Wind vector prediction method using artificial intelligence network and analysis apparatus
A wind vector prediction method includes receiving, by an analysis device, a weather video; inputting, by the analysis device, a first base video of a first time point for at least one water vapor absorption channel among a plurality of water vapor absorption channels included in the weather video and a reference video of a time point different from the first time point into a first learning network, and predicting, by the analysis device, a wind vector of a second time point after the first time point for the at least one water vapor absorption channel based on information output from the first learning network.
1 . A wind vector prediction method comprising the steps of:
receiving, by an analysis device including a processor and a memory, a weather video;
inputting, by the analysis device, a first base video of a first time point for at least one water vapor absorption channel among a plurality of water vapor absorption channels included in the weather video and a reference video of a time point different from the first time point into a first learning network; and
predicting, by the analysis device, a wind vector of a second time point for the at least one water vapor absorption channel based on information output from the first learning network, the second time point for the at least one water vapor absorption channel being after the first time point for the at least one water vapor absorption channel,
wherein the first learning network is a learning network that outputs video information of the second time point using the first base video and the reference video,
wherein the first learning network comprises:
a first encoder for receiving the first base video, and generating a first feature map;
a second encoder for receiving a residual video of the first base video and the reference video, and generating a second feature map;
a convolutional long short-term memory (LSTM) for receiving the second feature map, and generating a third feature map of the second time point; and
a decoder for receiving a value obtained by summing the first feature map and the third feature map, and generating information on the wind vector, and
wherein the first learning network uses the first base video of the first time point for a channel indicating ground radiance to generate an attention map identifying a cloud region, the attention map being analyzed in at least one of the first encoder, the second encoder or the convolutional long short-term memory (LSTM) of the first learning network by assigning separate weights to cloud regions and cloudless regions identified from spectral difference.
2 . The method of claim 1 , wherein the reference video is a video of the first base video of a past time point, and the first learning network predicts the wind vector based on a difference between the first base video and the reference video, or the reference video is a video of the first base video of a future time point, and the first learning network predicts the wind vector of a time point between the first base video and the reference video based on the difference between the first base video and the reference video.
3 . The method of claim 1 , further comprising the steps of:
predicting, by the analysis device, the wind vector of the second time point by inputting a second base video of a third time point after the second time point in the at least one water vapor absorption channel and the reference video into a second learning network; and
predicting, by the analysis device, a final wind vector by averaging the wind vector predicted based on the first base video and the wind vector predicted based on the second base video.
4 . The method of claim 1 , wherein the first learning network is any one among a generative adversarial network (GAN), an adaptive convolution network (ACN), and a convolutional encoder-decoder.
5 . The method of claim 1 , wherein the information output from the first learning network is a video for the wind vector or a flow vector for the wind vector.
6 . The method of claim 1 , wherein the attention map is generated using a segmentation network including a convolutional encoder-decoder trained to segment cloud regions based on spectral differences between the ground radiance channel and the water vapor absorption channel.
7 . A wind vector prediction method comprising the steps of:
receiving, by an analysis device including a processor and a memory, a video of a channel indicating a ground radiance and a video of a water vapor absorption channel;
inputting, by the analysis device, the video of the channel indicating the ground radiance into a segmentation network including a convolutional encoder-decoder to distinguish a cloud region and a cloudless region in the video of the water vapor absorption channel;
generating, using the segmentation network by the analysis device, a segmentation map including a wind vector for the cloudless region using the video of the water vapor absorption channel;
inputting, by the analysis device, a first base video of a first time point for the water vapor absorption channel and a reference video of a time point different from the first time point into a first learning network; and
predicting, by the analysis device, a wind vector of a second time point after the first time point based on information output from the first learning network,
wherein the first learning network is a learning network that outputs video information of the second time point using the first base video and the reference video,
wherein the first learning network comprises:
a first encoder for receiving the first base video, and generating a first feature map;
a second encoder for receiving a residual video of the first base video and the reference video, and generating a second feature map;
a convolutional long short-term memory (LSTM) for receiving the second feature map, and generating a third feature map of the second time point; and
a decoder for receiving a value obtained by summing the first feature map and the third feature map, and generating information on the wind vector, and
wherein the first learning network uses the segmentation map as an attention map identifying the cloud region, the attention map being analyzed in at least one of the first encoder, the second encoder, the convolutional long short-term memory (LSTM), or the decoder of the first learning network by assigning separate weights to cloud regions and cloudless regions identified from spectral difference.
8 . The method of claim 7 , wherein the first learning network performs interpolation prediction or extrapolation prediction.
9 . The method of claim 7 , further comprising the steps of:
predicting, by the analysis device, the wind vector of the second time point by inputting a second base video of a third time point after the second time point and the reference video in the water vapor absorption channel into a second learning network including a convolutional encoder-decoder; and
predicting, by the analysis device, a final wind vector by averaging the wind vector predicted based on the first base video and the wind vector predicted based on the second base video.
10 . An analysis apparatus for predicting a wind vector using an artificial intelligence network, comprising:
an input device for receiving an input of a first base video of a first time point for at least one water vapor absorption channel and a reference video of a time point different from the first time point;
a storage device for storing a learning network that outputs video information of a second time point using the first base video and the reference video; and
a computing device including a processor and a memory for inputting the first base video and the reference video into the learning network including a convolutional encoder-decoder, and predicting a wind vector of the second time point after the first time p oint for the at least one water vapor absorption channel based on output information,
wherein the learning network performs interpolation prediction or extrapolation prediction, and
wherein the first learning network comprises:
a first encoder for receiving the first base video, and generating a first feature map;
a second encoder for receiving a residual video of the first base video and the reference video, and generating a second feature map;
a convolutional long short-term memory (LSTM) for receiving the second feature map, and generating a third feature map of the second time point; and
a decoder for receiving a value obtained by summing the first feature map and the third feature map, and generating information on the wind vector, and
wherein the first learning network uses the first base video of the first time point for a channel indicating ground radiance to generate an attention map identifying a cloud region, the attention map being analyzed in at least one of the first encoder, the second encoder or the convolutional long short-term memory (LSTM) of the first learning network by assigning separate weights to cloud regions and cloudless regions identified from spectral difference.