Time series data is important for analyzing change for Earth observation and conservation efforts. Time series data involves capturing information about specific locations on Earth over the course of several months, years, and even decades. This information can then be used to track and manage changes in the environment.
Why time series data is used for Earth observation
This Earth observation could include changes in land use, deforestation rates, the expansion of urban areas, or the melting of ice caps and glaciers. By analyzing these changes, researchers can understand the pace and scale of environmental impact, which is essential for conservation efforts.
Time series data on temperatures, precipitation levels, sea ice extent, and carbon dioxide concentrations are vital for modeling climate change. These models can predict future climate conditions and help in planning adaptation and mitigation strategies for climate change. Other environmental metrics that are tracked over time are used to predict how climate change is driving natural disasters such as hurricanes, floods, wildfires, and droughts.
The health of ecosystems and biodiversity levels can also be monitored using time series data. This includes tracking changes in vegetation cover, water quality, and the presence or absence of certain species over time. Such data is critical in assessing the impact of human activities on natural habitats and in devising strategies to protect and preserve biodiversity.
Using time series data from different sources
Time series data is often collected by different research groups at different temporal scales and with differing spatial resolutions. Even Landsat, the longest Earth observation satellite imagery program, has collected data at different imagery resolutions as satellite remote sensing data collection sensors have improved.
Free weekly newsletter
Fill out your e-mail address to receive our newsletter!
By entering your email address you agree to receive our newsletter and agree with our privacy policy.
You may unsubscribe at any time.
In order to produce meaningful tracking of changes over time, researchers must often create customized models that factor in the the differences between datasets. This is often a very time-consuming practice. I recently attended a presentation at Stanford University on artificial intelligence and Earth observation by Will Marshall, the CEO of Planet. In describing one project that looked at damaged buildings in Ukraine, Marshall noted that the model used to track changes took months to develop as building styles are different around the world.
A versatile time series model
A collaboration between researchers from Harvard University, MIT Lincoln Laboratory, and the University of Virginia have recently submitted a preprint paper to arXiv about a new model called UniTS. UniTS stands for Unified Time Series model and it was developed as a versatile tool that can handle a wide range of time series tasks without the need for time-consuming and individualized adjustments. The goal of UniTS is be accepted as a tool to simplify the the modeling process for time series analysis that is dependent on disparate data sets.
UniTS’ performance was analyzed using 38 different sets of data from various areas and the researchers found that it outperformed existing task-specific and natural language-based models. The researchers found that the particular strengths of UniTS was in making predictions, categorizing data, filling in missing information, and spotting odd data points. In one test, it improved the accuracy of predictions by 10.5% compared to the best previous model, highlighting its precision in forecasting.
The authors of the paper hope that UniTS will enhance researchers’ capacity to streamline the analysis of temporal patterns from a collection of datasets, whether it’s for financial forecasting, healthcare diagnostics, or environmental conservation.
References
Gao, S., Koker, T., Queen, O., Hartvigsen, T., Tsiligkaridis, T., & Zitnik, M. (2024). UniTS: Building a Unified Time Series Model. arXiv preprint arXiv:2403.00131. DOI: arXiv:2403.00131.
