通过卫星云图人们可以获取不同的云型信息，其中对流云极易产生强对流灾害性天气，影响航空安全飞行，严重时对人们的生活与生命造成不可挽回的后果。目前以人工经验判断为主的手段主观性强且效率低。因此，本文提出了一种改进的SSD目标检测算法，能高效且较准确地检测对流云，对天气预测以及航空安全飞行等有实际意义。

本文以检测速度和精度综合表现较好的SSD模型为基础算法，利用VOC公共数据集对模型的检测能力进行测试，针对其不足提出了三种改进策略，然后把改进的SSD模型应用于卫星云图中对流云的检测，将其命名为cloud_detection模型。

针对SSD模型分层预测时用于预测的特征图没有再利用的问题，本文采用特征融合方法改善网络结构，通过特征连接和对应元素相加两种方式实现conv4_3和fc7层的融合，然后在融合层上生成新的特征金字塔进行多尺度检测,以此增加模型特征层之间的联系，充分利用不同网络层的特征信息，从而提高检测准确率。

针对模型靠前层的小目标特征明显但语义信息较少的问题，本文在上一改进后加入特征增强模块，该模块在参考多分支卷积和残差网络思想的基础上，引入了扩张卷积层来拓展靠前层特征的感受野，使得网络学习到的特征信息更丰富，提高检测精度的同时降低小目标的漏检。

针对SSD网络中的在线困难样本挖掘机制剔除了易分样本，使得易分样本无法进一步提升训练效果的问题，本文去除原来的在线困难样本挖掘机制，利用焦点损失函数来平衡正负样本数量，增强模型背景分辨能力的同时降低漏检率。

由于没有公开的卫星云图数据集，本文选取风云二号卫星红外云图，自建数据集记为fy_cloud，对其进行编号、标记、分组等处理后训练并测试模型。实验结果表明，cloud_detection模型在fy_cloud数据集中的准确率相比SSD模型提高了16.02%。通过各个时段的卫星云图识别结果来看，对流云整体检测结果较好，漏检率和误检率均较低。

Through satellite cloud images, people can obtain different cloud information, among which convective clouds are easy to produce strong convective disastrous weather, affect aviation safety flight, and cause irreparable consequences to people's production and life when serious.At present, the method of artificial experience judgment is subjective and inefficient.Therefore, this paper proposes an improved SSD target detection algorithm that can efficiently and accurately detect convective clouds, which is of practical significance for weather prediction and aviation safety flight.

This paper based on the SSD model with good detection speed and precision, first uses the VOC common data set to test the detection ability of the model, in view of its shortcomings, three improvement strategies are proposed. Then the improved SSD model is applied to the detection of convective clouds in satellite cloud images and named as cloud_detection model.

To solve the problem that the feature map used for prediction in hierarchical prediction of SSD model is not reused, this paper uses feature fusion method to improve the network structure, and realizes the fusion of conv4_3 and fc7 layers by feature connection and corresponding element addition. Then a new feature pyramid is generated on the fusion layer for multi-scale detection, so as to increase the relationship between the model feature layers,so that the feature of multiple layers is better utilized and the detection accuracy is also improved.

Aiming at the problem that the small target features in the front layer of the model are obvious but the semantic information is less,this paper adds feature enhancement module after the last improvement. Based on the idea of multi-branch convolution and residual network, A dilated convolution layer is introduced to expand the receptive field of features, which makes the feature information learned by the network richer, improves the detection accuracy and reduces the miss detection of small targets.

Aiming at the problem that the online difficult sample mining mechanism in the SSD network eliminates the easy sample, which makes the easy sample unable to further improve the training effect,this paper removes the original online difficult sample mining mechanism, introduces the focal loss function into the SSD position regression loss function to balance the problem of positive and negative sample, enhances the model background resolution ability and reduces the miss detection rate.

Because there is no public satellite cloud map data set, this paper selects the infrared cloud map of FY2 satellite, and the self-built data set is recorded as fy_cloud,Then the model is trained and tested after numbering, marking, grouping.The experimental results show that the accuracy of the cloud_detection model in the fy_cloud data set is 16.02% higher than that of the SSD model. According to the recognition results of satellite cloud images in each period, the overall detection results of convective cloud are good, the missing detection rate and false detection rate are low.

[1]翁理国, 张旭, 夏旻, 等. 关于卫星气象云图准确检测的仿真研究[J]. 计算机仿真, 2019, 36(01): 429-436.

[2]刘漩. 一次持续性强对流天气过程的成因分析[J]. 海峡科技与产业, 2017(03): 71-72.

[3]卢宏涛, 张秦川. 深度卷积神经网络在计算机视觉中的应用研究综述[J]. 数据采集与处理, 2016, 31(01): 1-17.

[4]朱亚平, 程周杰, 刘健文. 一次锋面气旋云系中强对流云团的识别[J]. 应用气象学报, 2009, 20(04): 428-436.

[5]赵文化, 单海滨. 基于红外窗区与水汽通道对流云团识别方法研究[J]. 气象, 2018, 44(06): 814-824.

[6]陈青青, 李彪, 汤志亚, 等. 地基可见光全天空云图云量图像处理识别方法[J]. 气象科技, 2017, 45(06): 1006-1010.

[7]Li P, Dong L, Xiao H, et al. A cloud image detection method based on SVM vector machine[J]. Neurocomputing, 2015, 169: 34-42.

[8]李俊杰, 傅俏燕. “高分七号”卫星遥感影像自动云检测[J]. 航天返回与遥感, 2020, 41(02): 108-115.

[9]Bai T, D Li, Sun K, et al. Cloud Detection for High-Resolution Satellite Imagery Using Machine Learning and Multi-Feature Fusion[J]. Remote Sensing, 2016, 8(9): 715.

[10]Zhang Q, Yu Y, Zhang W, et al. Cloud Detection from FY-4A's Geostationary Interferometric Infrared Sounder Using Machine Learning Approaches[J]. Remote Sensing, 2019, 11(24): 3035.

[11]Hughes M, Daniel H. Automated Detection of Cloud and Cloud Shadow in Single-Date Landsat Imagery Using Neural Networks and Spatial Post-Processing[J]. Remote Sensing, 2014, 6(6): 4907-4926.

[12]张春桂, 周乐照, 林炳青. 基于卫星和雷达资料的对流云团识别跟踪[J]. 气象科技, 2017, 45(03): 485-491.

[13]金炜, 符冉迪, 范亚会, 等. 采用多模糊支持向量机决策融合的积雨云检测[J]. 光学精密工程, 2014, 22(12): 3427-3434.

[14]Francis A, Sidiropoulos P, Muller J P. CloudFCN: Accurate and Robust Cloud Detection for Satellite Imagery with Deep Learning[J]. Remote Sensing, 2019, 11(19): 2312.

[15]安捷, 马尽文. 基于全卷积网络的遥感图像自动云检测[J]. 信号处理, 2019, 35(04): 556-562.

[16]翁理国, 孔维斌, 夏旻, 等. 基于深度极限学习机的卫星云图云量计算[J]. 计算机科学, 2018, 45(04): 227-232.

[17]Xie F, Shi M, Shi Z, et al. Multilevel cloud detection in remote sensing images based on deep learning[J]. IEEE Journal of Selected Topics in Applied Earth Observations and Remote Sensing, 2017, 10(8): 3631-3640.

[18]Chai D, Newsam S, Zhang H K, et al. Cloud and cloud shadow detection in Landsat imagery based on deep convolutional neural networks[J]. Remote Sensing of Environment, 2019, 225: 307-316.

[19]Li W, Zou Z, Shi Z. Deep Matting for Cloud Detection in Remote Sensing Images[J]. IEEE Transactions on Geoscience and Remote Sensing, 2020, PP(99): 1-13.

[20]裴亮, 刘阳, 谭海, 等. 基于改进的全卷积神经网络的资源三号遥感影像云检测[J]. 激光与光电子学进展, 2019, 56(05): 226-233.

[21]Yin Z, Ling F, Foody G M, et al. Cloud detection in Landsat-8 imagery in Google Earth Engine based on a deep neural network[J]. 2020. 11(12): 1181-1190.

[22]Li Y, Chen W, Zhang Y, et al. Accurate cloud detection in high-resolution remote sensing imagery by weakly supervised deep learning[J]. Remote Sensing of Environment, 2020, 250.

[23]张家强, 李潇雁, 李丽圆, 等. 基于深度残差全卷积网络的Landsat 8遥感影像云检测方法[J]. 激光与光电子学进展, 2020, 57(10): 364-371.

[24]Zhang J, Zhou Q, Wu J, et al. A Cloud Detection Method Using Convolutional Neural Network Based on Gabor Transform and Attention Mechanism with Dark Channel Subnet for Remote Sensing Image[J]. Remote Sensing, 2020, 12(19): 3261.

[25]郑益勤, 杨晓峰, 李紫薇. 深度学习模型识别静止卫星图像海上强对流云团[J]. 遥感学报, 2020, 24(01): 97-106.

[26]查少均, 金炜, 何彩芬, 等. 多通道特征融合Y型全卷积网络的对流云检测[J]. 光电子·激光, 2019, 30(10): 1068-1078.

[27]Cai L, Hu Y, Dong J, et al. Audio-Textual Emotion Recognition Based on Improved Neural Networks[J]. Mathematical Problems in Engineering, 2019, 2019(6): 1-9.

[28]Zhang C, Pan X, Li H, et al. A hybrid MLP-CNN classifier for very fine resolution remotely sensed image classification[J]. Isprs Journal of Photogrammetry & Remote Sensing, 2018, 140: 133-144.

[29]Zahran E, Khater M. Modified extended tanh-function method and its applications to the Bogoyavlenskii equation[J]. Applied Mathematical Modelling, 2016, 40(3): 1769-1775.

[30]Yarotsky, Dmitry. Error bounds for approximations with deep ReLU networks[J]. Neural networks: the official journal of the International Neural Network Society, 2017, 94: 103.

[31]Weiwei S, Yihong G, Xiaoyu T, et al. Fine-Grained Image Classification Using Modified DCNNs Trained by Cascaded Softmax and Generalized Large-Margin Losses[J]. IEEE Transactions on Neural Networks and Learning Systems, 2019, 30: 683-694.

[32]范钦民. 基于多层特征融合的SSD目标检测[D]. 西南交通大学, 2018.

[33]Wang B, Gu X, Li M, et al. Temperature error correction based on BP neural network in meteorological wireless sensor network[J]. International Journal of Sensor Networks, 2017, 23(4): 265.

[34]黎静华, 黄乾, 韦善阳, 等. 基于S-BGD和梯度累积策略的改进深度学习方法及其在光伏出力预测中的应用[J]. 电网技术, 2017(10): 209-217.

[35]Girshick R, Donahue J, Darrell T, et al. Rich feature hierarchies for accurate object detection and semantic segmentation[C]// CVPR. IEEE, 2014: 580-587.

[36]He K, Zhang X, Ren S, et al. Spatial Pyramid Pooling in Deep Convolutional Networks for Visual Recognition[J]. IEEE Transactions on Pattern Analysis & Machine Intelligence, 2014, 37(9): 1904-16.

[37]Girshick R. Fast R-CNN[C]//Proceedings of 2015 IEEE International Conference on Computer Vision, 2015: 1440-1448．

[38]Ren S, He K, Girshick R, et al. Faster R-CNN: Towards real-time object detection with region proposal networks[C]//Advances in Neural Information Processing systems, 2015: 91-99.

[39]Redmon J, Divvala S, Girshick R, et al. You Only Look Once：Unified,real-time object detectiono[C]//Proceedingsof CVPR 2015, 2015: 779-788.

[40]Liu W, Anguelov D, et al. SSD: Single shot multibox detector[C]//Proceedings of European Conference on Computer Vision. Springer International Publishing, 2016: 21-37.

[41]Bodla N, Singh B, Chellappa R, et al. Soft-NMS - improving object detection with one line of code[C]//Proceedings of IEEE International Conference on Computer Vision, 2017: 5562-5570.

[42]Fu C Y , Liu W , Ranga A , et al. DSSD : Deconvolutional Single Shot Detector[J]. 2017.arXiv preprint arXiv:1701.06659.

[43]Ren J, Chen X, Liu J, et al. Accurate Single Stage Detector Using Recurrent Rolling Convolution[C]//Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition. 2017:752-760.

[44]张思宇, 张轶. 基于多尺度特征融合的小目标行人检测[J]. 计算机工程与科学, 2019, 41(09): 1627-1634.

[45]杨丽娟, 李利. 基于双线性插值的内容感知图像缩放算法仿真[J]. 计算机仿真, 2019, 36(12): 244-248.

[46]温捷文, 战荫伟, 李楚宏, 等. 一种加强SSD小目标检测能力的Atrous滤波器设计[J]. 计算机应用研究, 2019, 36(03): 861-865.

[47]Li Y, Tang S, Lin M, et al. Implicit Negative Sub-Categorization and Sink Diversion for Object Detection[J]. IEEE Transactions on Image Processing, 2018, 27(4): 1561-1574.

[48]Chu J, Guo Z, Leng L. Object Detection Based on Multi-Layer Convolution Feature Fusion and Online Hard Example Mining[J]. IEEE Access, 2018: 19959-19967.

[49]张建明, 刘煊赫, 吴宏林, 等. 面向小目标检测结合特征金字塔网络的SSD改进模型[J]. 郑州大学学报(理学版), 2019, 51(03): 61-66+72.

[50]高友文, 周本君, 胡晓飞. 基于数据增强的卷积神经网络图像识别研究[J]. 计算机技术与发展, 2018, 28(08): 62-65.

[51]张雪松, 庄严, 闫飞, 等. 基于迁移学习的类别级物体识别与检测研究与进展[J]. 自动化学报, 2019, 45(07): 1224-1243.