Abstract
Convolutional Neural Networks (CNN) possess many positive qualities when it comes to spatial raster data. Translation invariance enables CNNs to detect features regardless of their position in the scene. However, in some domains, like geospatial, not all locations are exactly equal. In this work, we propose localized convolutional neural networks that enable convolutional architectures to learn local features in addition to the global ones. We investigate their instantiations in the form of learnable inputs, local weights, and a more general form. They can be added to any convolutional layers, easily end-to-end trained, introduce minimal additional complexity, and let CNNs retain most of their benefits to the extent that they are needed. In this work we address spatio-temporal prediction: test the effectiveness of our methods on a synthetic benchmark dataset and tackle three real-world wind prediction datasets. For one of them, we propose a method to spatially order the unordered data. We compare the recent state-of-the-art spatio-temporal prediction models on the same data. Models that use convolutional layers can be and are extended with our localizations. In all these cases our extensions improve the results, and thus often the state-of-the-art.
Introduction
Standard Convolutional Neural Networks (CNNs) are designed with translation invariance as a core principle. By sharing weights across the entire spatial domain, CNNs can efficiently detect features like clouds or pressure fronts regardless of their specific coordinates. While this "location agnosticism" is a strength in image recognition, it presents a significant limitation in geospatial tasks such as wind forecasting.
In meteorology, geographical context is paramount. A sensor located on a coastal cliff will exhibit different wind patterns than one situated in a forested valley, even under identical regional weather conditions. Standard CNNs are structurally incapable of distinguishing these site-specific nuances. This paper introduces Localized CNNs, which augment traditional architectures with learnable, site-specific parameters to capture local "quirks" without sacrificing the global feature extraction capabilities of convolutions.
Methodology: Localizing the Convolutional Layer
The authors propose three primary architectural building blocks to introduce localization into CNNs. These modules can be integrated into existing state-of-the-art models with minimal overhead.
Learnable Inputs (LI)
The Learnable Inputs approach introduces static parameter maps, or latent variables, that are optimized during the training process. These maps are concatenated to the input data at each time step. By doing so, the network is provided with a persistent "local signature" for every grid cell. This allows subsequent convolutional filters to learn how to interpret global weather patterns in the context of specific local features.
Local Weights (LW)
Local Weights involve the use of locally connected layers where weights are not shared across the spatial grid. These act as site-specific filters that can adjust the importance of certain inputs or compensate for local biases, such as a sensor that consistently over-reports wind speed due to its physical placement.
Grid Embedding for Unordered Data
For datasets where sensors are not arranged in a regular grid (e.g., weather station networks), the authors developed a Mutual Information (MI) based grid embedding. This method employs an evolutionary algorithm to map irregularly spaced sensors onto a 2D grid based on data correlations, ensuring the spatial relationships are preserved in a way that is compatible with CNN architectures.
Experimental Setup and Datasets
The proposed methods were evaluated across four diverse datasets:
- Bouncing Ball (Synthetic): A proof-of-concept task where a ball moves within a bounded box. Localization is critical here because the physical behavior changes at the boundaries.
- WIND (NREL): Real-world wind speed data from a 10x10 grid of turbines in Indiana, USA.
- METAR: Hourly reports from 57 weather stations across the US East Coast, representing non-grid data.
- Copernicus: A spatially diverse European dataset covering a 10x10 region with significant distances between sites, emphasizing the need for broad local context.
Results and Discussion
The integration of localized features consistently improved the performance of baseline models across all real-world datasets.
- WIND Dataset: Localized models outperformed standard CNNs across all forecasting horizons (5 to 60 minutes). For a 10-minute forecast, the LI + LW combination achieved an RMSE of 0.4998, compared to the standard CNN's 0.5009.
- METAR Dataset: Utilizing the optimized grid embedding, the Persistent LI + LW model reached a Mean Absolute Error (MAE) of 1.106, significantly outperforming both the baseline CNN (1.119) and the previous state-of-the-art DL-STF (1.180).
- Copernicus Dataset: The localized version of the PSTN model achieved the best overall results with an RMSE of 1.420 for 6-hour forecasts, compared to 1.433 for the original PSTN.
A key insight from these results is that "breaking symmetry" is beneficial for geospatial physics. While weight sharing prevents overfitting, allowing for "local expertise" enables the model to account for non-uniformities in the physical environment. Furthermore, localized features were found to be particularly effective at "patching" errors in grid representation or sensor alignment.
Conclusion
Localized CNNs offer a powerful yet simple extension to standard convolutional architectures for geospatial applications. By allowing models to learn site-specific characteristics through Learnable Inputs and Local Weights, researchers can significantly improve forecasting accuracy. Future work will explore incorporating additional meteorological variables like wind direction and pressure, as well as investigating the generalization of these local features across different seasons.
Summary updated