Position: Mission Critical – Satellite Data is a Distinct Modality in Machine Learning
Authors: Esther Rolf, Konstantin Klemmer, Caleb Robinson, Hannah Kerner
ICML 2024 | Conference PDF | Archive PDF | Plain Text | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Theoretical | This position paper argues that satellite data constitutes a distinct modality for machine learning research and that we must recognize it as such to advance the quality and impact of Sat ML research across theory, methods, and deployment. We outline critical discussion questions and actionable suggestions to transform Sat ML from merely an intriguing application area to a dedicated research discipline that helps move the needle on big challenges for machine learning and society. |
| Researcher Affiliation | Collaboration | 1Harvard Data Science Initiative and Center for Research on Computation and Society, Harvard University 2University of Colorado, Boulder 3Microsoft Research 4Microsoft AI for Good Research Lab 5School of Computing and Augmented Intelligence, Arizona State University. |
| Pseudocode | No | The paper is a position paper and does not contain any pseudocode or algorithm blocks. |
| Open Source Code | No | The paper is a position paper and does not describe a new methodology or provide a link to open-source code for its own content. |
| Open Datasets | Yes | For example, the text dataset of snapshots from the Common Crawl database used to train GPT-3 was 45 TB (and 570 GB after filtering) (Brown et al., 2020). LAOIN-5B, the largest paired text-image dataset today, is 220 TB (Schuhmann et al., 2022). The ILSVRC 2012 Image Net dataset (Deng et al., 2009), widely used for vision model benchmarking and pre-training, is 150 GB. |
| Dataset Splits | No | The paper discusses 'Spatially aware holdout and cross-validation methods' as a general concept in Section 2.3 for evaluating Sat ML models ('Spatially aware holdout and cross-validation methods have been designed to test how models might perform outside the regions in which they were trained. These include blocking or buffering the distance between a train and test set (Figure 3) or parametrically varying the distance between train and test data (Roberts et al., 2017; Le Rest et al., 2014; Pohjankukka et al., 2017; Airola et al., 2019; Wang et al., 2023a)'). However, since this is a position paper and does not present its own experiments, it does not specify concrete training/test/validation dataset splits used for its own work. |
| Hardware Specification | No | The paper discusses the need for efficient processing of large data volumes in Sat ML and general hardware implications, but it does not specify any hardware (e.g., GPU/CPU models, memory) used by the authors for experiments, as the paper is a position paper and does not conduct original experiments. |
| Software Dependencies | No | The paper discusses general software limitations for Sat ML (e.g., 'common ML libraries lack support for many-channel satellite images. For example, Torch Vision and other libraries with pretrained models assume images are 3-channel'), but it does not list specific software dependencies with version numbers used for the paper itself, as it is a position paper. |
| Experiment Setup | No | The paper is a position paper and does not describe an experimental setup with hyperparameter values or training configurations for its own work. |