Open Earth Monitor — Global Workshop 2023

Accurate mapping and monitoring of land dynamics is critical for climate change mitigation, biodiversity conservation, and epidemic prevention. With increasing data availability and processing capacities, there is a growing desire to move from periodic mapping of forest disturbances to continuous monitoring systems capable of providing timely information on forest disturbances to a variety of stakeholders. Many algorithms and approaches have been proposed in the research community to address this near real-time monitoring challenge. Their performance is typically demonstrated based on case studies over test areas or simulated datasets. However, when it is available, the software provided with the research papers only offers limited operational capacity. Individual software is often primarily developed to support the research experiment and consequently not optimized for speed or deployment at scale. In addition, implementation in different programming languages or the absence of a common interface to operate the algorithm make interoperability and comparisons exercises challenging. Inspired by the great success of scikit-learn that provides a standard interface to a large pool of optimized models and is widely acknowledged as the de-facto standard for machine learning in python, we developed the nrt python package. The package is designed for near real-time monitoring of disturbances in satellite image time-series. Five monitoring algorithms from the scientific literature on change detection (EWMA, CuSum, MoSum, CCDC, IQR) are implemented and exposed via a common API (Application Programming Interface). All the provided algorithms are optimized for fast computation thanks to the use of vectorized expressions or numba's JIT (Just In Time) compiler. Additionally, ongoing monitoring instances can be saved to disk and reloaded anytime, allowing for effortless operational deployment. The presentation will detail the functionalities of the package, the characteristics of the implemented algorithms, and illustrate its potential via a regional deployment covering several Sentinel2 tiles. The supporting ecosystem of tools, composed of dashboards for interactive use, parameters selection, and exploration of generated alert will also be presented.