The Group on Earth Observations Global Agricultural Monitoring Initiative (GEOGLAM) was initially launched by the Group of Twenty (G20) Agriculture Ministers in Paris, June 2011 during the French G20 Presidency. The goal of GEOGLAM is to increase market transparency and improve food security by producing and disseminating relevant, timely, and actionable information on agricultural conditions and outlooks on production at national, regional, and global scales. It achieves this by strengthening the international community’s capacity to utilize coordinated, comprehensive, and sustained Earth observations. Essentially GEOGLAM has delivered by bringing science based information and expertise to decision makers for better policies and programs in support of global food security, all under pinned by a solid research foundation focussed on operational outcomes.
In the 8 years since GEOGLAM was initiated the priorities of the national and international organizations we serve have evolved, particularly with respect to the United Nations 2030 Agenda for Sustainable Development, the Paris Climate Accord and the Sendai Framework. At the same time the technological landscape has shifted with access to open data and big data analytics. So as GEOGLAM looks to the future we seek to embrace these opportunities to address the challenging policy challenges before us.
The presentation will provide specific examples of how GEOGLAM activities have had a tangible impact on food security already. It will then provide an overview of GEOGLAM’s vision for the future by looking at how we are evolving our focus towards more quantified, multi-temporal metrics to address existing and emerging policy drivers, including the development of Essential Agricultural Variables (EAV’s), a Federated approach to information technology, and a new forward looking research agenda.

Global timely and reliable information on crop conditions and early warning of impending shortfalls of crop production, are critical components for achieving food security and ensuring sufficient, reliable food availability and access. This is the aim of the Sustainable Development Goal 2, which sets out to end hunger and ensure access by all people, in particular the poor, and people in vulnerable situations to sufficient food year-round (target 2.1). Earth Observations provide a significant contribution towards providing such crucial crop information, owing to their global, repeatable, and synoptic nature and ability to provide quantitative indicators of crop development through the growing season. The Crop Monitor for Early Warning (CM4EW), based largely on remotely sensed information, is a major step towards provision of such information. It was developed in the framework of the international G20 GEOGLAM program, in response to the pressing need for enhanced early warning of crop shortfalls, and for better coordination across the various agencies responsible for crop assessments in regions most at risk to food insecurity. It is focused on enhancing the use of Earth Observations for improved global, science-driven information on crop conditions and production, building on existing systems. This type of information is particularly important for the most vulnerable countries to food insecurity. However, such information is often incomplete or lacking. As such, the CM4EW was developed with the goal to reduce uncertainty and strengthen decision support by providing actionable information, on a monthly basis, to national, regional and agencies concerned with food security, through timely consensus assessment of crop conditions. In addition to SDG 2, the CM4EW supports Goals 1 (End Poverty), 3 (Health & Well-Being), 12 (Sustainable Production and Consumption), 13 (Climate Change Impacts), and 17 (Global Partnerships). This initiative brings together the main international food security monitoring agencies and organizations to develop monthly crop assessments based on satellite observations, meteorological information, field observations and ground reports, which reflect an international consensus. The first issue of the CM4EW bulletin was published in February of 2016, and in the short time since its launch, has become an internationally recognized source of reliable information on early warning and crop conditions and often used to inform government, and UN organizations decisions on food allocation, and humanitarian assistance. Based on its success, the Crop Monitor model is now being adapted for implementation within national ministries of agriculture in East Africa and Southeast Asia to enhance the use and uptake of remote sensing data into crop monitoring activities and better inform food security decisions at national scale.

The open source Sentinel-2 for Agriculture (Sen2-Agri) system automatically ingests and processes Sentinel-2 and Landsat 8 time series in a seamless way to derive four basic products for agriculture monitoring, identified as a priority by a large user community including the GEOGLAM one. These products are monthly multispectral cloud-free composite, up-to-date cropland mask, mid-season and end-of-season crop type maps, and vegetation status indicator as NDVI and LAI time series.

The Sen2-Agri system relies on streamlined processes based on machine learning algorithms and quality controlled in situ data. It was demonstrated during the 2016 and 2017 growing seasons over three entire countries (Ukraine, Mali, South Africa) and eight local sites distributed across the world. In situ data were collected for algorithm calibration and products validation purposes in a timely manner, thanks to very fruitful collaborations with local partners (including the JECAM network).

For most sites, the overall accuracy of the cropland mask at the end of the season was higher than 90%. Accuracy values higher than 80% were obtained as soon as the middle of the season, with a saturation plateau reached after one or two months. These results clearly show the feasibility to deliver accurate cropland masks as early as the mid-season, independently to the complexity of the cropping systems
The crop type maps depicting the 5 main crops for the considered study sites were also successfully validated: overall accuracy values higher than 80% and F1 Scores of the different crop type classes were most often higher than 0.65. Crop type maps production strongly depends on the availability and the quality of in situ crop type information available in a timely manner for the calibration. Mid-season crop type maps were also produced, with lower accuracy metrics but still reaching a quality level already useable for early operational monitoring.

The good accuracy estimates obtained for crop mapping can be related with the density and quality of the EO time series, with the quality of in situ data collected over each site, but also with the performance of the methods developed to specifically address large-scale agriculture monitoring applications. The challenge is indeed to deal with significant agro-climatic gradients and crop calendars diversity and overlapping, making it very difficult to define the start and the end of a growing season. First, the Sen2-Agri system relies on the concept of rolling cropland masks based on the last 12 months, with a monthly update. The cropland masks delivered each month from the middle of the season proved to be capable of capturing cropping calendar shifts along an agro-climatic gradient, as shown with the accuracy evolution of the cropland masks along the season. Continuous monitoring over more than year needs to be experimented to confirm these good results. A second key aspect of the crop mapping strategy is the use of an a priori stratification of the area of interest, allowing dealing with the Sentinel-2 wide-swath coverage encompassing agro-ecological gradients while being locally relevant thanks to the user knowledge.

The Sen2-Agri demonstration was continued during throughout 2017 and 2018, with nation-wide monitoring over Belgium and South Sudan. In South Sudan, the experiment is conducted with the World Food Program, which organizes national field campaigns to collect in situ data based on the GeoODK mobile application. In 2017, the Sen2-Agri system was also run over the cotton-belt region in Mali. While the 2016 experiment did not allow mapping the crop type in Mali because of significant geolocalization errors in the in situ crop type data, crop types were successfully mapped in 2017. F1 Scores of the three main crop types (mil, rice and cotton) were higher than 0.65, showing the fast learning curve to improve performances.

The Sen2-Agri system is also being uptaked by a growing users’ community, with more than 700 system downloads around the world: World Food Program in Malawi and Uganda, International Maize and Wheat Improvement Center in Mexico, Bangladesh, Iran and Malawi, Agricultural Research Council and Agriculture Ministry in South Africa (Western Cape Province), Australian Commonwealth Scientific and Industrial Research Organisation in Australia, Space Research Institute in Ukraine, Earth Observation Data Center providing a Sen2-Agri cloud solution, etc. Technical and thematic support is provided to these users through an active forum and a series of regular webinars and online trainings.

These experiences clearly highlight the operational agriculture monitoring capacity of the Sen2-Agri system to exploit in near real-time the observation acquired by the Sentinel-2 mission over very large areas. They pave the way for countrywide crop specific monitoring system at parcel level bridging the gap between parcel visits and national scale assessment, thus supporting market transparency while building national monitoring capacity as requested by the AMIS and GEOGLAM G-20 initiatives.

Yield information before harvests is required for decisions regarding the management of agricultural production and markets. Early information on crop emergence date is crucial for forecasting yield estimates; it enables crafting insurance products and food security policies. Current methods (e.g. field surveys) for collecting such information are time consuming and financially demanding at relevant and regional level. Remote sensing data is an alternative tool especially with the data fluxes made freely available to the global community by Sentinel-2 of the Copernicus program.

The objective of this study is to provide operational and early season estimates of crop emergence date at the field level using high-resolution remote sensing data. The innovative challenge is to deliver this information as early as possible after the emergence in a reliable way over large areas (i.e. regional-scale areas).

Thanks to a very comprehensive field datasets surveyed by an insurance company, six methods and four different metrics were compared for estimating crop emergence date using Sentinel-2 time series acquired over the Free State Province, South Africa, during two successive growing seasons. The datasets were provided for 359 fields including information on insured area, district, crop type, cultivar, density (plants/hectare), row width, report date, phenological stage at the time of observation (using a vegetative and reproductive growth stages scale), plant sowing and emergence dates, agricultural management practices including weed management. The datasets were randomly divided for calibrating and validating the used methods with 70% (251) and 30% (108 fields) respectively.
The metrics including vegetation indices (NDVI, MSAVI), Hue and Leaf Area Index were computed from Sentinel-2 data for constructing maize growth profiles through linear and logistic interpolation methods. The selected methods applied different criteria on the growth profiles to detect the more stable vegetation development stage in relation to the actual emergence date. The emergence date is, then, estimated by calibrating the time lag until the growth profile reaches the specified criteria. These criteria are the fixed or relative thresholds, maximum value, inflection point, highest growth rate and base logistic value. In particular, the threshold method considered the date at which the growth profile reached a certain value (fixed or adjusted in relative terms for each profile) for estimating the emergence date. In addition, the base logistic method provides the time lag date at which the line tangential to the inflection point intersected the fitted growth profile at its lowest values.


In general, results showed that the methods based on vegetation indices have high and more stable performances. Emergence date was estimated with deviation ranging mostly between 9 and 20 days according to the different methods. The algorithm combining the relative threshold method, MSAVI and logistic interpolation gave the best emergence date estimates for maize with lowest standard deviation (SD=8.3 days) and mean absolute deviation (MAD=7.4 days). The superiority of this algorithm to others was assured by estimating the emergence date of sunflower where SD was of 9.9 days and MAD was of 7.4 days. Furthermore, this experiment was applied once more during another crop season where the algorithm including relative threshold method, MSAVI and logistic interpolation showed a stable and high performance with SD and MAD recording 12.8 and 5.9 days respectively.

Consequently, it is concluded that the relative threshold method in combination with MSAVI and logistic interpolation method are the best for providing early estimate of crop emergence date with high stability over space and time.

A timely monitoring of agricultural land changes and cropping practices over space and time is one of the prerequisites identified by the GEOGLAM initiative to improve performances of Early Warning Systems for food security, hence contributing to the establishment effective and well-targeted management plans over agricultural territories. The use of remote sensing data in support of expert knowledge is a common practice in this context because of their objective look and the opportunity they offer to generate and update information on agricultural land dynamics in a timely and comprehensive manner. Three years into ESA Sentinel missions, providing optical and radar imagery at high spatial resolution and low revisit time, new scenarios are emerging in remote sensing for the characterization and monitoring of land surfaces, including the systematic generation of Earth Observation derived datasets such as Land Cover maps [1]. This is especially true for tropical countries, where climatic and environmental conditions have long limited the exploitation of satellite imagery for this task, in contexts where user needs are even more significant.

Indeed, Remote Sensing is continuously proving to be an effective tool for the collection of spatialized information on agricultural practices [2]. Taking advantage of our experience within the GEOGLAM-JECAM network, in the last years we have been developing a specific methodology to improve agricultural Land Cover mapping in challenging tropical agro-systems [3], typically dominated by the presence of smallholder rainfed agriculture. In order to address the specificities of such areas (small field sizes, landscape heterogeneity and fragmentation, strong cloudy conditions especially during the cultural season), our proposed solution leverages the coupling of Very High Spatial Resolution imagery (Spot6/7, Pléiades, ...) and optical decametric satellite image time series such as Sentinel-2 and Landsat 8, and is based on an object-based supervised classification approach. This work led to the development of a fully functional processing chain prototype, named MORINGA, and currently constitutes our major contribution to the Scientific Expertise Centre for Land Cover (CES OSO) of the French THEIA Land Data Services Centre.

More in detail, the proposed methodology has been designed to provide fine-scale agricultural land cover classification using multi-sensor satellite imagery. An overall scheme is depicted in the attached abtract figure. In the typical MORINGA workflow, the input dataset always includes a single VHR scene acquired at the peak of the growing season, when landscape heterogeneity is at its most. Several HR optical time series covering at least the annual cropping season are also taken into account, as well as digital elevation and slopes. Reference data for supervised classification is generally constituted by integrating in-situ data collection (mainly on cropland) and photo-interpretation. The general workflow can be easily extended to enable the use of other types of imagery (e.g. radar imagery).

The proposed general classification scheme consists of several main steps: (1) VHR segmentation to resort to a suitable fine-resolution object layer at plot or sub-plot scale; (2) the extraction of a large set of radiometric indices (vegetation, soil, water) from HR and VHR data plus VHR texture indices, (3) object-based featuring of both reference areas (for training) and segmented objects; (4) Random Forest (RF) model fitting and classification of the featured object layer. Several implementation details have been also managed to cope with cloudy acquisitions, the small field and landscape fragmentation constraints, as well as the intra-class variability due to the heterogeneity of cropping practices and plot conditions. This includes (1) a suitable multi-temporal gap-filling per time series as performed in [1], (2) the precise co-registration of the multi-sensor dataset and an optimized fine-scale object statistics computation, and (3) a specific training sample generation strategy based on the intersection between the object layer and polygons from the provided reference dataset.

In order to assess the performances of the methodology on several sufficiently contrasted tropical agro-systems, a specific protocol for testing has been conceived and agreed among several JECAM site contributors in the form of a proper cross-site experiment using, to the extent possible, corresponding input datasets to eventually point out strenghts and weaknesses related to site specificities and acquisition conditions. The target Land Cover nomenclature (hierarchical, with four nested level including a Crop vs. Non-Crop level) has been agreed-upon and harmonized at most among the various tropical JECAM sites. Tested sites include landscapes dominated by smallholder farming (Burkina Faso - Koumbia, Madagascar - Antsirabe) as well as more intensive and heterogeneous agricultural territories (Brazil - Sao Paulo, Reunion Island). These experiments proved, on one side, the general applicability and effectiveness of the methodology even at larger scales with respect to JECAM sites, but also allowed to consistently identify room for improvements.

[1] J. Inglada, A. Vincent, M. Arias, B. Tardy, D. Morin, and I. Rodes. Operational High Resolution Land Cover Map Production at the Country Scale Using Satellite Image Time Series. Remote Sensing, 9(1):95, 2017.

[2] MDPI (Ed.). Remote Sensing in Agriculture and Vegetation: High Resolution Image Time Series for Novel Agricultural Applications [Special Issue]. Remote Sensing, 2018. https://www.mdpi.com/journal/remotesensing/special_issues/highresolution_timeseries_agr

[3] V. Lebourgeois, S. Dupuy, E. Vintrou, M. Ameline, S. Butler, and A. Bégué. A Combined Random Forest and OBIA Classification Scheme for Mapping Smallholder Agriculture at Different Nomenclature Levels Using Multisource Data (Simulated Sentinel-2 Time Series, VHRS and DEM). Remote Sensing, 9(3):259, 2017.

The Anomaly hotspots of Agricultural Production decision support system (ASAP) was launched in June 2017 for making available timely early warning information based mainly on Earth Observation in an advanced and easy to use online platform. The information provided aims at supporting multi-agency early warning initiatives (such as the GEOGLAM Crop Monitor for Early Warning) and food security assessments including the IPC (Integrated Food Security Phase Classification) and the Cadre Harmonisé, which are the basis for the Global Report on Food Crises.
ASAP provides operationally two categories of products: (1) every month for 80 countries with high risk of food insecurity, an assessment of the warnings by a crop analyst in order to identify the so-called “agricultural hotspots and (2) every 10 days at global level, automatic warnings on agricultural drought conditions and their agro-climatic background indicators (rainfall and derived indices including a global water satisfaction index, temperature, biomass proxy, crop phenology).

In 2018 the ASAP system has been improved mainly in three ways:
1.) Addition of national and sub-national crop calendars based on comparison analysis between national level FAO crop calendars and remote sensing based land surface phenology
2.) Production of a global Water Satisfaction Index with crop coefficients and simulation start date based on land surface phenology
3.) Launch of a new version of the so-called “High Resolution Viewer”, an innovative visualization platform for high resolution satellite imagery visualisation and analysis based on Google Earth Engine (GEE).
The crop calendars aim mainly at facilitating the work of the ASAP analysts in their monthly agricultural production hotspots analysis. These calendars, which are also available to the public, complement the FAO crop calendars by identifying the crops that can be monitored with remote sensing indicators, as well as providing information on crop seasons variability at sub-national level.
The new Water Satisfaction Index adds more agro-climatic information to the ASAP system as compared to the simple use of rainfall indicators (e.g. the Standardised Precipitation Index). Model start dates and crop coefficient are based on remote sensing derived land surface phenology.

Finally the new high resolution viewer is a major step forwards in monitoring crop conditions at the parcel level and helps to exploit easy and rapid access to high resolution imagery from the EU Copernicus Sentinel 2 and the USGS Landsat sensors. This allows a fast qualitative comparison of arable crop- and grass-land areas between the current period and a reference period of choice (e.g. the previous year). It also includes a time series function which extracts NDVI time profiles and image chips for parcel level analysis. These features make the viewer suitable also for use outside the ASAP platform, such as for example for monitoring large scale land acquisitions (LSLAs) or project related land use changes at the parcel level.
The ASAP platform will continue to evolve and the next steps already identified include further improvements in the cloud based processing of high resolution satellite imagery shown by the high resolution viewer. Functional modules to select and display profiles and image chips from the Sentinel-1 geocoded backscattering coefficients are ready to be integrated into the operational analysis environment. Short-term extensions will include better cloud screening and compositing, as well as quantification methods, such as histogram analysis over a user defined area and customizable choice of indicators, primarily based on feedback from the analysts.

Thanks to the technical improvements mentioned above, we believe that the ASAP platform has further improved the way it provides access to rapid qualitative crop and rangelands information at the global scale and continues to complement the quantitative yield forecasts provided by the JRC’s Mars Crop Yield Forecasting (MCYF) system operated for Europe and neighbourhood countries. The web platform has been updated frequently and provides a modern and user friendly environment allowing quick access to the main information available. Finally the high resolution viewer is built on the computing power of the Google Earth Engine (GEE) and is a unique demonstration of global level cloud based access to high resolution satellite imagery for any point of the globe in near real time. high resolution satellite imagery and visual analysis for agricultural monitoring.