Search

In this study, the idea of using a fully symmetric U-Net deep learning model for forecasting a segmented image of high global aerosol concentrations is implemented. As the forecast relies on historical data, the model used a sequence of the last eight segmented images to make the prediction. For this, the classic U-Net model was modified to use ConvLSTM2D layers with MaxPooling3D and UpSampling3D layers. In order to achieve complete symmetry, the output data are given in the form of a series of eight segmented images shifted by one image in the time sequence so that the last image actually represents the forecast of the next image of high aerosol concentrations. The proposed model structure was tuned by the new FOX metaheuristic algorithm. Based on our analysis, we found that this algorithm is suitable for tuning deep learning models considering their stochastic nature. It was also found that this algorithm spends the most time in areas close to the optimal value where there is a weaker linear correlation with the required metric and vice versa. Taking into account the characteristics of the used database, we concluded that the model is capable of generating adequate data and finding patterns in the time domain based on the ddc and dtc criteria. By comparing the achieved results of this model using the AUC-PR metric with the previous results of the ResNet3D-101 model with transfer learning, we concluded that the proposed symmetric U-Net model generates data better and is more capable of finding patterns in the time domain.

Introduction/purpose: The paper presents a new state-of-the-art method that involves NASA satellite imagery with the latest deep learning model for a spatiotemporal sequence forecasting problem. Satellite-retrieved aerosol information is very useful in many fields such as PM prediction or COVID-19 transmission. The input data set was MODAL2_E_AER_OD which presents global AOT for every 8 days from Terra/MODIS. The implemented machine learning algorithm was built with ConvLSTM2D layers in Keras. The obtained results were compared with the new CNN LSTM model. Methods: Computational methods of Machine Learning, Artificial Neural Networks, Deep Learning. Results: The results show global AOT prediction obtained using satellite digital imagery as an input. Conclusion: The results show that the ConvLSTM developed model could be used for global AOT prediction, as well as for PM and COVID-19 transmission.

In order to better predict the high aerosol concentrations associated with air pollution and climate change, a machine learning model was developed using transfer learning and the segmentation process of global satellite images. The main concept of transfer learning lies on convolutional neural networks and works by initializing the already trained model weights to better adapt the weights when the network is trained on a different dataset. The transfer learning technique was tested with the ResNet3D-101 model pre-trained from a 2D ImageNet dataset. This model has performed well for contrail detection to assess climate impact. Aerosol distributions can be monitored via satellite remote sensing. Satellites can monitor some aerosol optical properties like aerosol optical thickness. Aerosol optical thickness snapshots were the input dataset for the model and were obtained from NASA’s Terra-Modis satellite; the output images were segmented by comparing the pixel values with a threshold value of 0.8 for aerosol optical thickness. Hyperparameter optimization finds a tuple of hyperparameters that yields an optimal model that minimizes a predefined loss function on given independent data. The model structure was adjusted in order to improve the performance of the model by applying methods and hyperparameter optimization techniques such as grid search, batch size, threshold, and input length. According to the criteria defined by the authors, the distance domain criterion and time domain criterion, the developed model is capable of generating adequate data and finding patterns in the time domain. As observed from the comparison of relative coefficients for the criteria metrics proposed by the authors, ddc and dtc, the deep learning model based on ConvLSTM layers developed in our previous studies has better performance than the model developed in this study with transfer learning.

The attention mechanism in natural language processing and self-attention mechanism in vision transformers improved many deep learning models. An implementation of the self-attention mechanism with the previously developed ConvLSTM sequence-to-one model was done in order to make a comparative evaluation with statistical testing. First, the new ConvLSTM sequence-to-one model with a self-attention mechanism was developed and then the self-attention layer was removed in order to make comparison. The hyperparameters optimization process was conducted by grid search for integer and string type parameters, and with particle swarm optimization for float type parameters. A cross validation technique was used for better evaluating models with a predefined ratio of train-validation-test subsets. Both models with and without a self-attention layer passed defined evaluation criteria that means that models are able to generate the image of the global aerosol thickness and able to find patterns for changes in the time domain. The model obtained by an ablation study on the self-attention layer achieved better outcomes for Root Mean Square Error and Euclidean Distance in regards to developed ConvLSTM-SA model. As part of the statistical test, a Kruskal–Wallis H Test was done since it was determined that the data did not belong to the normal distribution and the obtained results showed that both models, with and without the SA layer, predict similar images with patterns at the pixel level to the original dataset. However, the model without the SA layer was more similar to the original dataset especially in the time domain at the pixel level. Based on the comparative evaluation with statistical testing, it was concluded that the developed ConvLSTM-SA model better predicts without an SA layer.

Mathematical methods are the basis of most models that describe the natural phenomena around us. However, the well-known conventional mathematical models for atmospheric modeling have some limitations. Machine learning with Big Data is also based on mathematics but offers a new approach for modeling. There are two methodologies to develop deep learning models for spatio-temporal image prediction. On these bases, two models were built—ConvLSTM and CNN-LSTM—with two types of predictions, i.e., sequence-to-sequence and sequence-to-one, in order to forecast Aerosol Optical Thickness sequences. The input dataset for training was NASA satellite imagery MODAL2_E_AER_OD from Terra/MODIS satellites, which presents global Aerosol Optical Thickness with an 8 day temporal resolution from 2000 to the present. The obtained results show that the ConvLSTM sequence-to-one model had the lowest RMSE error and the highest Cosine Similarity value. The advantages of the developed DL models are that they can be executed in milliseconds on a PC, can be used for global-scale Earth observations, and can serve as tracers to study how the Earth’s atmosphere moves. The developed models can be used as transfer learning for similar image time-series forecasting models.