We present a deep learning architecture that reconstructs a source of data at given spatio-temporal coordinates using other sources. The model can be applied to multiple sources in a broad sense: the number of sources may vary between samples, the sources can differ in dimensionality and sizes, and cover distinct geographical areas at irregular time intervals. The network takes as input a set of sources that each include values (e.g., the pixels for two-dimensional sources), spatio-temporal coordinates, and source characteristics. The model is based on the Vision Transformer, but separately embeds the values and coordinates and uses the embedded coordinates as relative positional embedding in the computation of the attention. To limit the cost of computing the attention between many sources, we employ a multi-source factorized attention mechanism, introducing an anchor-points-based cross-source attention block. We name the architecture MoTiF (multi-source transformer via factorized attention). We present a self-supervised setting to train the network, in which one source chosen randomly is masked and the model is tasked to reconstruct it from the other sources. We test this self-supervised task on tropical cyclone (TC) remote-sensing images, ERA5 states, and best-track data. We show that the model is able to perform TC ERA5 fields and wind intensity forecasting from multiple sources, and that using more sources leads to an improvement in forecasting accuracy.

Visible satellite imagery (VIS) is essential for monitoring weather patterns and tracking ground surface changes associated with climate change. However, its availability is limited during nighttime. To address this limitation, we present a discrete variational autoencoder (VQVAE) method for translating infrared satellite imagery to VIS. This method departs from previous efforts that utilize a U-Net architecture. By removing the connections between corresponding layers of the encoder and decoder, the model learns a discrete and rich codebook of latent priors for the translation task. We train and test our model on mesoscale data from the Geostationary Operational Environmental Satellite (GOES) West Advanced Baseline Imager (ABI) sensor, spanning 4 years (2019 to 2022) using the Conditional Generative Adversarial Nets (CGAN) framework. This work demonstrates the practical use of a VQVAE for meteorological satellite image translation. Our approach provides a modular framework for data compression and reconstruction, with a latent representation space specifically designed for handling meteorological satellite imagery.

Good air quality is a critical determinant of public health, influencing life expectancy, respiratory health, work productivity, and the prevention of chronic diseases. This study presents a novel approach to classifying the Air Quality Index (AQI) using deep learning techniques, specifically convolutional neural networks (CNNs). We collected and curated a dataset comprising 11,000 digital images from three distinct regions in Indonesia—Jakarta, Malang, and Semarang—ensuring uniformity through standardized acquisition settings. The images were categorized into four air quality classes: good, moderate, unhealthy for sensitive groups, and unhealthy. We designed and implemented a CNN architecture optimized for AQI classification. The model achieved an impressive accuracy of 99.81% using K-fold cross-validation. In addition, the model’s interpretative capabilities were examined using techniques such as Grad-CAM, providing valuable insights into how the CNN identifies and classifies air quality conditions based on image features. These findings underscore the effectiveness of CNNs for AQI classification and highlight the potential for future work to incorporate a more diverse set of digital images captured from various perspectives to enhance dataset complexity and model robustness. The dataset is publicly accessible at https://doi.org/10.5281/zenodo.15727522.