Deep Random Features for Scalable Interpolation of Spatiotemporal Data
Authors: Weibin Chen, Azhir Mahmood, Michel Tsamados, So Takao
ICLR 2025 | Venue PDF | Archive PDF | Plain Text | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | We experiment on various remote sensing data at local/global scales, showing that our approach produce competitive or superior results to existing methods, with well-calibrated uncertainties. We evaluate the spatiotemporal deep random features (DRF) model on various remote sensing datasets and compare against various baselines to assess its ability to make predictions and quantify uncertainty. In our first experiment, we consider interpolation of synthetic data, and evaluate our model s ability to recover the ground truth. In our second and third experiments, we consider interpolation of real satellite data at local and global scales to test the robustness of our method. Details can be found in Appendix C. |
| Researcher Affiliation | Collaboration | 1University College London, UK 2Physics X, UK 3California Institute of Technology, USA |
| Pseudocode | No | The paper describes the methodology using mathematical equations and textual explanations, but it does not contain any clearly labeled pseudocode or algorithm blocks. |
| Open Source Code | Yes | CODE AVAILABILITY The code for reproducing our experiments is available in the following github repository https://github.com/totony4real/Deep Random Features. |
| Open Datasets | Yes | We use the freeboard observations along S3A, 3B and CS2 satellite tracks downloaded from Gregory et al. (2024a). Our final dataset contains 1,174,848 data points. In this final experiment, we used global sea level anomaly measurements from the Sentinel 3A satellite for the period March 1st 4th, 2020, resulting in 8,094,569 data points. |
| Dataset Splits | Yes | Our final dataset comprise 1,158,505 datapoints; we select 80% of these randomly for training and the remaining 20% for validation. We use 70% of the data (randomly sampled) for training, 15% for validation and the remaining 15% for testing. By considering four days of measurements, our final data consists of 8,094,569 datapoints. We use 80% for training, and 20% for validation. |
| Hardware Specification | Yes | All experiments are performed using the NVIDIA L4 GPU. |
| Software Dependencies | No | We use the Python library Bo Torch (Balandat et al., 2020) with the following configurations: ... The implementation is based on the Python package lightning-uq-box (Lehmann, 2024). Implemented using the Python package GPy Torch (Gardner et al., 2018). Implemented using the Python package GPy Torch with the spherical kernel implemented with the geometric-kernels package. |
| Experiment Setup | Yes | Optimiser: Adam Learning Rate: 0.001 Loss Function: Mean Squared Error (MSE) with L2 regularisation (equation 76) Number of Epochs: 1 Batch Size: 1024. For the DRF model used in this experiment, we use the following model configurations: Bottleneck size: B = 128 for both spatial and temporal layers. Hidden unit size: H = 1000 for both spatial and temporal layers. Number of spatial layers: Lx = 1, 2, 3, 4, 10, 20. For results in Table 1, we fix Lx = 4. Number of temporal layers: Lt = 1. |