两类遥感图像深度神经网络预测方法及比较

Two Types of Remote-sensing Image Deep Neural Network Prediction Methods and Comparison

遥感影像预测是一种利用不同影像之间的时空及光谱特征进行影像变换的应用方法,能够弥补卫星影像数据的缺失,对于遥感应用领域的发展具有重要的意义。影像的色彩迁移及时序预测就是遥感影像预测包含的研究角度。基于影像色彩迁移的循环一致性生成对抗网络(CycleGAN)是一种不需要其他额外信息就能将一张影像从源域映射到目标域的方法,基于影像时序预测的卷积长短时记忆网络(ConvLSTM)是一种利用多时序影像的状态信息进行不同影像之间映射的方法。通过对比CycleGAN及ConvLSTM网络在影像预测中的结果,进一步分析色彩迁移网络及时序预测网络在影像预测中的适用性。实验使用无人机影像(UAV)数据、均方根误差(RMSE)及结构相似度(SSIM)等影像质量评价指标,得出如下结论:①使用CycleGAN及ConvLSTM网络的预测结果都能与参考影像保持一致的光谱特征,ConvLSTM网络在遥感影像预测方面的性能优于CycleGAN网络;②根据时序影像的缺失情况,选择不同的网络进行时序影像预测可以提高预测结果的精度和质量。当然,对于影响CycleGAN和ConvLSTM网络适用性的其他因素,将在未来的工作中进一步研究和分析。

Remote-sensing image prediction is an application method that uses spatio-temporal and spectral characteristics of different images to transform images.It can make up for the missing satellite image data and has great significance for research and development of remote-sensing field.The color migration and time-series prediction of images are the research perspective of the remote-sensing image prediction.The Cycle-Consistent Generative Adversarial Network(CycleGAN)based on color migration of image is a method that can map an image from a source domain to a target domain without additional information.The Convolution Long Short-term Memory Network(ConvLSTM)based on time-series prediction of image is a method that can perform mapping between different images by using the state information of multiple time-series images.The predicted results of CycleGAN and ConvLSTM network in image prediction are compared,and further the applicability of color migration network and time-series prediction network to image prediction is analyzed.The experimental results,which were performed with unmanned aerial vehicle(UAV)images,root mean square error(RMSE)and structural similarity(SSIM),are summarized as follows:①CycleGAN and ConvLSTM network can maintain the consistent spectral characteristics between the predicted result and the reference image,and the performance of ConvLSTM network in remote sensing image prediction is better than CycleGAN network.②According to different missing situations of time-series images,choosing different networks for time-series image prediction can improve accuracy and quality of the predicted result.For other factors affecting the applicability of CycleGAN and ConvLSTM network,it will be further researched and analyzed in the future work.