Innovative on-demand land cover services for customised spatial and temporal needs

Innovative on-demand land cover services for customised spatial and temporal needs

3 Oct 2024

The wide range of land cover (LC) products needed by users can best be met by processing systems able to accommodate user-defined tailoring via user-provided parameters – e.g., the number, types and definition of classes, area of interest, temporal needs – and then generate on-demand the desired customised product. This candidate will demonstrate such an on-demand LC service by showcasing a number of tailored LC products with customised classes (e.g., related to tree genus, cover crop type, and forest management) at different spatial resolutions (100m and 10m).

A wide variety of ground-truth datasets for land cover classification exist in private and public data repositories and are used to train land cover classification models. Oftentimes, these datasets are subjected to privacy regulations or users would like to be able to use these datasets within their own applications without sharing the data. This is for instance relevant for national forest inventories (NFI data), that contain highly valuable information for training and validating remote sensing based models on forest composition and structure. Yet, the coordinates of the NFI data often cannot be shared publicly to ensure the data remains representative.

Furthermore, building and training a land cover classification model is a complex process requiring specialised expertise, which many users may lack. Training a model and performing inference on satellite data also involves handling large datasets, presenting additional challenges for users with limited network resources or infrastructure.

A cloud-based pipeline for training and inference of classification models

This prototype addresses the challenges above by demonstrating a cloud-based model training and inference pipeline that allows users to upload reference training data, extract training features from the Copernicus Data Space Ecosystem (CDSE) backend, train a machine learning classification model, and choose the area, time period, and spatial resolution of the final classification results. This will create a classification map tailored to the user’s needs and requirements, while ensuring the privacy of their reference data. Additionally, the prototype anticipates further customisation options, such as model selection and parameter adjustments, providing users with greater control over the final classification outcomes.

Training
Training data Extract remote sensing data over your training locations in the cloud while maintaining privacy of the data
Model Train your classification model using with custom training labels. Here, customisation of model selection and parametrisation is foreseen
Inference
Area of interest Use the trained model to compute a classification map over your preferred area of interest, time period, and spatial resolution in the cloud and download the resulting map
Time period
Spatial resolution

This versatile cloud-based pipeline for the training and inference of classification models has been implemented using openEO. OpenEO is a platform that offers standardised interfaces for easy access to and processing of Earth Observation data across various infrastructures, including the Copernicus Data Space Ecosystem. Specifically, the model and inference pipeline has been built using the General Framework for Mapping (GFMAP) package that standardises and simplifies the usage of many common processes and functions used within openEO.

To test and showcase the workflow, we applied the pipeline to map forest composition (dominant tree genus), by training Random Forest and Catboost models using a set of annual Sentinel-1 and Sentinel-2 statistics, DEM, and climate data. In the next phase, we plan to evaluate and extend the usability of the prototype, focusing on a broader range of applications beyond tree genus classification. Specifically, the next phase will target forest management mapping at 10m and 100m, as well as cover crop type mapping. Furthermore, we will further improve the cost and processing efficiency of the prototype while expanding its flexibility to meet various classification needs by offering more model parameterisation options.

Workflow for prototype C11, including user-interactive inputs. In future phases of the project, local implementations will be adapted to a fully cloud-based (openEO) workflow.

In conclusion, this prototype focuses on developing and testing a cloud-based classification pipeline, covering key components such as training data preparation, feature extraction, model training and inference within openEO, enabling on-demand land cover mapping.

Cover image credit: European Union, Copernicus Sentinel-2 imagery

Would you be interested to follow the results of EvoLand and the development of this candidate prototype? Sign up for our newsletter (here below or via this form), follow us on LinkedInMastodon or X (Twitter). 

Questions about the article or the project? E-mail us at contact@evo-land.eu 

== 

This article is part of a series providing more details on all 11 EvoLand candidate prototypes.

Previous article in the series: 

Improving GPP estimates over cropland and grassland by using the Sentinel-2 – EvoLand (evo-land.eu)

 

Newsletter

Stay current! Subscribe to our EvoLand newsletter!