PSIque: Next Sequence Prediction of Satellite Images using a Convolutional Sequence-to-Sequence Network

Seungkyun Hong, Seongchan Kim, Minsu Joh, Sa-Kwang SongarXiv 2017 · arXiv

Abstract

Continuous satellite imagery is vital for weather forecasting and disaster management, yet data gaps frequently occur due to sensor failures or celestial events. This paper introduces PSIque, a convolutional sequence-to-sequence (Seq2Seq) autoencoder designed to predict future or missing satellite weather images. By integrating Convolutional LSTMs with symmetric skip connections, PSIque generates high-dimensional visual forecasts from historical data, providing a robust tool for "nowcasting" and data recovery.

Introduction

Modern meteorological services rely on a steady stream of satellite data. However, remote sensing systems are susceptible to interruptions from image merging issues, satellite repositioning, or solar/lunar eclipses. While traditional Numerical Weather Prediction (NWP) models are powerful, their high computational cost and reliance on massive datasets make them less suitable for rapid, responsive nowcasting. PSIque addresses this by treating weather forecasting as a video prediction task, generating plausible future satellite images based solely on historical visual time-series data.

Methodology

The core of PSIque is a Sequence-to-Sequence (Seq2Seq) autoencoder framework. Standard Seq2Seq models often use 1D LSTMs, which require flattening images and thus destroying critical spatial relationships. PSIque overcomes this by utilizing Convolutional LSTMs (ConvLSTMs) in both the encoder and decoder.

Architecture of the PSIque model featuring ConvLSTM cells and selective symmetric skip connections.

To preserve fine-grained visual details often lost in the autoencoder's bottleneck, the authors introduced Symmetric Skip Connections (SkipConx). These auxiliary links connect convolutional layers in the encoder directly to corresponding deconvolutional layers in the decoder. The research explored various topologies, finding that linking multiple layers (specifically layers 1, 3, and 5) optimized information propagation for both low-level features (edges) and high-level structures (storm patterns).

Experimental Setup

The model was trained and evaluated using the COMS-1 Satellite Meteorological Imager (MI) Dataset, comprising four-channel infrared images from 2011 to 2017. Images were resized to 256 × 256 × 4 pixels. The dataset was structured into approximately 26,490 sequence instances, where the model consumes 4 hours of historical data to predict the subsequent 2 hours of weather transitions.

Results and Discussion

The experimental results highlight two critical factors for successful satellite image prediction: spatial awareness and memory bypass.

Quantitative Performance

The transition from standard fully-connected LSTMs (C2L2D) to ConvLSTMs (C2CL2D) resulted in a massive performance leap, with Mean Squared Error (MSE) dropping from 7537.9 to 215.0. The addition of symmetric skip connections further refined the output, with the L1/3/5 configuration achieving the best overall performance (MSE 197.2).

Qualitative Analysis

Visually, the baseline C2L2D model produced unusable noise. In contrast, the PSIque model with L1/3/5 skip connections generated coherent, realistic cloud cover patterns that closely mirrored ground truth observations.

Qualitative comparison of predicted satellite images across different model configurations against the ground truth observation.

The use of multi-level connections proved essential; by passing features from multiple depths, the decoder could reconstruct both the broad atmospheric structures and the sharper details necessary for a plausible forecast.

Conclusion and Future Work

PSIque demonstrates that convolutional Seq2Seq architectures with skip connections can effectively predict complex meteorological transitions. While the current results are promising, the authors identify image resolution as a primary area for improvement. Future research will explore the integration of differentiable memory mechanisms and Generative Adversarial Networks (GANs) to produce sharper, high-resolution outputs suitable for granular weather situation prediction.

Summary updated

BibTeX
@misc{hong2017psique,
  author = {Seungkyun Hong and Seongchan Kim and Minsu Joh and Sa-Kwang Song},
  title = {{PSIque: Next Sequence Prediction of Satellite Images using a Convolutional Sequence-to-Sequence Network}},
  year = {2017},
  eprint = {1711.10644},
  archivePrefix = {arXiv},
  url = {https://arxiv.org/abs/1711.10644},
}
Figure