Here we present a Python tutorial for manipulating the time series of satellite images, analysing variations and detecting possible anomalies. To do this, we take the example of forest diebacks due to the proliferation of bark beetles in recent years in a large part of Europe.
keywords: remote sensing, time-series, sentinel, dieback, anomalies detection
The tutorial is divided into 5 steps:
- 01_vi_datacube.ipynb: we propose here to search through a STAC catalog the available images according to several criteria (start date, end date, cloudiness, etc.), calculate some vegetation indices (VI), apply mask and export the datacube as file.
- 02_vi_index.ipynb: from the VI time-series, we experiment a gap filling technique to remove the masked pixels and estimate a VI value instead. Then we present different ways for displaying satellite datacubes.
- 03_vi_time-series_plot.ipynb: we analyze the VI time-series with the NRT package. The objective is to fully understand how this tool works, by visualizing through plots the time-series values, the modeling by harmonious function and the more or less early detections depending on the methods used.
- 04_optimized_params.ipynb: we propose here to optimize the parameters to minimize the difference between the reference date and the detection date.
- 05_produce_anomaly_map.ipynb: we use the NRT package to produce an image of the diebacks spread.
The tutorial uses the time-series acquired by the Sentinel-2 constellation (Copernicus programme). For practical reasons, but also to learn how to manipulate datacubes, we will be using the free Microsoft Planetary Computer Cloud service. This service is based on recent standards (STAC, COG) that enable spatial data to be discovered and processed effectively online.
However, depending on your needs, you can change provider as long as it complies with the STAC standards.
The tutorial provides three vector files, so you can experiment the NRT package by masking non-forest areas and assessing the results with a few reference data.
bd-foret_p33.gpkg: limits of the studied area, a plot of coniferous forest in the North of France. Taken from BD Foret V2 ©IGNplot01_samples.shp: set of 14 reference points visually interpreted on Sentinel-2 time-series. It gives the start date of dieback (due to bark beetles) at the yellow or red stage (approximation). Produced by ©Kenji Ose (2024).plot01_samples_2.shp: set of 50 reference points visually interpreted on Sentinel-2 time-series. It gives the start date of dieback (due to bark beetles) at the yellow or red stage (approximation). Produced by ©Kenji Ose (2024).
The tutorial uses a classic time-series analysis approach. This involves two main stages:
- a fitting stage in which, part of the univariate (i.e. spectral index) time series, considered as "normal", is used to model a harmonic function. For our example, it describes the seasonal variations in a vegetation index for a healthy forest.
- a monitoring stage that consists of comparing the predictions of the harmonic function with the actual values in order to detect any anomalies (in this case diebacks).
There are many Python packages for analysing time-series, but relatively few for easily exploiting satellite data. Here we present two interesting examples:
-
the Fordead package, developed by INRAE (TETIS laboratory): this tool dedicated to the analysis of forest dieback based on time-series of Sentinel-2 images has proven its effectiveness. The Health Forest Department of the French Ministry of Agriculture has used it for several years to produce dieback monitoring maps linked to the bark beetle health crisis in France. As this tool is very well documented, we recommend you to consult the dedicated web pages directly: https://fordead.gitlab.io/fordead_package/
-
the NRT package, developed by the Joint Research Center (JRC) D.1. Unit: this tool is intended to be standard in order to gradually integrate different anomaly detection methods on time-series of satellite images. Therefore, unlike Fordead, the NRT package is not dedicated to forestry issues and can possibly be tested on other types of land cover. Detailed documentation is available at the following address: https://nrt.readthedocs.io/en/latest/index.html
All of the exercises presented here are based on the NRT package. Some modules have been added to improve the ergonomics and understanding of this tool.
The tutorial is designed in the form of Jupyter notebook files. You must first install the Python NRT package (follow the documentation for this tool). Depending on the environment chosen, you may need to complete with some packages dedicated to the analysis of spatial data.
The tutorials are written by Kenji Ose.
For any help, comment, etc., please use the issue tracker.
GNU GPLv3