-MobileNetV2卷积神经网络作为主干模型。与传统的Xception模型相比,这种轻量化的方法大大减少了参数量,从而提高了模型的训练速度和推理时间。为了弥补主干模型轻量化后捕获图像特征的细节方面的不足,在主干网络后引入CA坐标注意力模块,这个模块能够更好地关注特征图中不同空间位置之间的关系,从而更深入地理解图像中云的相对位置,有助于提高模型对图像特征的理解能力,进而提高云检测的精度。同时还提出了一种PPM-ASPP模块,该模块能够在多个不同尺度上提取特征,更好地适应不同类型和不同大小的云。实验效果评价表明:本文所用的遥感影像云检测方法训练时间少、误判低、云检测精度高,适用于高分辨率的光学遥感影像云检测任务。

In order to solve the current problem of low accuracy,high misjudgment,and omission of cloud detection in optical remote sensing images.In this paper,a multi-scale lightweight convolutional neural network is proposed to realize the cloud detection task by adopting the lightweight LM-DeeplabV3+framework proposed in this paper and combining with the ZY-3 satellite image data to extract features with a larger sensing field and multiple spatial scales.Firstly,this paper presents an improved lightweight L-MobileNetV2 convolutional neural network as the backbone model.Compared with the traditional Xception model,this lightweight approach greatly reduces the number of parameters,which improves the training speed and inference time of the model.Second,to make up for the deficiency in capturing the details of image features after the lightweight of the backbone model,a CA(Coordinate Attention)coordinate attention module is introduced after the backbone network,which can pay better attention to the relationship between different spatial locations in the feature map,thus providing a more in-depth understanding of the relative positions of the clouds in the image,and helping to improve the model's understanding of the image features,which helps to improve the model's ability to understand the image features,and thus improve the accuracy of cloud detection.Meanwhile,this paper also proposes a PPM-ASPP module,which can extract features at multiple different scales to better adapt to different types and sizes of clouds.The evaluation of the experimental results shows that the remote sensing image cloud detection method used in this paper has less training time,low misjudgment,high cloud detection accuracy,and is suitable for high-resolution optical remote sensing image cloud detection tasks.