Remote Sensing area classification based on AI image recognition

Province of South Holland

Based on image recognition of satellite imagery and analysis, changes in natura2000 areas are understood and mapped in detail. We do this for ecological purposes.

Last change on 22nd of November 2024, at 11:34 (CET) | Publication Standard 1.0

• Publication category: Other algorithms
• Impact assessment: Field not filled in.
• Status: In use

General information

Theme

Nature and Environment

Begin date

Field not filled in.

Contact information

Voor vragen en opmerkingen kan je terecht bij: [e-mail]. Wil je bezwaar maken, dan kan je terecht bij de Juridische afdeling van dienst beheer organisatie. https://www.zuid-holland.nl/contact/

Link to publication website

https://github.com/Provincie-Zuid-Holland/satellite_images_nso_extractor https://github.com/Provincie-Zuid-Holland/satellite-images-nso-datascience https://github.com/Provincie-Zuid-Holland/satellite_images_nso_tif_model_iterator https://github.com/Provincie-Zuid-Holland/vdwh_ahn_processing

Link to source registration

NSO, Scikit

Responsible use

Goal and impact

By using smart algorithms, we can map and analyse changes in a natural area. This allows us to predict the effect on the environment and take action. Using image recognition on satellite images and analysis, changes in Natura 2000 areas, a European network of protected natural areas, are visualised and mapped in detail. This is done to support ecological purposes, such as nitrogen policy monitoring.

The province of South Holland is monitored using various sensors (satellite images, LIDAR, IR, multispectral, microwave, etc.) from satellites, aircraft and helicopters. Much of this data is available for free. Exploiting such area-wide measurement series can give a big boost to monitoring natural areas.

Interpreting remote sensing data into ecologically relevant insights can be an important building block for a digital twin of nature. These measurement sets can serve as the 'skeleton' or framework on which the digital twin can be further built. Drone imagery, combined with species recognition, also provides a means of obtaining highly detailed data. This increases coverage and it allows for better updating within the province.

Habitat biodiversity is lower than desired. At European level, it is stipulated that government agencies have the task of improving this in Natura 2000 areas. Using image recognition, changes in these areas are visualised and mapped in detail. This allows plant species to be recognised and monitored, for example to identify the spread of invasive species and take remedial measures against nitrogen. Within nature reserves, certain plant species are sensitive to nitrogen. Excess nitrogen causes these plant species to be displaced by other, less desirable species, such as nettle grasses. This process, such as grassing, can have negative consequences for the biodiversity of dune areas, for example. The resolution of the images is 50 to 30 centimetres of raw data from NSO, with aggregations up to 3-4 metres for ecological applications. There is no impact on people as they are not recognisable on the satellite images (subject to privacy check NSO).

Considerations

Open and clear communication on the use of satellite imagery is essential for resident acceptance and confidence in this technology. Use is for nature management and nitrogen policy purposes only.

Human intervention

Yes by ecologists.

Risk management

The model is 90% watertight. Human intervention (ecologist) is needed, to check things. It is also good to look at drone images additionally. The accuracy is 90% of image recognition with F1 score.

Operations

Data

This dataset contains snapshots of automatically generated vegetation structure classifications of the Natura 2000 sites, based on SuperView and Pleiades Neo satellite images from NSO. The images are made available every one to four months and segmented into vegetation structure classes by a model. The analysed images go back to 2019.

The data are visualised in an ArcGIS Operations Dashboard, which allows comparison of all segmented snapshots.

Links to data sources

• NSO: https://www.spaceoffice.nl/nl/
• Scikit: https://scikit-learn.org/

Technical design

Scikit learn: package models open source within Python.

Model: Pixed-based random forest for recognising vegetation structures.