Co-Production of a 10-m Cropland Extent Map for Continental Africa using Sentinel-2, Cloud Computing, and the Open-Data-Cube

A central focus for governing bodies in Africa is the need to secure the necessary food sources to support their populations. It has been estimated that the current production of crops will need to double by 2050 to meet future needs for food production. Higher level crop-based products that can assist with managing food insecurity, such as cropping watering intensities, crop types, or crop productivity, require as a starting point precise and accurate cropland extent maps indicating where cropland occurs. Current continental cropland extent maps of Africa are either inaccurate, have too coarse spatial resolutions, or are not updated regularly. An accurate, high-resolution, and regularly updated cropland extent map for the African continent is therefore recognized as a gap in the current crop monitoring services. Using Digital Earth Africa’s Open Data Cube platform, and working in conjunction with multiple regional African geospatial institutions, we co-develop a 10 metre resolution cropland extent map over the African continent using a Random Forest machine learning classiﬁer and an annual time-series of Sentinel-2 satellite images. Members of the regional African geospatial institutions

Digital Earth Africa (https://www.digitalearthafrica.org/) (DE Africa) operates a digital infrastructure in Africa that aims to provide free and open access to Earth observation (EO) data and services to all, and build capacity across Africa to use EO based insights to address sustainable development challenges.
The 10 metre resolution cropland extent map is DE Africa's first continental service for agriculture and is expected to serve as a basis for higher level crop monitoring and management products.We have co-developed this service with our partners across the continent.Members of the regional African geospatial institutions (RCMRD, OSS, Afrigist, AGRHYMET, SANSA, and NADMO) were instrumental in defining the specifications of the product, in developing and implementing a continental scale reference data collection strategy, and assisting with iterative model building.
We have classified cropland using an annual Sentinel-2 time series and a Random Forest machine learning model.The cropland extent product is among the highest resolution and highest accuray products of its type for the continent of Africa.The product comes packaged with three layers: a pixel-based classification, a pixel-based cropland probability layer, and an object-based segmentation filtered classification.All the components of the service: models, reference data, code, and results are open source and freely available online through Digital Earth Africa's mapping and analysis platforms.
A full description of the dataset, including details on how it was made, the validation results, and how to access the different datasets can be seen on the deafrica user guide.(https://docs.digitalearthafrica.org/en/latest/data_specs/Cropland_extent_specs.html) The cropland extent product is in the final stages of development, with southern and central Africa still to complete.The expected completion date is early 2022.The completed regions are available on all of Digital Earth Africa's platforms (see the section below for links and details).11/01/2022, 10:43 am DE Africa has a number of freely available platforms for accessing and analysing our datasets.

Exploring datasets
The video embedded below shows an example of exploring the product using DE Africa's mapping portal.
A direct link with some of the cropland extent layers pre-loaded can be accessed by following this link.(https://maps.digitalearth.africa/#share=s-y3JyeK8agnTdwegsSLhlYIhdrqw)
Sample polygons for each AEZ were assessed using Collect Earth Online (CEO) tool developed by NASA SERVIR.Analysts used image interpretation to classify each polygon as either 'crop', 'non-crop', 'mixed', or 'unsure'; where the 'crop' and 'non-crop' labels were only selected if the sample region is homogeneously crop or non-crop.Monthly Sentinel-2 mosaics, a two year NDVI time-series (2018-06 to 2020-06), and Bing 'Aerial' basemaps were available to assist the image interpretation.The video below shows an example of this procedure using the Collect Earth Online Tool [VIDEO] https://res.cloudinary.com/amuze-interactive/video/upload/vc_auto/v1638239626/agu-fm2021/BA-E4-89-84-15-50-25-89-AE-A8-EF-C5-32-7D-19-EE/Video/Collect_Earth_Online_-_30_November_2021_pbdscc.mp4 Cropland in the reference data is defined as "a piece of land of minimum 0.01 ha that is sowed/planted and harvestable at least once within 12 months after the sowing/planting date".
This definition excludes non-planted grazing lands and perennial crops which can be difficult for satellite imagery to differentiate from natural vegetation.
In addition to the samples collected through CEO, additional samples were digitized through a GIS application in locations where the classifier did poorly.This process was done in an iterative fashion, incrementally improving classifications through targeted training data collection.
In total, >24,000 training data samples were collected, with an additional ~1,800 samples isolated as an independent validation dataset (see the figure in the slideshow in the bottom right of the poster).
The accuracy of the method for collecting reference samples described above was independently evaluated by Radiant Earth (https://www.radiant.earth/).Radiant Earth's team developed a visualisation app to validate a random subset of the labels collected using the CEO tool.The app retrieves Airbus SPOT imagery over the area of interest in four 6-month windows (covering the two-year of the training data specification).For each AEZ, the app would sequentially retrieve the images for each individual polygon, visualize all the available imagery and overlay the polygon on top of it.A member of Radiant's team would then interpret the class (Crop, No-Crop, Mixed) based on the guideline and examples provided by Digital Earth Africa, and record that in the app.After validating all polygons, a new GeoJSON file would be generated with the additional property for validation labels.The results of this independent validation of DE Africa's reference samples are shown in the table below.The overall accuracy is 96.3 %, indicating that the reference samples are of a high quality and fit-for-purpose.