Agriculture monitoring and yield estimation are important for food security, both globally and locally. Understanding year-to-year variability in crop yield and it’s relationship to meteorological conditions is especially important for countries and regions where the yield losses due to environmental factors are already observed and further predicted under the impact of climate change. One of such countries is India, where wheat is a staple food for a large part of the population and where the wheat production is the second largest in the world. Wheat yields in India have been steadily increasing since 1960s and 1970s due to benefits of the Green Revolution. However, in recent years the wheat yield has been unstable as major crop yield losses were attributed to unfavorable meteorological conditions. Including various environmental variables in the crop yield modeling can be challenging, as their impact on the final yield is complex and varies depending on the crop growth stage. For example, moderate rainfall is beneficial for crops, but extensive and untimely rainfalls can lead to huge yield losses, as in 2015, which points out the importance of accounting for such non-linear processes.

In this work, we exploit the relationship of meteorological conditions during the whole growing season simultaneously with the satellite data on vegetation, and the wheat yield in the Wheat Belt in India. We use GLDAS 2.1 data as the meteorological input and MODIS data for the vegetation remote sensing input. We show that adding satellite information on crop is crucial for yield estimation, as it carries information on both crop phenology, as well as on the crop response to the meteorological conditions. We apply the convolutional neural networks (CNNs) that can model non-linear processes and can extract important features on their own, without prior knowledge or human effort in feature design. This also means that we do not force assumptions on which time is the most important for the final crop yield. By doing so, we can include in the analysis the whole time series of multiple input variables at a high temporal resolution (4 days), and model nonlinear interactions among them. This is especially important for the studied region, as the timing of wheat planting and ripening vary across the Indian counties and inter-annually due to farmer practices and various meteorological conditions.

Furthermore, we analyze the CNNs in terms of important features and important time windows for yield estimation, which shows that these vary across space and time. By combining meteorological and satellite vegetation data with CNNs this work may help to disentangle the complex interactions between the features in the time series of the input data and the wheat yield.

As the global demand for agricultural commodities increases, major agricultural agencies have developed Crop Yield Forecasting Systems (CYFS) to better monitor and forecast crop production. These systems cope with the consequences of production variability in terms of food security, market development and trade. As of 2019, major systems are little or never based on Earth observation (EO) data, mostly because they use only low spatial resolution sensors (250m-1km), unable to capture field-specific observations and therefore to provide a valuable crop monitoring information. The spatial and temporal resolutions of the Sentinel-2 mission has changed the perspective and may revolutionized the use of EO data for crop monitoring. This study aimed to design an innovative methodology for a Sentinel-2 (and Landsat 8)-based CYFS to monitor crop yield at subnational level.
The developed system relied on a large number of yield metrics calculated for each crop field based on: (i) Leaf Area Index (LAI) derived from the Harmonized Landsat Sentinel-2 (HLS) data set (Claverie et al. 2018, RSE) and (ii) meteorological data sets (air temperature, radiation, precipitation, potential evaporation) derived from the NCEP NARR (for US) and the ECMWF ERA-Interim (for Europe) data sets. In addition, two calibrated crop growth models were used in this research: (i) the complex and mechanistic WOFOST model (de Wit et al. 2012, AFM) and (ii) the semi-empirical SAFY model (Duchemin et al. 2008, AWM). For each crop field, most relevant crop parameters of both models were calibrated using LAI data in order to derive model-based simulated yields that served as additional yield metrics. All yield metrics were finally averaged at subnational level and confronted with the NASS and MARS yield statistics. The yield estimation model relied on calibrated multi-linear regressions between aggregated yield metrics and subnational yield statistics.
The system was tested over fives US states (Kansas, Washington, Oklahoma, Colorado, Montana), two European countries (Belgium and France) and four seasons (2014-2017). The seasons 2014-2016 were used to calibrate the multi-linear regressions, while the season 2017 served for evaluation. The correlation analysis of individual yield metrics versus yield statistics allowed to discriminate non-relevant indices and to maintain only those showing a r² higher than 0.4. Performances of the system to predict yield after harvest showed relative uncertainties obtained over season 2017 were around 5% at state (for US) or national (for Europe) level, and around 10% at county level. The performances of such EO-based CYFS were unmatched to this day at this scale.
Nonetheless, the analysis raised some limitations. The length of the EO archive (four seasons for Landsat 8, two seasons for Sentinel-2A) did not provide enough substance of the capacity of the system to estimated yield in extreme cases. Then, the use of crop yield data at subnational level and the aggregation process reduced the number of data and the spread of the data. Despite the fact that the model was able to map yield estimation for each crop field, no validation was done so far due to lack of local yield data.

Rice is the primary food crop for more than half of the world’s population and responsible for over 50% of the daily calorie intake for most Asian people. Asia’s biggest rice producers are facing increasing pressure to ensure food security due the consequences of population and economic growth. At the same time demand for rice imports is increasing and agricultural areas are confronted with urban encroachment and the limits of yield increase. Paddy fields in Asia are often located in coastal lowlands and river deltas, which are vulnerable to sea level rise, floods and droughts, salt water intrusion and other hazards. The floodplains of coastal river deltas in Asia are among the most productive rice producing regions in the world and are experiencing a reduction in rice yield due to increased salinity in the ground water as a result of sea-level rise. Despite the importance of knowledge about rice production the countries official land cover products and rice production statistics are of varying quality and unable to provide timely information about the quantity of rice production.

The growth of rice plants can be monitored with time series from Synthethic Aperture Radar (SAR) sensors. In regions with high rice production the rice plants are commonly grown under agronomic flooding, where rice is transplanted or seeded into flooded fields, followed by the emergence of the plant through the water surface and a rapid increase in biomass. The application of SAR time series to monitor the growth of rice plants has been limited by a lack of sensor data. Since the launch of Sentinel-1A and 1B the data availability increased massively and the access to this free and open archive allows for the creation of temporally dense time series for the monitoring of rice growth in Asia.

We present a study which used Sentinel-1 SAR time series to estimate rice production in the Mekong Delta, Vietnam, for multiple years and intra-annual rice growing seasons. For each growing season the paddy rice area was mapped by generating object based time series utilizing a superpixel segmentation algorithm and a phenology based decision tree classifier. The rice yield and production was estimated per object with a random forest regression model trained on per-season yield data. Yield reference data was collected by surveying 357 rice farms distributed over the Mekong Delta. The results show good agreement with the reference data and rice production data published by Vietnam’s Statistic Offices at the district level and point towards the potential of Sentinel-1 time series for transparent and timely, post-harvest, rice production estimation.

A consortium led by the Rothamsted Research, working with Agricompas and Pixalytics in the UK alongside CIAT (International Centre for Tropical Agriculture), Cenipalma (Colombian Oil Palm Research Center), FEDEARROZ (Colombian National Rice Growers Federation) and Solidaridad in Colombia to develop an Ecological Production Management Information System (EcoProMIS). This 38-month project is supported by a grant of £3.9M from the UK Space Agency’s second call of its International Partnership Programme (IPP), a £152M multi-year programme funded by the Global Challenges Research Fund launched in 2015, which promotes sustained partnerships to deliver deep and impactful change.
The project aims to support Colombian rice and oil palm farmers in improving productivity and stabilising incomes while producing responsibly through mitigating climate change and protecting the environment. The longer-term project outcome being an Agricompas Ltd led public-private partnership of farmers, rice and palm growers’ federations, research institutes and companies that provide near real-time knowledge.
State-of-the-art field data collection technologies will be used to feed crop models, which will provide information on crop performance and help establish the gaps between the current situation and potential yields on an automated basis. Satellite and drone Earth Observation (EO) data combined with environmental and crop data will additionally be used to understand the impacts of crop and ecosystem management practices on biodiversity, greenhouse gas emissions, and the socio-economic conditions of stakeholders.
Initial EO activities have focused on developing Leaf Area Index (LAI), Land Surface Temperature (LST) and Normalized Difference Vegetation Index (NDVI) products from the Copernicus Sentinel-2 and USGS/NASA Landsat-8 satellite missions, to both assess productivity and the status of crop development. The satellite EO products will be compared to drone derived products and ground-based measurements to provide an understanding of crop´s anomalies and uncertainties. This presentation will focus on the preliminary results obtained from the initial data collection exercise carried out during the 2018/19 growing season.

Yield estimation and forecasting, at the field scale, would allow farm managers to plan their agronomic operations, e.g., sowing, tillage or fertilization, on the basis of expected yield potential. At the district scale, it would be useful for public and private organization, for commercial or planning purposes. In order obtain in-season crop yield predictions, biophysical variables retrieved from remotely sensed data can be assimilated into crop growth models. Data assimilation is a group of methods allowing to combine in an optimal way different information types, dynamically integrating mathematical descriptions of processes embedded into deterministic models and physical information obtained from observations. It allows to update model simulations with observed data, e.g. from remote sensing, within a systematic estimation structure. The most advanced assimilation methods include a framework for the incorporation of model and observation errors and the quantification of prediction uncertainty.
The Ensemble Kalman Filter (EnKF) method has been shown to be particularly suitable for the assimilation of remotely sensed data into crop models and has been extensively assessed for this purpose. The efficiency of EnKF has proved to be affected by a range of factors, such as the number of observations ingested into the process. Many studies have shown that using a higher number of observations brings about an improved accuracy of estimation. Other factors such as errors and uncertainties in the remote sensing observations, the variables that are retrieved as well as crop models formulation errors and parameters uncertainties play also an important role.
The project presented in this study, carried out in the framework of the Dragon 4 Programme supported by the European Space Agency (ESA) and the Ministry of Science and Technology (MOST) of China, aims at targeting these methodological issues, affecting the operational application of data assimilation schemes to the precision farming context.
A study was carried out in Maccarese (Central Italy) for the purpose of assessing the effect of temporal and spatial resolution of multispectral satellite imagery on the accuracy of durum wheat yield data estimation at the field scale from EnKF data assimilation. Field campaigns to measure biophysical variables were carried out from January 2018 to April 2018, for different growing conditions of the crop, at dates close to Sentinel-2 and Venus satellite acquisitions. The latter is a French-Israeli mission combining a high spatial (5 m) and temporal (2 days) resolution, with spectral bands quite similar to Sentinel-2. Initially, an assessment of different hybrid methods, i.e. based on the training of non-parametric regression algorithms with PROSAIL model simulations, was made, in order to identify the most accurate technique for the retrieval of biophysical variables from satellite imagery. Subsequently to a validation study with ground data, the retrieval technique was applied to all the cloud-free images available for the wheat season, notably 25 for Sentinel-2 and 26 for Venus. Among all the biophysical variables, fractional vegetation cover (FCOVER) was estimated with a smaller error (RRMSE of 9.5%) as compared to LAI (RRMSE of 19%). The former variable was thus employed for the data assimilation using the Acquacrop crop model developed by the Food and Agriculture Organization of the United Nations (FAO), since this model uses FCOVER as the main dynamic crop state variable. Recently an open source code version, Aquacrop-OS, has been released, thus allowing the application of sequential data assimilations methods, such as the Ensemble Kalman Filter, which was previously impossible.
EnKF was thus applied to Aquacrop-OS using a) only observations from Sentinel-2, b) only observations from Venus, or c) observations from both satellites, for durum wheat fields of the Maccarese farm. The results, in terms of grain yield prediction, were compared to fields where yield map data had been collected from the combine harvester at the end of the season. In general, benefits from an increased spatial and temporal resolution for field scale yield prediction mapping could be observed, but these came at the cost of an increased computational demand, therefore suggesting that a balance would be required for an operational application in the context of precision agriculture.

Pierre Guillevic (1, 2), Jean-Claude Roger (1, 2), Inbal Becker-Reshef (1), Varaprasad Bandaru (1), Alisa Coffin (3), Andrew French (3), Jerry Hatfield (3), Michael Humber (1), Sergii Skakun (1, 2), Belen Franch (1, 2), Matt Hansen (1), Brian Thomas (4), Eric Vermote (2), Chris Justice (1), César Izaurralde (1), Jan Dempewolf (1), Catherine Nakalembe (1), Michelle Cryder (3), Charles Sanchez (5), Jaehak Jeong (6), Carlos Di Bella (7), Maria Eugenia Beget (7), Siza Tumbo (8), Clement Adjorlolo (9), Pierre-Louis Frison (10)

(1) University of Maryland, Department of Geographical Sciences, (2) NASA GSFC
(3) USDA ARS, (4) University of Pittsburg ,
(5) University of Arizona , (6) Texas, A&M,
(7) INTA, Argentina, (8) SUA, Tanzania,
(9) SANSA, South Africa, (10) University Paris-Est, France


The objective of this NASA project is to develop an integrated modeling platform using Earth Observations to simulate large-scale hydrological processes and food production in support of water use decision and agricultural policy. The decision tool will provide timely and relevant assessments of crop growth, agricultural water use and available water supply from local to regional scales, with the capability to predict the impact of different climate and water resources management scenarios on food production. The approach is based on the use of field and Earth observations to calibrate an eco-hydrological model, and climate simulations to forecast crop condition and yield through the crop season. The expected primary outcomes and results for stakeholders is potential best water management practices to improve water use efficiency and prevent excessive losses of water and nitrogen to the environment. These will be provided to in-country end-users within various relevant Ministries and Departments, in order to inform regional and national decisions. Changing precipitation patterns, droughts of increasing duration and severity and growing demand for food have significant implications on water management practices, spurring an increased need for irrigation in a sustainable manner and better monitoring of water resources for the agricultural sector in general. Understanding the context and water resource constraints that stakeholders and smallholder farmers face in climate sensitive regions is critical to target and understand the potential impact of water use management. We focus on a diverse set of regional pilot cases in Argentina, South Africa, Tanzania and the United States that are representative of a range of agricultural systems, water issues and climate impacts. The application is integrated in the GEOGLAM project, an existing widely-used web-based platform which will aid synthesizing the project outcomes and expert domain knowledge to provide timely and actionable water resources and crop condition information to assist local, regional and national water management decisions. At the end, the application will be interactive and will offer the end-user organizations the possibility to submit new water resource management scenarios online for impact evaluation. The project is funded by the Water Resources NASA Applied Sciences Program. This presentation will provide an overview of the project objective and methods used to assess water resources and will present first results.