What question did this study set out to answer?

This research aims to establish participatory workflows for developing national-scale machine learning classification models using open-source tools.

May 3, 2026Open Access

Free and open-source machine learning workflows for co-creating national-scale classification models through country-driven QField surveys and Digital Earth Pacific

NMN. Metherall NMN. Metherall NMN. Metherall

Key Points

This research aims to establish participatory workflows for developing national-scale machine learning classification models using open-source tools.
Conducted a participatory workshop in Tonga in July 2023 with representatives from seven Ministries.
Utilized FOSS tools like QGIS, QField, and Scikit-Learn within the DE Pacific Analytical Hub for model development.
Engaged over 240 participants across 10 Pacific Island Countries for training and data collection.
Successful generation of land use land cover models, with detailed accuracy assessments provided.
Positive evaluations on the participatory mapping processes from the workshop.
Demonstrated capacity building and ongoing learning processes within the Pacific region.

Abstract

This paper highlights the Free and Open-Source Software (FOSS) datasets, analytical tools and workflows used for co-creating localised national-scale machine learning classification models in Digital Earth Pacific (DE Pacific). The case study includes the participatory workflows used within the DE Pacific Land Cover Assessment Skills Transfer (LCAST) Workshop in the Kingdom of Tonga in 2023. The FOSS tools used were QGIS, QField, GeoPandas, ODC STAC-, Xarray, Pandas, Scikit-Learn, NumPy and Folium through the DE Pacific Analytical Hub Jupyter environment. These workflows supported participatory processes to gather inputs into the calibration and validation of machine learning workflows as seen in the Land Use Land Cover (LULC) examples of the LCAST workshop. The workshop held over one week in July 2023 included representatives from seven Ministries of Tonga. The methods covered the ‘full-cycle’ of workflows for generation of LULC models, from field survey data collection to computer labs for data processing and analyses. The results highlight the LULC mapping products, accuracy assessments, the workshop evaluation as well as the broader lessons learned about the intrinsic value of the participatory mapping processes. Since 2023, these workflows have been used in LCAST and similar country-driven workshops with more than 240 participants across 10 Pacific Island Countries and Territories (PICTs) or Large Ocean States (LOS) including Fiji, Republic of the Marshall Islands, Tuvalu, Palau, Cook Islands, Papua New Guinea, Vanuatu, New Caledonia, the Solomon Islands and the Kingdom of Tonga. These FOSS approaches may contribute to the long-term continuity of two-way learning processes and inputs needed for the co-creation, calibration and validation for modelling and mapping as well as the building of local capacity and capabilities in earth observations in the Pacific. The paper outlines these workflows in detail through a visual flowchart and provides access to GitHub for replication of results.

Ask AI

Helpful

Bookmark

View Full Paper