The International Production Assessment Division (IPAD) of the U.S. Department of Agriculture’s (USDA’s) Foreign Agricultural Service (FAS) is mandated to provide the most objective, timely, and accurate assessment of global agricultural outlook and current crop conditions affecting global food supply and food security. IPAD monitors agricultural productivity at national to global scale and supports the G20 Agricultural minsters request for increased market transparency. Monthly global and national production estimates are published by IPAD in the World Agricultural Production (WAP) circular which is released simultaneously with the World Agricultural Supply and Demand Estimates (WASDE) report from USDA's World Agricultural Outlook Board (WAOB). Both WAP and WASDE reports summarize global crop production on a monthly basis and data from these reports are used by the U.S. Office of Management and Budget (OMB) as Principle Federal Economic Indicators.

IPAD uses a convergence of evidence methodology to estimate area, yield, and crop production during the Inter-Agency Commodity Estimates Committee process. IPAD provides an operational crop monitoring system based on medium to coarse satellite remote sensing for the sake of consistency with the existing long-term archives. The multi-source methodology incorporates weather, soil moisture, and satellite-derived vegetation indices in the context of crop geography and crop calendars. Fieldwork, input from FAS’ network of overseas attachés, economic analysis, and other sources add to the methodology. Crop production estimates are updated monthly and are incorporated into USDA’s World Agricultural Supply and Demand Estimates which provides information for commodity price discovery.

Desktop software that supports the Foreign Agricultural Service (FAS) crop analyst’s mission to estimate crop production has relied on time-intensive manual processes and the storage of terabytes of satellite imagery, meteorological, and ancillary data. To offset these disadvantages FAS has funded the development of the Global Agricultural and Disaster Assessment System (GADAS): in-line with U.S. Federal Government’s directive to consider moving to the cloud “first and foremost.”
GADAS, a “citizen-facing”, enterprise web Geographic Information System (GIS), allows FAS crop analysts to meet their mission more efficiently while taking advantage of the cloud. GADAS consumes hundreds of web services, from both inside FAS as well as from external commercial providers, including hazards and disaster layers (a partnership between FAS and the Pacific Disaster Center). GADAS web services cover many disciplines: weather, vegetation index, crop conditions, disasters, land cover, infrastructure, elevation, and satellite imagery (Landsat and MODIS). ESRI technology is at the core; however, custom functionality enabled performance of mission-critical tasks. Functionality includes comparative visualization as well as analysis, including the embedding of imagery classification functions directly in the application. GADAS sends requests to external web services, which perform the analysis, and then results are displayed. Classification of imagery can be accomplished without the need for FAS to store any imagery data.

GADAS supports the mission of the Washington-based FAS crop analysts, but because of the global extent of its data and cloud application, a broad diversity of users within other countries can be supported. Training for FAS’ overseas personnel at U.S. Embassies and several foreign agricultural ministries, starting in Honduras, will improve collaboration. Fostering collaboration among agricultural ministries aids USDA’s decision-making process for estimating the global agricultural supply and demand of field crops.

Area A - Board 81

Present and future earth observation (EO) missions are key to the design of regionally appropriate solutions to fulfill the raising food demands and ensure food security while keeping the environmental impacts to a minimum. Studies on crop phenology and crop yield forecasting from satellite information have been traditionally focused on the use of optical vegetation indices such as the enhanced vegetation index (EVI). These indices are sensitive to the amount of photosynthetic active vegetation, which depends on the biotic and abiotic conditions that affect crop status and, ultimately, determine final yield. More recently, several studies have shown that microwave satellite time series of vegetation optical depth (VOD) carry information about crop growth and yield in a variety of agro-ecosystems. VOD is a measure of the attenuation of soil emissivity through vegetation; it is sensitive to the amount of living biomass as well as to the amount of water stress experienced by the vegetation.

In this study, we propose a combination of optical MODIS EVI and microwave SMAP VOD and statistical machine learning methods for crop yield predictions in the US Corn Belt (8 states, see Fig 1). Agricultural survey data from the National Agricultural Statistics Service (USDA-NASS) are used as a benchmark. We fixed a temporal window from April to October of 2015 to predict the crop yield of that year. This 7-month temporal window is chosen to include the growing and senescence season of all crops in all states, and ensure predictions could be performed before the harvest is ended.

Linear regression (LR) and its nonlinear natural extension called kernel ridge regression (KRR) models were build and compared. Time series of EVI and VOD satellite data extracted at the county-scale conformed the input feature space. A first experiment was undertaken to predict the total crop yield at the county scale without differentiating crop types. The purpose of this experiment was to provide an assessment of the impact of crop heterogeneity on the models and the prediction capabilities in mixed crops. A set of experiments where then set up to build yield prediction models for the most important crops grown in the region, namely corn, soybean and wheat.

An error assessment for the different techniques and configurations was performed. Results show that the use of information from the two vegetation products leads to improved predictions that the use of each index separately. Maps of predicted yield show a good agreement with survey data (R² > 0.7, relative error < 20%)(see Fig 2). A phenological analysis shows that VOD and EVI times series during the growing season are in phase, with a lag explained by the fact that maximum greenness (from EVI) precedes maximum biomass (from VOD).

This work supports the use of satellite time series information as opposed to arbitrary unidimensional metrics, such as the maximum or mean of the vegetation indices in crop yield prediction models. Also, it shows the use of new microwave EO information alongside well-known optical indices can potentially improve present agricultural monitoring and yield prediction capabilities. This stresses the importance of ensuring the continuity of passive microwave observations beyond SMOS, SMAP, and AMSRE, which is now being proposed via the Copernicus Imaging Microwave Radiometer (CIMR) high priority candidate mission.

Area A - Board 82

The aim of the study was to examine the applicability of soil-vegetation parameters calculated from Sentinel2, Sentinel-1, Sentinel-3, PROBA-V satellites data on crop monitoring, its phenological stage and yield prognosis. The extensive field measurements have been carried out simultaneously to satellites observations in order to elaborate the best relationship between satellite data and in-situ measured LAI which further has been applied for the modeling of crop yield and biomass. The prognosis model based on meteorological data with the periodical input of LAI for wheat yield prognosis have been applied. Additionally classification of crops over JECAM site in the NUTS4 district was performed for 2016 -2018 to choose the fields with wheat crop. Also the data base of NOAA/AVHRR satellite has been used to apply the crop conditions finding the trends in crop development for the same area during 1997-2018.
The relationships between satellite-based and ground measured LAI values were studied in a temporal profile and conclusions on precision of remotely sensed LAI estimation were elaborated. The Simplified Surface Energy Balance model (SSEBop) parameterized for operational applications as an innovative method allowing the estimation of evapotranspiration (ET) using remotely sensed data (Ts;NDVI) and global weather datasets have been applied to estimate the crop growth conditions which influence crop yield prediction. The most recent innovations in the retrieval methodologies of biophysical variables allowing the monitoring of vegetation state and dynamics have been presented using Sentinel1 data. In the research the fusion of optical data and radar data have been presented. The analysis of phenology stages have been examined through the NDVI and temperature accumulation during the crop growth. The conditions in 2016, 2017 and 2018 differed. The severe drought in 2018 has been examined through different values of soil - crop parameters deduced from S2, S1. The conditions were presented and compared with these deducted from the 20 years of NOAA/AVHRR data in order to adapt the actual satellite-based algorithms to NOAA/AVHRR characteristics of soil-vegetation conditions in order to present the time series and continuity for the actual data.
The presentation will give the input to EuroGEOSS on Agriculture/Food together with other scientists what will give the opportunity to consider mutual way of the involvement to GEOSS. The observations at the JECAM site has been carried out through the ESA Projects: Land Products Validation and Characterisation in support to Proba-V, S-2 and S-3 missions and Service for Drought Monitoring .

Area A - Board 83

Remote sensing data allows the monitoring of land cover at local, regional and global scale. In the last years, the availability and increase of satellite images with diverse spatial and spectral resolutions improved the monitoring of agricultural sectors in the world. However, the annual crop mapping at the large scales has drawbacks due to the complex crop dynamic and the available training data from study area, which limits the classifier reproducibility in other years and places. Space and time generalization have been studied to map forest and crop types on years and locations not used in the calibration of the classifier, but few space-time generalization studies were applied to crop monitoring, especially for sugarcane areas. A regional mapping of sugarcane areas is done in São Paulo (SP), the main sugarcane producer state in Brazil. A space and time generalized classifier is developed for mapping sugarcane areas with Landsat images time-series and random forest algorithm applied on objects after segmenting the image to close-to-field scale polygons. The potential of spectral indices and bands related to near infrared and short wave infrared regions, that have good performance to identify sugarcane areas as underlined in Luciano et al. (2018), is evaluated and combined with phenological indices to map sugarcane areas at regional scale. The random forest is calibrated at 10 sites located across SP region, with 6 years of training data (2009 to 2014). A global space and time classifier is created with all sites calibration and applied to the SP region for an independent year (2015). The thematic map of sugarcane areas of year 2015 is shown in the attached abstract figure. The importance of the main predictions variables are evaluated using the random forest algorithm, following by a feature selection to reduce the number of variables and keep the most important ones for easier and faster application at large scales, even though the feature selection slightly reduced the accuracies. The global model is compared with the reference map and regional official statistics from the Brazilian government, with obtained R² at higher than 0.9 at municipality scale. Due to the recent availability of Sentinel-2 images, there are new possibilities for improving these maps. A global space-time generalized model allows the monitoring of sugarcane areas without new training data and with low cost and resources, being an important tool for sustainable crop and environmental management.

Luciano, A.C. dos S., Picoli, M.C.A., Rocha, J.V., Franco, H.C.J., Sanches, G.M., Leal, M.R.L.V., le Maire, G., 2018. Generalized space-time classifiers for monitoring sugarcane areas in Brazil. Remote Sens. Environ. https://doi.org/10.1016/j.rse.2018.06.017

Area A - Board 84

The Copernicus programme with Sentinel-1 and Sentinel-2 has opened a new chapter in monitoring agriculture production on a country level. The ESA EOStat project presented here exploits these new capabilities, aiming at: (1) creation of a system of EO-based support for acquisition of statistical information for agricultural production in Poland, (2) providing an assessment of economic viability of the system application in the operational work of the statistical office, as well as (3) providing the guidelines for further uptake of EO-based services for collection of agricultural statistics. The project also strives to provide a suite of evaluated and documented tools for automatic processing of satellite data and derivation of information relevant for agricultural production assessment (i.e., crop maps and crop condition). Here, we demonstrate how Sentinel 1, Sentinel-2 and Sentinel-3 can be employed to build the system for gathering statistical information for agriculture.

We test several crop type classification techniques using solely Sentinel-1 or Sentinel-2 data (also employing Sen2Agri system), and we develop a new data fusion approach. We aim to recognise 9 types of crops. The proposed classification approach is based on a set of indices which reflect phenological changes in crop growth. The classification algorithm is trained and validated using various sources of reference data: in-situ campaigns, farmers’ declarations, and Paying Agency in-situ control results. We test the sensibility of the classification methods to quality of source data, amount of training and testing points, etc. Moreover, we apply Sentinel-3 OLCI and SLSTR to provide an assessment of apparent crop growth conditions, and operational crop yield forecasting. Forecasting is based on a statistical relation of apparent crop conditions and the long-term average derived from long-term NOAA/AVHRR data record. To assure the consistency of these two data sources, the data are intercalibrated by means of spectral convolution. In addition, we develop an early warning system of potential losses in crop production due to natural hazards such as drought, frost kill or floods.

Area A - Board 85

Grassland ecosystems are among the most valuable environments in Europe hosting a variety of protected habitats and species under the have been under increasing pressure in Europe. With the introduction of the Renewable Energy Act (EEG, Erneuerbare Energien Gesetz) the use of renewable energy including biomass has strongly increased in Germany since 2000. This has put increasing pressure on grassland caused by land use intensification and land cover changes. Aim of this study was to test and discriminate grassland evolutions over three time steps: 2016, 2006 and 1990.
In the past the distinction between grassland and crops was very difficult, due to the lack of multiple images reflecting essential phenological development stages of grassland and crops. Current active and passive satellite systems are able to provide several images within one year representing the complete vegetation period, which was not possible before. Embedded in a study to assess these changing energy market effects on grassland and arable land distribution in Germany, baseline maps were produced using Sentinel-2 time series data from 2015/2016.
For the production of these maps we tested the Sen2Agri software system, freely provided by the ESA, at a national German scale. The applied random forest classifiers yielded very good results for the arable and grassland land cover maps with an overall mapping accuracy of 91% (+/- 0.7%). It showed the high potential for operational mapping of grassland and arable land changes in Germany. In this presentation we will focus on the experiences gained with the Sen2Agri System and present the outcomes of a final validation with an independent data set. Using field reference information from over 7000 sample points based on the LUCAS 2015 and other field data sources, the maps where validated. The results showed a slight overestimation of grassland in Germany, which was contributed to the slightly different class definitions and resolutions.

Area A - Board 86

In mountainous regions, agricultural management is a key strategy to preserve ecosystems stability and local economies. Vegetation phenology is a significant indicator of crop health and productivity. Notably, changes of the start and end of the growing season might have a significant impact on fragile mountain crops, with a reduction of ecosystems resilience.
Phenological phases monitoring is therefore relevant in the decision-making process and allows to better understand the impact of climate variability. In addition, by monitoring the mowing activities, the productivity of mountain meadows can be assessed.
Optical remote sensing has been widely used for monitoring crop dynamics. Most of the studies using satellite images for crop monitoring focused on the use of optical images, exploiting the reflectance of visible and near-infrared (NIR) radiation and the emittance of infrared thermal radiation (IR).
However, cloud contamination and topographical effects significantly affect optical data.
In contrast to optical images, Synthetic Aperture Radar (SAR) can provide a systematic source of data on surface changes that are insensitive to cloud contamination. At the same time, however, the radar signal may be difficult to be interpreted because the total backscattering of the radar is a sum of the backscattering of vegetation and soil. However, the VH channel has proved effective in following phenological steps on meadow areas.
A synergic integration of microwave and optical sensors could give more accurate results in the agricultural monitoring in mountain regions. Combining Sentinel-1 and Sentinel-2, with a revisiting time of six and five days, respectively, offer an unprecedented opportunity to monitor crop phenology.
To assess the accuracy of satellite processed products, a field validation is necessary. To overcome traditional approaches, proximal sensors have become a key method for validating phenological phases.
Indexes derived from digital cameras offer the possibility to record high-resolution spatial and temporal data in remote areas.
In this study, we analyzed the vegetation phenology in the alpine areas of South-Tyrol (Italy) by constructing time-series from both Sentinel-1 and Sentinel-2.
Here, we investigated the impact of different parameters on SAR signal, performed a correlation analysis between sensors and presented a new methodology to detect phenological stages using σ⁰ SAR time series. A sensitivity analysis has been done of VV and VH polarized images covering the time span between 2015 and 2017, together with the NDVI from Sentinel-2 and ground sensors. After a noise removal using different techniques, several filters were applied to SAR and Optical time series.
From the modelled values we extracted the Start of Season (SOS), Maximum, End of Season (EOS) and harvesting time, and we validated our results using information from NDVI sensors and phenocams. Furthermore, SAR sensitivity to vegetation component has been analysed in relation to altitude, slope and plant geometry.

From our study we can assume that multi-temporal SAR signal and specifically the VH polarization can be used to detect phenological dynamics in grassland and in particular, multi-temporal SAR signal is well correlated to the NDVI from Sentinel-2 and ground observations of vegetation indices. From our analyses, the VH channel loses sensitivity with increasing altitude, while the VV channel is affected by the soil component and cannot be used at high altitudes. In addition, the geometry of the vegetation influences the results in the VH channel, making phenological phases detection difficult at altitudes above 2000 m a.s.l.

Area A - Board 87

Time series of remote sensing vegetation indices and vegetation biophysical parameters are of primary importance for vegetation monitoring, and represent an essential component of the existing crop monitoring and yield forecasting operational services (Baruth et al., 2006; Fritz et al., 2018, Lopez-Lozano & Baruth, 2018). Medium and low-resolution platforms were at the core of those components, as near real-time monitoring of vegetation requires a high revisiting time. The launch of the Sentinel-2 platforms opens new avenues for agricultural applications, providing for the first time freely accessible high-resolution imagery with 5-day revisiting time. That may enable the operational use of high-resolution imagery for near real-time and crop-specific crop monitoring and yield forecasting in Europe.

This contribution aims to investigate the feasibility of producing continuous Sentinel-2 vegetation index time-series for croplands in Europe. More specifically, addresses the impact of noise in vegetation products and the expected availability of cloud-free imagery in the reliability of continuous pixel-level time-series with Sentinel-2 imagery. The resulting time-series could be used for different purposes: crop classification, identification of key phenological stages or its use on empirical or deterministic models for yield forecasting.

For these purposes an operational chain to produce top-of canopy (TOC, equivalent to L2A) reflectance was created to further retrieve calibrated NDVI imagery. The atmospheric correction was performed using the Simplified Model for Atmospheric Correction algorithm (SMAC, Rahman and Dedieu, 1994) using as dynamic input the Aerosol Optical Density product (AOD, MYD04_3k collection 6) from the MODIS platform. The solutions implemented in the Sen2cor algorithm for cloud and shadow detection were incorporated as well in the operational chain.

Five different study sites in Europe, corresponding each one to a single Sentinel-2 grid tiles, were selected as they are representative of the agro-climatic variability of Europe: north-west of Spain, south-west of Romania, central France, northern Germany; and an agricultural area border between Latvia and Lithuania. These five sites, present as well a wide variability of cloud contamination (Whitcraft et al., 2015).

For each tile, the 2016 and 2017 available Sentinel-2A and 2B Top Of Atmosphere (level L1C) products were downloaded from the Copernicus Open Access Hub data warehouse to obtain NDVI. An experimental dataset of reference NDVI time-series were created selecting the pixels with the highest availability of cloud free data (at least 80% of the total available images) and applying a gap-filling and smoothing algorithm consisting in a polynomial function applied over a 125-days moving window, followed by a Savitzky-Golay filter. These synthetic time-series represent an approximation to the “true” NDVI without noise and no missing data due to cloud cover. The NDVI values were, first, contaminated with a Gaussian relative noise to model satellite and sensor factors (e.g. uncertainties in atmospheric correction, sensor calibration, etc.) and from atmospheric cloud conditions, secondly, gaps were simulated in the time-series to account for cloud coverage, following an approach similar to Kandasamy et al. (2013). In order to simulate realistically the presence of clouds on each study site, cloud probability was modelled using a MSG-SEVIRI sunshine duration product, computed from 15-minutes pixel-based cloud masks. The simulated time-series with noise and missing data due to cloud contamination were smoothed and gaps were filled applying the Whittaker Filter. Therefore, a different algorithm has been chosen to generate the reference time-series and to reconstruct them, to avoid possible bias. The reliability of the time-series reconstruction was assessed comparing the filtered NDVI values with the reference time-series.

Results are discussed in view of their applicability for large-scale monitoring systems, by considering: 1) the spatial quality of the smoothed products; 2) the effect of gap proportion and gap distribution in the reliability of reconstructed NDVI in the different study sites; and 3) the suitability of a methodology based on the smoothing and gap-filling on Sentinel-2 data for continuous crop monitoring across Europe, based on the expected fraction of missing data.



References:
- Baruth, B., Royer, A., Klisch, A. and Genovese, G., 2008. The use of remote sensing within the MARS crop yield monitoring system of the European Commission. Int. Arch. Photogramm. Remote Sens. Spat. Inf. Sci, 37, pp.935-940
- Fritz, S., See, L., Bayas, J.C.L., Waldner, F., Jacques, D., Becker-Reshef, I., Whitcraft, A., Baruth, B., Bonifacio, R., Crutchfield, J. and Rembold, F., 2018. A comparison of global agricultural monitoring systems and current gaps. Agricultural Systems.
- Kandasamy, S., Baret, F., Verger, A., Neveux, P. and Weiss, M., 2013. A comparison of methods for smoothing and gap filling time series of remote sensing observations–application to MODIS LAI products. Biogeosciences, 10(6), pp.4055-4071.
- López-Lozano, R. and Baruth, B., 2018. An evaluation framework to build a cost-efficient crop monitoring system. Experiences from the extension of the European crop monitoring system. Agricultural Systems.
- Rahman, H. and Dedieu, G., 1994. SMAC: a simplified method for the atmospheric correction of satellite measurements in the solar spectrum. Remote Sensing, 15(1), pp.123-143.
- Whitcraft, A.K., Vermote, E.F., Becker-Reshef, I. and Justice, C.O., 2015. Cloud cover throughout the agricultural growing season: Impacts on passive optical earth observations. Remote sensing of Environment, 156, pp.438-447.

Area A - Board 88

The main objective of EO4AGRI is to catalyse the evolution of the European capacity for improving operational agriculture monitoring from local to global levels based on information derived from Copernicus satellite observation data and through exploitation of associated geospatial and socio-economic information services.
• EO4AGRI assists the implementation of the EU Common Agricultural Policy (CAP) with special attention to the CAP2020 reform, to requirements of Paying Agencies, and for the Integrated Administration and Control System (IACS) processes.
• EO4AGRI works with farmers, farmer associations and agro-food industry on specifications of data-driven farming services with focus on increasing the utilization of EC investments into Copernicus Data and Information Services (DIAS).
• EO4AGRI addresses global food security challenges coordinated within the G20 Global Agricultural Monitoring initiative (GEOGLAM) capitalizing on Copernicus Open Data as input to the Famine Early Warning System Network (FEW-NET).
• EO4AGRI assesses information about land-use and agricultural service needs and offers to financial investors and insurances and the potential added value of fuelling those services with Copernicus information.

The EO4AGRI team consists of 11 organizations, complementary in their roles and expertise, covering a good part of the value-chain with a significant relevant networking capital as documented in numerous project affiliations and the formal support declarations collected for EO4AGRI. All partners show large records of activities either in Copernicus RTD, governmental functions, or downstream service operations. The Coordinator of EO4AGRI is a major industrial player with proven capacities to lead H2020 projects.

The EO4AGRI project methodology is a combination of community building; service gap analysis; technology watch; strategic research agenda design and policy recommendations; dissemination (incl. organization of hackathons).

This project has received funding from the European Union’s Horizon 2020 research and innovation programme under grant agreement No 821940. EO4AGRI started on 1st of November 2018 and lasts for 24 months.

Area A - Board 89

The Group on Earth Observations Global Agricultural Monitoring (GEOGLAM) Initiative, through its global and regional networks, aims to enhance national capacity to use Earth observations for agricultural monitoring. In turn, these national systems can support GEO's objective of a global system of systems, with south-south, north-south, and triangular cooperation, collaboration, and knowledge and technology flow as key characteristics.

To date, while there are numerous capacity development initiatives and programs active throughout the Earth observations world, there is little coordination between them and very limited sharing of resources or experiences. Further, there is a general dearth of longer-term, holistic, and comprehensive visions for how to develop sustained national capacity for EO usage in an institutional context. In this vein, GEOGLAM and in particular its regional network Agricultural Monitoring in the Americas is working to develop a comprehensive model for capacity development in the agricultural monitoring. Such a plan would be tailored for GEOGLAM needs, but its design and framework would be broadly applicable and could therefore serve as a model for designing capacity development initiatives and activities in other thematic and geographic areas. Such a model would include guidance and documentation on:
• Identifying stakeholders and undertaking a needs-assessment;
• How to trace backwards from a needs-assessment to data and skills requirements and gaps in order to identify concrete steps and relevant actors;
• “Best-practices” for knowledge and technology transfer, particularly for “operationalizing” research and development outcomes;
• How to develop and deploy training materials in a bi- or multi-national context, particularly using a “co-design” framework;
• What to ask participants before, during, and after a training workshop or longer-term knowledge and technology transfer relationship, in order to determine what needs to occur, what is or is not working, and what modalities and activities were most and least successful;
• “Best-practices” for stewardship of training materials (e.g. manuals, datasets, software tools) with an emphasis on discoverability and accessibility, so as to maximize the reuse of high-quality materials throughout the community. This would draw upon lessons learned in the GEOCAB Portal development and deployment exercise;
• The order of operations and timelines for carrying out each of the above activities, including how to identify and work in multinational, collaborative settings.
Each of these activities would leverage the experiences, resources, and existing documentation from related activities including NASA SERVIR, ARSET, and DEVELOP programs, the Committee on Earth Observation Satellites Working Group on Capacity Development, GEO’s Capacity Building Coordination Working Group, and examples from scientific conferences and the gray, white, and peer-reviewed literature.

This presentation will cover GEOGLAM's capacity development activities thus far and describe progress and next steps toward accomplishing the development of this comprehensive model for capacity development.

Area A - Board 90

Argentina is participating in the Joint Experiment for Crop Assessment and Monitoring (JECAM) since the early beginning of the network in 2010. A site was implemented in the Rolling Pampas (San Antonio de Areco, Buenos Aires province) with the main objective of testing methodologies for crop type mapping and estimation of crop biophysical parameters. As many JECAM sites, the original study area was about 100×100 Km, probably related to available processing capacity and high resolution image availability. During last time, field data was obtained without interruption and classification techniques were tested and refined and accurate annual crop type maps were generated for the study area. With recent advances in image processing capabilities and improvements in image availability and access related to the launch of new satellites (e.g. Sentinel 1 and 2), it became more feasible the mapping of larger areas with high resolution images, something necessary for crop type identification at field level. Nevertheless, the need for in situ data for calibrating and testing products is still a challenge. For current campaign (2018-2019) we expanded our JECAM like on-road surveys along main agricultural areas of Argentina. Surveys were coordinated by two INTA units located in Center Argentina (Hurlingham, Buenos Aires province) and North Argentina (Salta province). Total area was divided using a zonation based mainly in crop presence and abundance. For each zone, geo-located points of crop types and no cropland areas are being registered twice a year (during winter and summer agricultural campaigns) for the generation of training and validation datasets. Supervised classifications are being performed using Landsat 8, Sentinel 2 and Sentinel 1 thought the Google Earth Engine platform. Results of this project will fulfill information that is scarce in Argentina, and will allow characterization and forecasting of agricultural production using an objective methodology with certain level of accuracy. Replication of this study for historical periods will help in the analysis of agricultural impacts, allowing the characterization of land use transitions and crop rotation or monoculture occurrence.

Area A - Board 91

The Federal District is one of the 27 federative units in Brazil, located in the Cerrado region. According to (Assad, 1994), this biome presents two well-defined seasons, one drought, which begins in the month of May, ending in September, and another rainy season, which runs from October to April. This climatic characteristic allows the occurrence of two grain crops. In Federal District spring-planted crops there is the cultivation of soybeans, maize and beans from October. Most of the grain area is occupied by soybeans. Cultivation of the summer-planted crops occurs shortly after harvesting the spring-planted ones, in the same area. Among summer-planted grains, the largest area is cultivated with maize. The main focus of this study was to estimate by statistical sampling the summer-planted maize area, using the spring-planted crop map for the 2017/18 season.

The mapping of the spring-planted crop was performed through Sentinel-2 images. There was elaboration of RGB - 432 (true color) compositions of 10 meters spatial resolution that corresponded to the crop development period - September/17 to February/18. The mapping was performed by manual vectoring on a 1:10.000 scale through visual interpretation. The identification of cultivation areas occurred by analyzing the spectral behavior according to the crops stages. In September and October, during soil preparation and sowing stage, the spectral response is typically of soil exposed. During the development stage, from November to January, the spectrum of crops with high vegetative vigor results in differentiation from other classes such as pasture and natural vegetation. In the ripening and harvesting stages, in February, senescent or already absent vegetation produces characteristic spectral response.

The sampling method was adapted from (Potapov P. et al, 2017). Using two-stage sampling, it was possible to estimate area from sample points on the field.
On the first stage of the sampling, all the crop area on Federal District was divided on blocks of 90x90km. Those blocks were classified on three strata: with less then 25% of mapped area inside the block, between 25% and 75% of mapped area inside the block and more than 75% of mapped area inside the block. Using Neymman allocation procedure, were randomly selected 3 blocks from the first stratum, 17 from the second and 15 from the third.

On the second stage, 12 points on each of the blocks were randomly selected. On the field work, each point was verified and notes were taken about what crop it represents. After collecting this data, is possible to estimate the area for each block, each stratum and the total area.

Define Αh,i to be the area of cropland on block i of stratum h, and ph,i to be the estimate proportion of maize from the sample points within the area of block i of stratum h. The area of maize in which block is estimated as Y_h,i = A_h,i * p_h,i. For each stratum, the area is calculated by Y_h=(N_h/n_h)*∑_i(Y_h,i). Finally, the total area of maize is estimated by Y = ∑_h Y_h.

For Federal District, the area of summer-planted maize on season 2017/18 was estimated in 35,884 hectares. The official Brazilian data for the same period is 35,800 hectares (Conab, 2018). The results shows that this method is pretty accurate and is possible to estimate summer-planted crop area using spring-planted crop map for stratifying, as the cropland area used on the two crops is almost the same.

Reference:
1) Assad, E.D. (ed.) Chuva nos cerrados. Análise e espacialização. Embrapa/SPI. Brasília. 423p 1994.
2) Xiao-Peng Song, Peter V. Potapov, Alexander Krylov, LeeAnn King, Carlos M. Di Bella, Amy Hudson, Ahmad Khan, Bernard Adusei, Stephen V. Stehman, Matthew C. Hansen, National-scale soybean mapping and area estimation in the United States using medium resolution satellite imagery and field survey, Remote Sensing of Environment, Volume 190, 2017, Pages 383-395, ISSN 0034-4257, https://doi.org/10.1016/j.rse.2017.01.008.
3) COMPANHIA NACIONAL DE ABASTECIMENTO. Acompanhamento da Safra Brasileira de Grãos. Brasília, v.5, Safra 2017/18.n.12, décimo segundo levantamento, set .2018.p.106.

Area A - Board 92

The Asia-RiCE initiative (http://www.asia-rice.org) has been organized to enhance rice production estimates through the use of Earth observation satellites data, and seeks to ensure that Asian rice crops are appropriately represented within GEO Global Agriculture Monitoring (GEO-GLAM) to support FAO Agriculture Market Information System (FAO-AMIS). Asia-RiCE is composed of national teams that are actively contributing to the Crop Monitor for AMIS and developing technical demonstrations of rice crop monitoring activities using both Synthetic Aperture Radar (SAR) data (Radarsat-2 from 2013; Sentinel-1 and ALOS-2 from 2015 and extend such activity from a province estimation to national estimation. And also, using GCOM-W, GPM, MODIS, Himawari and other satellites, JAXA provides agro-met information to Asia to support monthly rice crop outlook in cooperation with ASEAN food security information project (AFSIS).
This paper reports GEOGLAM activity accomplishment in Asia using Japan’s satellites in 2018.

Area A - Board 93

The constellation of Sentinel satellites launched by the European Space Agency in recent years offers an unprecedented access to free satellite images at high spatial resolution (10-20 m) and frequent revisit (5-12 days). The Sentinel-1 constellation offers weather-independent day and night observations of the whole globe land mass every 12 days in both ascending and descending tracks. Few areas, such as the Mekong Delta in Vietnam, are even updated every 6 days.
A new rice crop growth monitoring service has been developed by SarVision and the University of Wageningen with Sentinel-1 satellite data. The service is based on a generic methodology to map crop age, fallow age and crop growth stage using dense time series of radar data. The approach is a modification of Kalman filtering, capable to handle several rice models simultaneously. It uses an empirical radar signature database to estimate crop starting date and crop age. An empirical look up table database is used to convert the crop age to the crop growth stage. Growing rice fields are automatically mapped and classified according to the rice growth stage (seedling, tillering, booting, flowering, milking, ripening, harvested), see Fig. 1. The complete processing chain is fully automated, allowing the update of maps maximum two days after image acquisition.
1. Method description
Sentinel-1 IWS SLC images are speckle filtered and ortho-rectified into geocoded VH and VV bands with a 15 meter geometric resolution. The thematic algorithm uses VH and VV bands as inputs to generate rice extent and growth stage maps.
Time series of radar images are well suitable to identify rice crops and monitor its development. This is partly caused by the characteristics of the radar temporal backscatter signal, which is distinct from most other crops because of the inundation of rice fields. Figure 2 shows the temporal profile for 3 models of rice between transplantation and harvest, which have been extracted on rice fields during 3 consecutive rice growing seasons in the Mekong Delta. Radar backscatter values are computed on each field polygon for each polarisation (VH, VV) and plotted against time after transplanting/sowing, creating crop models. Transplanting/sowing dates, harvest dates, in-season growth stages and field boundaries are recorded through field work.
The algorithm analyses how well the observed multi-temporal backscatter signal fits with the models temporal profile in order to classify pixels as rice or non-rice and estimate the crop age. However, in practice growing patterns show variations due to multiple crop cycles, time shifts in crop starting dates, variations in length of the cropping season and variations in the amplitude of the temporal profile. Kalman filtering is used to follow such variations, using information from previous and future images to estimate the position of one observation in the growing season.
Once the crop age is estimated, it is converted into the crop growth stage using an empirical look up table derived from field records. In the Mekong Delta, a total of nine rice crop models and look up tables are used.
2. Validation
Validation of the rice growth stage maps is carried out by comparing mapped growth stage to recorded growth stage on rice fields. The validation dataset consists in geo-located polygons on which growth stage information is collected periodically. Growth stages are translated in a numerical scale, from 3_seedling to 9_harvested).
A simple linear interpolator is used for deriving the growth stage at the time of Sentinel-1 data acquisition since the Sentinel-1 acquisition dates may not be coincident with the in-situ data measurements. This assumption constitutes an uncertainty source because the temporal evolution of growth stages may not be linear in many cases. Therefore, interpolated growth stages are recorded with one decimal number. For instance, for a field measured 4 at t0 and 5 at t0+6 days, the growth stage will be recorded as 4.5 if the satellite acquisition date is t0+3 days. In this way, is obtained a validation dataset with, for each polygon, an interpolated value of the growth stage at each date of sentinel-1 data acquisition.
For each polygon, statistics of pixel values from the growth stage maps were extracted. Pixels which intersects with boundaries of the polygon were excluded. The “mapped growth stage” assigned to the polygon was the mode value, i.e. the value that occurs most often.
Then, the “mapped growth stage” was compared to the growth stage interpolated from the field measurements. For each pair of [field polygon Id, map date] values, the mapped growth stage was plotted against the field data growth stage. The determination factor (R2) and the root mean square error were computed (RMSE). Figure 3 shows results obtained over the Mekong Delta in Vietnam, with a R2 of 0.94 and RMSE of 0.38 (1 unit corresponds to 1 growth stage).
3.Applications
The technology is being operational and implemented in Vietnam (https://sat4rice.wordpress.com/) over the Mekong Delta, and Bangladesh (http://idss.com.bd/) over the whole country. It is being developed in Myanmar, Indonesia and Tanzania. After an initial set up phase of calibration and validation using local field data, it can be scaled up quickly to cover large regions, thanks to the global availability of Sentinel-1 data.
It provides near real time in-season geo-statistics on rice cultivated areas and growth stages, feeding digital online platforms and mobile applications. Extension officers have access to up to date information for prioritizing advisory tasks and react to actual situations in the field. Information on pests and diseases risks is extrapolated from crop growth stage maps, weather data and agronomic knowledge. Agro-businesses have access to new analytics to improve logistics, increase knowledge on their market and engage farmers. Failed rice crops can be tracked and related to flooding events, past and current rice crop cycles are recorded, with the potential to support finance and insurance services for rice farmers. In season estimates of start of season and peak of season dates, integrated to crop yield models, have the potential to support early crop yield estimation and forecasting.
The technology has been developed with the support of the G4AW program (https://g4aw.spaceoffice.nl/en/) funded by the Dutch Ministry of Foreign Affairs and implemented by the Netherlands Space Office.

Area A - Board 94

In the big data era, various kinds of satellite data are made easily and/or freely available in the world. Therefore, the crop type mapping with these satellite data has strongly attracted the attention from the remote sensing researchers. As a new comer, the high resolution satellite series, short name for GF, are being developed in China. GF-1 data has a 16 meters ground sampling in 4 bands, such as blue, green, red and near infrared spectra. In Europe, the Copernicus project ensures the stable Sentinel satellite series and provides multispectral and 10-meter resolution optical satellite images to the worldwide end users. USGS’s LANDSAT 8 data has been accessible free of charge online since 2013. These satellite images become the rich data sources for the crop type mapping with the machine learning algorithm nowadays. In support of the provincial agricultural monitoring, we have developed an approach to use GF, Sentinel 2 and Landsat 8 images to mapping crop types in irrigation area of the yellow river of Ningxia, China. Field sample photos were taken with the GPS camera in summer 2017 and 2018 respectively and thereafter the crop types for the ground truth data were interpreted with a software, named GPS Photo Data Processor. With the support of these ground truth samples, more samples for the training and validation were further visually added over a clear sky image in key crop growth stage. The Random Forest was used as the classifier for this study as many literatures have reported that the RF algorithm overperformances other algorithms in many cases, such as SVM, Maximum Likelihood. The classification results of crop type map were evaluated with the error confusion matrix, in particular, OA(overall accuracy) and F1 Score. Sentinel 2A/B images during the growing season in 2017 and 2018 were collected and processed via the Sent2Agri system. Landsat 8 images were downloaded from the USGS websites. And the GF satellite images were collected from CRESDA in China. All these data were further processed and finally spatially congruent. The performance for crop type mapping with time series of each of these 3 kinds of data sources was analyzed and compared. The preliminary results show that the accuracies were between 84-93%. The accuracy of crop type mapping with GF data was the lowest due to less bands and other limitations. The accuracy of crop type mapping with all bands of Sentinel 2A/B reached the highest due to more key bands and higher resolution while the accuracy of crop type mapping from Landsat 8 was slight lower than Sentinel 2. The utilization of huge volume of the high resolution satellite images is challenging to the researchers. The paper will give the detailed report in due time.

Area A - Board 95

CST (Crop Statistics Tool) is a standalone freeware developed by the JRC MARS project for estimating crop yields in advance of harvest using yield indicators derived from crop models, weather or remote sensing data.
For relating yield of a given crop and area with yield indicators available up to certain time in the season, CST requires past yield data, usually collected at administrative level, and indicator data for at least 10-15 years. Currently two main analyses are implemented: (1) regression analysis in which yield statistics are linearly regressed against the indicators observed at a given dekad and (2) scenario analysis whereby CST looks for the past years which are most similar to the target year according to selected indicators up to the time of forecast, and combines their yields to produce a forecast. CST also determines if a so-called technological time trend is present in the yield data, in which case the indicators are used to predict the yield deviation from this trend.
CST presents many advantages for the operational analysis of yield – indicators relationships by crop analysts: it allows importing indicator data in predefined formats and cumulating these indicators over time; for the linear regression analysis, CST produces standard statistics (including R-square, adjusted R-square, residual standard deviation, root mean square error of prediction...), confidence intervals around the forecast, statistical indicators, tests, and graphs; it also allows testing a given model over several dekads and regions in batch processing so as to determine the best model and timing of forecast.
The potential use of CST is here illustrated with results obtained over various African countries at regional level (South Africa, Ethiopia, Senegal, Burkina Faso, Algeria…) using the regression approach and easily accessible indicators such as dekadal NDVI data or rainfall estimates data derived from satellite observations. Assuming the available yield statistics are reliable, these tests show contrasted behaviours: in some countries, an operational yield forecasting system can be set up using the currently available indicators while in others, additional research on new indicators or models is needed.

Area A - Board 96

In Africa, up-to-date, reliable and detailed agricultural statistics (area, yield and production) are essential for implementing efficient Early Warning Systems (EWS). However, their yearly production relies almost exclusively on labor-intensive surveys that are both constrained regarding representativeness and reliability. Furthermore, several examples of EWS failing to identify the emergence of localized famines (eg. Ethiopia in 1999-2000, Malawi in 2001-2002 or Niger in 2005) highlights the outstanding challenges faced in detecting local food insecurity on time. In this context, up-to-date cropland and crop type maps generated using remote sensing data have often been identified as potential tools to accurately monitor food production at a local level. However, challenges are still wide in heterogeneous smallholder farming systems, like in sub-Saharian countries, mainly due to the small field size, the heterogeneous management practices, the landscape fragmentation, and the widespread presence of trees within the fields.
This study presents the first large-scale 10-m crop type map in smallholder farming systems over southern Mali region (135,000 km²). The classification algorithms relied on Sentinel-2 and Landsat 8 optical image times series acquired during the 2017 growing season and was run using the ESA-Sen2-Agri processing system. To train the random forest models and validate the maps, 6591 field polygons were collected around 28 villages spread over the study area through an intensive windshield survey (2 weeks). Finally, in order to assess the spatial reliability and consistency of the maps, this paper questions the classic map validation method (i.e. using a random split between training and validation polygons) by comparing it to an iterative “Leave-4 villages out” (L4O) approach.

When using the classic validation method, results show that the 10-m cropland mask reaches an overall accuracy (OA) of 96% and cropland F-Score of 95%. With the “L4O” validation approach, which presents less spatial correlation between training and validation datasets, these metrics decrease slightly to an average 91.2% (+-0.05)%. OA and cropland F-score of 81 (+-13)%.

The 10-m crop type map showed acceptable results (with the L4O approach) for cotton (F-score= 64 (+-13)%), rice (F-score= 51 (+-31)%) and dried cereals (F-score= 55% (+-10%) --I do not have this result, I think that you can put 55% for now and we will adjust this) (sorghum and millet). However larger differences are observed between the classic validation and the “L4O” approach. This indicates a larger model over-fitting effect in the case of multiple classes models and strengthen the need of high quality training datasets for large area mapping (across topo-sequence, crop types, climate (influencing the number of observation), agricultural practices, …).

These preliminary results are very promising to improve crop monitoring within smallholder farming systems where farmers really depend on agriculture to meet their basic needs.

Area A - Board 97