-
采用BP神经网络反演法和CNN反演法实现GIIRS大气温湿度廓线反演,两种方法实现流程如图2所示。先将训练样本集中大气温度、湿度和臭氧廓线数据作为RTTOV辐射传输模式输入,模拟得到GIIRS辐射通道亮温值,从中选择出1.2节中225个反演通道的亮温数据,分别组成温度廓线反演样本对和湿度廓线反演样本对。之后将反演样本对均匀间隔选取1000个样本对作为独立的测试样本,剩余的11528个样本对作为训练样本,并用BP神经网络法和CNN法分别建立反演模型,通过对两种模型框架及参数不断的优化和调整,得到最佳的反演模型。最终用1000个独立测试样本的亮温数据作为反演模型的输入,得到反演的大气温湿度廓线,将反演结果与测试样本廓线“真值”进行反演结果精度测试,并对两种反演算法进行比较和分析。
-
BP (Back propagation Neural Network)神经网络是目前人工神经网络算法中应用最广泛的模型,能够实现任意精度的连续函数映射,有效用于复杂的非线性函数的逼近[10]。文中采用的是一个三层的前馈网络,包括输入层、隐含层和输出层。选择的225个通道的模拟亮温值作为该BP神经网络的输入层,对应的101层大气温度、湿度廓线分别作为输出层,即输入层有225个节点,输出层有101个节点。隐含层设置为一层,网络性能的好坏受隐含层节点数设置的影响,如果隐含层节点数太少,会使得信号资料不够,则必定会影响网络的效果;而隐含层节点数太多,则会使得训练花费更多的时间,降低业务效率。基于以前的研究[19-21],目前应用较广泛的关于隐含层节点数选取的方法有三种,分别为公式(1)~(3),经过试验及比较,对于隐含层节点数的设置主要参考了公式(3),即隐含层节点为108个。
$$ h={\mathrm{log}}_{2}T $$ (1) $$ h=\sqrt{mn} $$ (2) $$ h=\sqrt{(0.43mn+0.12{m}^{2}+2.54n+0.77m+0.35)}+0.51 $$ (3) 式中:T 为训练样本的个数;
$ h $ 为隐含层的节点数;$ m $ 为输出层的节点数;$ n $ 为输入层的节点数。使用Newff函数创建神经网络,带动量的梯度下降法traingdx作为网络的训练函数,双曲正切S型传递函数tansig作为网络的激活函数。神经网络训练参数如表1所示。
表 1 BP神经网络训练参数
Table 1. Training parameters of BP neural network
Parameter Set value Attributes Net.trainParam.epochs 10000 Training times Net.trainParam.goal 0 Training goal Net.trainParam.lr 0.01 Learning rate Net.trainParam.mc 0.95 Momentum factor Net.trainParam.show 25 Number of intervals displayed Net.trainParam.min_grad 1×10−6 Minimum performance gradient -
卷积神经网络(Convolutional Neural Networks, CNN)在图像识别与分类、语义分割、智能驾驶等领域已有优异的表现[22-23],它是一种包含了卷积计算并且有深度结构的前馈神经网络,是深度学习算法代表之一。它也是一种特殊的深层神经网络模型,其特殊性主要体现在两方面:一方面是它的神经元之间的连接是非全连接的;另一方面是在同一层中某些神经元间的连接权重是共享的。它的非全连接和权值共享的网络结构降低了网络模型的复杂度,减少了权值的数量。
CNN的结构主要包括输入层、隐含层和输出层。其中隐含层主要包括卷积层、池化层、全连接层这三种常见的网络层,隐含层是实现特征提取的关键技术。卷积层的主要功能就是对输入进行特征提取,它的内部包含多个卷积核,卷积层内的每个神经元都与前一层中位置接近的区域的多个神经元相连,区域的大小取决于卷积核的大小。卷积核在进行卷积工作时相当于滤波器按设定的步长在输入图片上进行滑动操作,输入图片与卷积核进行卷积计算,并产生与卷积核个数相同的特征映射图。卷积运算如下:
$$ {y}^{(l+1)}={k}^{\left(l\right)}{x}^{\left(l\right)}+{b}^{\left(l\right)} $$ (4) 式中:
$ {x}^{\left(l\right)} $ 为上层的输出;$ {k}^{\left(l\right)} $ 为$ l $ 层中的某个卷积核;$ {b}^{\left(l\right)} $ 为偏置。则该层的输出为:$$ {x}^{(l+1)}=f\left({x}^{\left(l\right)}\right) $$ (5) 式中:
$ f\left(*\right) $ 为激活函数。池化层相当于一个降采样的过程。输入的图片在经过卷积层内不同卷积核的卷积运算后,输出的特征图会被传递到池化层进行特征选择和信息过滤,池化层通过对不同位置的特征作聚合统计,得到维度较低的统计特征[24],目前常用的池化操作有最大值池化和平均值池化。全连接层从网络结构上讲等同于传统神经网络中的隐含层,位于卷积神经网络隐含层的最后部分,与卷积层和池化层的局部连接方式不同的是全连接层的每一个节点都需与上一层的所有节点相连,从而将前面几层提取到的局部特征进行非线性组合综合得到全局特征。卷积神经网络输出层的上一层通常都是全连接层,其工作原理和结构与传统的前馈神经网络的输出层相同。
-
采用的模型包含一个输入层,四个卷积层、两个池化、一个全连接层和一个回归输出层,卷积层和池化层交替设置构成一个多层神经网络,框架结构如图3所示,图中卷积神经网络的输入层为每个样本的225个通道亮温值,可以将其视为宽为1的一维图像,每个像素点为每个通道的亮温,所以输入层大小为
$ \text{225}\text{×1}\text{} $ ,输出层为$ \text{1}\text{×101} $ 的大气温度、湿度标签,用一维卷积核进行卷积操作,图中深色部分为卷积核大小。卷积层主要提取样本数据特征,将会输出若干个特征图(feature maps),是卷积神经网络中最核心的部分。由于深度学习的CNN网络框架复杂、模型参数众多,实际应用中,为了得到最优的模型参数,需对其进行反复调试。以温度反演模型为例,计算不同参数设置时网络的训练误差。参考指标为验证均方根误差RMSE、温度反演RMSE和网络训练时间,RMSE越小、训练时间越短则网络性能越好。其中:验证RMSE为网络提供的检验参数,温度反演RMSE为用训练库中均匀间隔选取的1000对独立测试样本统计的反演误差。以第一卷积层(conv_1)为例,测试选取不同卷积核大小时的参考指标如表2所示,其中第一列为卷积核大小,第二列为验证RMSE,第三列为独立检验样本的温度反演RMSE,第四列为网络训练时长,发现卷积核越大则网络训练时间越长,而验证RMSE和温度反演RMSE则越小,综合考虑参考指标和输入信号的大小,最终第一卷积层的卷积核大小设置为
$ \text{5}\text{×1} $ 的一维卷积核。测试第一卷积层最优输出特征图(feature maps)个数的参考指标如表3所示,同表2,发现输出特征图数越多则验证和温度反演RMSE越小、网络训练时间越久,但是输出特征图数从50到60时,验证RMSE和温度反演RMSE不再减小,训练时间却明显加长,所以最终选择第一卷积层输出特征图数为50。表 2 第一卷积层不同卷积核大小时网络训练参考指标
Table 2. Network training reference indicators for different convolution kernel sizes of conv_1
Convolution
kernel sizeValidation
RMSETemperature retrieval
RMSE/KTraining
time$ 3\times 1 $ 0.8521 3.4625 12'48'' $ 4\times 1 $ 0.8329 3.4217 13'14'' $ 5\times 1 $ 0.8298 3.4137 14'20'' $ 6\times 1 $ 0.8274 3.3871 15'00'' 表 3 第一卷积层不同输出特征图数时网络训练参考指标
Table 3. Network training reference indicators for different output feature maps of conv_1
Number of output
feature mapsValidation
RMSETemperature retrieval
RMSE/KTraining
time20 0.7916 3.2888 16'14'' 30 0.7816 3.2427 22'47'' 40 0.7730 3.2245 27'32'' 50 0.7595 3.1698 27'55'' 60 0.7617 3.1756 35'80'' 其他网络参数最优选择与上述方法类似,不再赘述。每一个卷积层都选用同样大小的滤波器(即卷积核大小
$ \text{5}\text{×1} $ 的一维卷积核),卷积核移动的步幅为1,池化层采用大小为$ \text{2}\text{×1} $ 、步幅为2的平均值池化方式。卷积层中每个神经元的输入通过激活函数映射到输出端,增加网络模型的非线性,常用的激活函数有sigmoid、tanh和ReLU函数,其中sigmoid和tanh函数由于都存在饱和的问题,容易造成梯度消失,而ReLU函数为不饱和的非线性函数,可以避免梯度消失、梯度爆炸等问题,同时在网络训练时具有更快的收敛速度,因此激活函数选为ReLU函数。为了加快网络的训练速度,网络训练优化器选Adam[14],与传统的优化器相比,Adam具有更高的计算效率以及较低的内存需求。文中CNN网络模型完整构架及参数如表4所示。表 4 CNN模型完整构架
Table 4. Complete structure of CNN model
Name Type Output Parameter imageinput Image input $ \text{225×1×1} $ - conv_1 Convolution $ \text{225×1×50} $ $ \text{5}\text{×1×1×50} $ batchnorm_1 Batch normalization $ \text{225×1×50} $ $ \text{1}\text{×1×50} $ relu_1 ReLU $ \text{225}\text{×1×50} $ - avgpool_1 Average pooling $ \text{112×1×50} $ - conv_2 Convolution $ \text{112×1×100} $ $ \text{5}\text{×1×50×100} $ batchnorm_2 Batch normalization $ \text{112×1×100} $ $ \text{1}\text{×1×100} $ relu_2 ReLU $ \text{112×1×100} $ - avgpool_2 Average pooling $ \text{56×1×100} $ - conv_3 Convolution $ \text{56×1×100} $ $ \text{5}\text{×1×100×100} $ batchnorm_3 Batch normalization $ \text{56×1×100} $ $ \text{1}\text{×1×100} $ relu_3 ReLU $ \text{56×1×100} $ - conv_4 Convolution $ \text{56×1×100} $ $ \text{5}\text{×1×100×100} $ batchnorm_4 Batch normalization $ \text{56×1×100} $ $ \text{1}\text{×1×100} $ relu_4 ReLU $ \text{56×1×100} $ - dropout Dropout $ \text{56×1×100} $ - fc Fully connected $ \text{1×1×101} $ $ \text{101}\text{×5600} $ regressionoutput Regression output - - -
将上述均匀间隔选取的1000个独立的测试样本,分别用BP神经网络法和CNN法建立的模型进行大气温度、湿度垂直廓线反演,并统计反演误差。精度评价指标有平均误差ME、均方根误差RMSE和平均相对误差MRE,定义分别为:
$$ ME=\frac{1}{{N}_{s}}\sum _{i=1}^{{N}_{s}}({y}_{i}-{x}_{i}) $$ (6) $$ RMS E=\sqrt{\frac{1}{{N}_{s}}\sum _{i=1}^{{N}_{s}}{({x}_{i}-{y}_{i})}^{2}} $$ (7) $$ MRE=\dfrac{\dfrac{1}{{N}_{s}}\displaystyle\sum _{i=1}^{{N}_{s}}\left|({x}_{i}-{y}_{i})\right|}{\dfrac{1}{{N}_{s}}\displaystyle\sum _{i=1}^{{N}_{s}}{x}_{i}}\times 100 {\text{%}} $$ (8) 式中:
$ {x}_{i} $ 为检验样本廓线;$ {y}_{i} $ 为GIIRS反演的大气廓线;$ {N}_{s} $ 为检验样本个数。针对不同地表类型和是否有云的影响对BP法和CNN法分别建立了三种网络模型:第一方案为对样本廓线不做任何方式的分类;第二方案将样本分为陆地和海洋,分别进行建模和反演;第三方案参考ZHA10方法[25]将样本分为有云和晴空分别建模和反演。
图4显示了整层大气(0.005~1100 hPa)的温度反演值和“真值”的散点分布图,其中图4(a)不分类、图4(b)陆地/洋面分类、图4(c)晴空/云分类为BP神经网络法得到的散点图;图4(d)~(f)使用的是CNN算法,横坐标为测试样本温度值,纵坐标为反演温度值。从图中可以看出,两种方法反演值和“真值”有较高的一致性(相关系数高达0.99),均分布在直线y=x(图中红线)的两侧。BP神经网络法三种分类方案的相关系数R均为0.989,平均误差ME分别为0.01 K、0.006 K、0.008 K,均方根误差RMSE分别为3.96 K、4.03 K、3.92 K,平均相对误差MRE分别为0.95%、0.98%、0.94%;CNN法三种分类方案的R均为0.99,ME分别为−0.02 K、−0.0008 K、0.01 K,RMSE分别为2. 90 K、3.03 K、3.07 K,MRE分别为0.65%、0.69%、0.69%。从R、ME、RMSE和MRE都可以看出,CNN法优于BP神经网络法。图5与图4相似,显示了水汽混合比的散点图。同样,CNN法反演的R、ME、RMSE和MRE均优于BP神经网络法。至于不同分类方案,BP神经网络法用方案三(即将样本分为有云和晴空)时得到的反演结果较其他两种方案稍有提高,均方根误差较小;而CNN法不对样本数据进行任何分类时其反演均方根误差最小。
图 4 温度反演散点图。(a)~(c) BP神经网络三种分类方案;(d)~(f) CNN三种分类方案
Figure 4. Temperature scatter diagram of retrieval. (a)-(c) Three classification schemes of BP neural network; (d)-(f) Three classification schemes of CNN
图6和图7分别显示了检验样本统计的温度反演平均误差、均方根误差和平均相对误差垂直廓线,其中(a)、(b)、(c)分别为三种不同分类方案,虚线代表BP神经网络法,实线代表CNN法,图6中红色代表平均误差,黑色代表均方根误差,图7则为平均相对误差。从图中可以看出,两种算法的平均误差在所有高度上相当,均较小;相较于BP神经网络法,CNN法就均方根误差RMSE和平均相对误差MRE而言均有明显的改进,在高层10~200 hPa改进较大,三种分类方案RMSE改进的最大值分别为1.15 K、1.06 K和1.02 K,MRE改进的最大值分别为0.45%、0.42%和0.39%,对流层低层600~1000 hPa均方根误差减小较小,RMSE减小了约0.6 K,MRE减小了约0.17%。为进一步分析不同分类方案的差别,两张图中的(d)分别给出了CNN法三种分类方案时温度反演均方根误差和平均相对误差垂直廓线(即(a)~(c)图中的黑色实线),其中实线代表方案一,点画线代表方案二,虚线代表方案三。可以看出,方案一反演结果较其他两种方案在所有高度层均方根误差和平均相对误差都偏小,而分类方案二和三则在不同高度上各有优劣。
图 6 温度反演误差廓线。(a)~(c) 三种分类方案,红色代表偏差,黑色代表均方根误差,虚线代表BP神经网络法,实线代表CNN法;(d) CNN三种分类方案均方根误差廓线,实线方案一,点画线方案二,虚线方案三
Figure 6. Error profile of retrieval for temperature. (a)-(c) Three classification schemes, red is the bias, black is the root mean square error, dotted line is the BP neural network method, and the solid line is the CNN method; (d) Root mean square error profile of the three classification schemes of CNN, the solid line is the first scheme, the dotted line is the second scheme, and the dashed line is the third scheme
图 7 温度反演平均相对误差廓线。(a)~(c) 三种分类方案,虚线代表BP神经网络法,实线代表CNN法;(d) CNN三种分类方案平均相对误差廓线,实线方案一,点画线方案二,虚线方案三
Figure 7. Mean relative error profile of retrieval for temperature. (a)-(c) Three classification schemes, dotted line is the BP neural network method, and the solid line is the CNN method; (d) Mean relative error profile of the three classification schemes of CNN, the solid line is the first scheme, the dotted line is the second scheme, and the dashed line is the third scheme
图8和图9分别给出了水汽混合比反演的平均误差、均方根误差和平均相对误差垂直廓线,同样,CNN法比BP神经网络法反演的湿度廓线误差小,RMSE在对流层低层500~1000 hPa改进较大,三种分类方案分别平均改进了0.43 g/kg、0.41 g/kg和0.34 g/kg,MRE在各层改进均较大,三种分类方案分别在各层平均改进了7.06%、7.41%和5.56%。两张图中的(d)同样分别显示了不同分类方案水汽混合比均方根误差和平均相对误差垂直廓线,对CNN法而言,在所有高度层方案一反演精度均高于其他两种分类方案,尤其低层500~1000 hPa改进最大。
Atmospheric temperature and humidity profile retrievals using a machine learning algorithm based on satellite-based infrared hyperspectral observations
-
摘要: 星载红外高光谱垂直探测仪GIIRS (Geostationary Interferometric Infrared Sounder)能够实现大气温度和湿度参数高垂直分辨率的观测,为数值天气预报提供精度更高的初始场。基于GIIRS观测辐射值采用BP神经网络(Back Propagation Neural Network)法和深度学习的卷积神经网络(Convolutional Neural Networks, CNN)法反演大气温度、湿度垂直廓线,重点在于CNN法模型的构建与参数的优化,得到反演精度最高的网络模型配置。将训练样本根据不同地表类型和是否有云的影响分为三种方案(方案一:不分类、方案二:陆地/洋面分类、方案三:晴空/有云分类),分别进行建模、反演和检验。结果表明两种反演算法都有较好的反演精度,相对而言CNN法在所有高度层上反演偏差、均方根误差和平均相对误差均较小,反演精度更高。CNN法温度反演在高层10~200 hPa改进较大,三种分类方案改进的最大值分别为1.15 K、1.06 K和1.02 K;湿度反演在对流层低层500~1000 hPa改进较大,三种分类方案分别平均改进了0.43 g/kg、0.41 g/kg和0.34 g/kg。BP神经网络法方案三时(即分晴空和云时)温度和水汽混合比廓线反演精度最好;CNN算法方案一时(即不对样本数据进行任何分类)反演精度最高。Abstract: The satellite-based infrared hyperspectral Geostationary Interferometric Infrared Sounder (GIIRS) can achieve high vertical resolution observations of atmospheric temperature and humidity parameters, which provide a more accurate initial field for numerical weather forecasting. Based on GIIRS observation radiation, a back propagation (BP) neural network and deep learning convolutional neural networks (CNNs) are used to retrieve atmospheric temperature and humidity profiles, and the focus is on the construction of the CNN model and the optimization of parameters, thus obtaining the network model configuration with the highest retrieval accuracy. The training samples are divided into three schemes according to different surface types and the influence of whether there are clouds (scheme 1: no classification, scheme 2: land or ocean surface, scheme 3: clear or clouds) and modelling, retrieving and testing. The results show that the two retrieval algorithms both have good retrieval precision. Relatively speaking, the CNN method has a smaller retrieval bias, root-mean-square error and mean relative error at all altitudes, and the retrieval precision is higher. The temperature retrieval of the CNN method is greatly improved in the high level at 10-200 hPa, and the maximum values of the three classification schemes are 1.15 K, 1.06 K, and 1.02 K, respectively, and the humidity retrieval of the CNN method also shows improvement in the lower troposphere at 500-1000 hPa, and the averages of the three classification schemes are 0.43 g/kg, 0.41 g/kg, and 0.34 g/kg, respectively. The third scheme (clear or clouds) of the BP neural network method has the best retrieval precision of temperature and water vapour mixing ratio profiles, and the first scheme (no classification of sample data) of the CNN algorithm has the most accurate retrieval results.
-
图 6 温度反演误差廓线。(a)~(c) 三种分类方案,红色代表偏差,黑色代表均方根误差,虚线代表BP神经网络法,实线代表CNN法;(d) CNN三种分类方案均方根误差廓线,实线方案一,点画线方案二,虚线方案三
Figure 6. Error profile of retrieval for temperature. (a)-(c) Three classification schemes, red is the bias, black is the root mean square error, dotted line is the BP neural network method, and the solid line is the CNN method; (d) Root mean square error profile of the three classification schemes of CNN, the solid line is the first scheme, the dotted line is the second scheme, and the dashed line is the third scheme
图 7 温度反演平均相对误差廓线。(a)~(c) 三种分类方案,虚线代表BP神经网络法,实线代表CNN法;(d) CNN三种分类方案平均相对误差廓线,实线方案一,点画线方案二,虚线方案三
Figure 7. Mean relative error profile of retrieval for temperature. (a)-(c) Three classification schemes, dotted line is the BP neural network method, and the solid line is the CNN method; (d) Mean relative error profile of the three classification schemes of CNN, the solid line is the first scheme, the dotted line is the second scheme, and the dashed line is the third scheme
表 1 BP神经网络训练参数
Table 1. Training parameters of BP neural network
Parameter Set value Attributes Net.trainParam.epochs 10000 Training times Net.trainParam.goal 0 Training goal Net.trainParam.lr 0.01 Learning rate Net.trainParam.mc 0.95 Momentum factor Net.trainParam.show 25 Number of intervals displayed Net.trainParam.min_grad 1×10−6 Minimum performance gradient 表 2 第一卷积层不同卷积核大小时网络训练参考指标
Table 2. Network training reference indicators for different convolution kernel sizes of conv_1
Convolution
kernel sizeValidation
RMSETemperature retrieval
RMSE/KTraining
time$ 3\times 1 $ 0.8521 3.4625 12'48'' $ 4\times 1 $ 0.8329 3.4217 13'14'' $ 5\times 1 $ 0.8298 3.4137 14'20'' $ 6\times 1 $ 0.8274 3.3871 15'00'' 表 3 第一卷积层不同输出特征图数时网络训练参考指标
Table 3. Network training reference indicators for different output feature maps of conv_1
Number of output
feature mapsValidation
RMSETemperature retrieval
RMSE/KTraining
time20 0.7916 3.2888 16'14'' 30 0.7816 3.2427 22'47'' 40 0.7730 3.2245 27'32'' 50 0.7595 3.1698 27'55'' 60 0.7617 3.1756 35'80'' 表 4 CNN模型完整构架
Table 4. Complete structure of CNN model
Name Type Output Parameter imageinput Image input $ \text{225×1×1} $ - conv_1 Convolution $ \text{225×1×50} $ $ \text{5}\text{×1×1×50} $ batchnorm_1 Batch normalization $ \text{225×1×50} $ $ \text{1}\text{×1×50} $ relu_1 ReLU $ \text{225}\text{×1×50} $ - avgpool_1 Average pooling $ \text{112×1×50} $ - conv_2 Convolution $ \text{112×1×100} $ $ \text{5}\text{×1×50×100} $ batchnorm_2 Batch normalization $ \text{112×1×100} $ $ \text{1}\text{×1×100} $ relu_2 ReLU $ \text{112×1×100} $ - avgpool_2 Average pooling $ \text{56×1×100} $ - conv_3 Convolution $ \text{56×1×100} $ $ \text{5}\text{×1×100×100} $ batchnorm_3 Batch normalization $ \text{56×1×100} $ $ \text{1}\text{×1×100} $ relu_3 ReLU $ \text{56×1×100} $ - conv_4 Convolution $ \text{56×1×100} $ $ \text{5}\text{×1×100×100} $ batchnorm_4 Batch normalization $ \text{56×1×100} $ $ \text{1}\text{×1×100} $ relu_4 ReLU $ \text{56×1×100} $ - dropout Dropout $ \text{56×1×100} $ - fc Fully connected $ \text{1×1×101} $ $ \text{101}\text{×5600} $ regressionoutput Regression output - - -
[1] Wang Ying, Huang Yong, Huang Siyuan. A preliminary study of the retrieval methods for atmosphere temperature and humidity profiles [J]. Remote Sensing for Land & Resources, 2008, 20(1): 23-26. (in Chinese) [2] Liu Hui, Dong Chaohua, Zhang Wenjian. New characteristics of satellite infrared atmospheric detector development over the world [J]. Meteorological Science and Technology, 2006, 34(5): 600-605. (in Chinese) [3] Qiu Jinhuan, Chen Hongbin, Wang Pucai, et al. A prospect on future atmospheric remote sensing [J]. Chinese Journal of Atmospheric Sciences, 2005, 29(1): 131-136. (in Chinese) [4] Smith Sr W L, Weisz E, Kireev S V, et al. Dual-regression retrieval algorithm for real-time processing of satellite ultraspectral radiances [J]. Journal of Applied Meteorology and Climatology, 2012, 51(8): 1455-1476. doi: 10.1175/JAMC-D-11-0173.1 [5] Guan Yuanhong, Ren Jie, Bao Yansong, et al. Research of the infrared high spectral (IASI) satellite remote sensing atmospheric temperature and humidity profiles based on the one-dimensional variational algorithm [J]. Trans Atmos Sci, 2019, 42(4): 602-611. (in Chinese) [6] Zhu L, Bao Y, Petropoulos G P, et al. Temperature and humidity profiles retrieval in a plain area from Fengyun-3D/HIRAS sensor using a 1D-VAR assimilation scheme [J]. Remote Sensing, 2020, 12(3): 435. doi: 10.3390/rs12030435 [7] Chakraborty R, Maitra A. Retrieval of atmospheric properties with radiometric measurements using neural network [J]. Atmospheric Research, 2016, 181: 124-132. doi: 10.1016/j.atmosres.2016.05.011 [8] Van Damme M, Whitburn S, Clarisse L, et al. Version 2 of the IASI NH3 neural network retrieval algorithm: near-real-time and reanalysed datasets [J]. Atmospheric Measurement Technique, 2017, 10(12): 4905-4914. doi: 10.5194/amt-10-4905-2017 [9] Kolassa J, Gentine P, Prigent C, et al. Soil moisture retrieval from AMSR-E and ASCAT microwave observation synergy. Part 2: Product evaluation [J]. Remote Sensing of Environment, 2017, 195: 202-217. doi: 10.1016/j.rse.2017.04.020 [10] Guan Li, Liu Yang, Zhang Xuehui. Application of artificial neural network algorithm in retrieving atmospheric temperature profiles from hyperspectral infrared data [J]. Trans Atmos Sci, 2010, 33(3): 341-346. (in Chinese) [11] Liu Yang, Guan Li. Study on the inversion of clear sky atmospheric humidity profiles with artificial neural network [J]. Meteorological Monthly, 2011, 37(3): 318-324. (in Chinese) [12] Huang P, Guo Q, Han C, et al. An improved method combining ANN and 1D-Var for the retrieval of atmospheric temperature profiles from FY-4A/GIIRS hyperspectral data [J]. Remote Sensing, 2021, 13(3): 481. [13] Milstein A B, Blackwell W J. Neural network temperature and moisture retrieval algorithm validation for AIRS/AMSU and CrIS/ATMS [J]. Journal of Geophysical Research:Atmospheres, 2016, 121(4): 1414-1430. doi: 10.1002/2015JD024008 [14] Mahngren Hansen D, Laparra V, Nielsen A A, et al. Statistical retrieval of atmospheric profiles with deep convolutional neural networks [J]. ISPRS Journal of Photogrammetry and Remote Sensing, 2019, 158: 231-240. doi: 10.1016/j.isprsjprs.2019.10.002 [15] Seemann S W, Borbas E E, Knuteson R O, et al. Development of a global infrared land surface emissivity database for application to clear sky sounding retrievals from multispectral satellite radiance measurements [J]. Journal of Applied Meteorology and Climatology, 2008, 47(1): 108-123. doi: 10.1175/2007JAMC1590.1 [16] Zhu L, Li J, Zhao Y, et al. Retrieval of volcanic ash height from satellite-based infrared measurements [J]. Journal of Geophysical Research:Atmospheres, 2017, 122(10): 5364-5379. doi: 10.1002/2016JD026263 [17] Saunders R, Hocking J, Turner E, et al. An update on the RTTOV fast radiative transfer model (currently at version 12) [J]. Geoscientific Model Development, 2018, 11(7): 2717-2737. doi: 10.5194/gmd-11-2717-2018 [18] Gambacorta A, Barnet C D. Methodology and information content of the NOAA NESDIS operational channel selection for the cross-track infrared sounder (CrIS) [J]. IEEE Transactions on Geoscience and Remote Sensing, 2013, 51(6): 3207-3216. doi: 10.1109/TGRS.2012.2220369 [19] Yu P, Shi C, Yang L, et al. A new temperature channel selection method based on singular spectrum analysis for retrieving atmospheric temperature profiles from FY-4A/GIIRS [J]. Advances in Atmospheric Sciences, 2020, 37(7): 735-750. doi: 10.1007/s00376-020-9249-9 [20] Wanas N, Auda G, Kamel M S, et al. On the optimal number of hidden nodes in a neural network [C]//Conference Proceedings of IEEE Canadian Conference on Electrical and Computer Engineering, 1998, 2: 918-921. [21] Gao Daqi. On structures of supervised linear basis function feedforward three-layered neural networks [J]. Chinese Journal of Computers, 1998, 21(1): 80-86. (in Chinese) [22] Zhou Hongqiang, Huang Lingling, Wang Yongtian. Deep learning algorithm and its application in optics [J]. Infrared and Laser Engineering, 2019, 48(12): 1226004. (in Chinese) doi: 10.3788/IRLA201948.1226004 [23] Xue Shan, Zhang Zhen, Lv Qiongying, et al. Image recognition method of anti UAV system based on convolutional neural network [J]. Infrared and Laser Engineering, 2020, 49(7): 20200154. (in Chinese) doi: 10.3788/IRLA20200154 [24] Niu Yaxi, Ji Xiaoping. Image retrieval algorithm based on convolutional neural network [J]. Computer Engineering and Applications, 2019, 55(18): 201-206. (in Chinese) [25] Zhang J, Chen H, Li Z, et al. Analysis of cloud layer structure in Shouxian, China using RS92 radiosonde aided by 95 GHz cloud radar [J]. Journal of Geophysical Research, 2010, 115: D00K30.