Quantifying methane emissions and rice productivity in the Mekong Delta with a simultaneous data assimilation scheme of L/C-band SAR data and ground observation

Hironori Arai 1,2*, Thuy Le Toan 1, Wataru Takeuchi 3, Kei Oyoshi 4, Hoa Phan 1, Stephan Mermoz 5, Alexandre Bouvet 1, Lam Dao Nguyen 6, Tamon Fumoto 7, Kazuyuki Inubushi 8
1 Centre d’Etude Spatiale de la BIOsphère-Université Paul Sabatier-Centre National de la Recherche Scientifique-Institut de Recherche pour le Développement-Centre National d’Etudes Spatiales (CESBIO-UPS-CNRS-IRD-CNES), 18 av. Ed. Belin, 31401 Toulouse CEDEX 9, France; thuy.letoan@cesbio.cnes.fr, thi-hoa.phan@cesbio.cnes.fr, Alexandre.Bouvet@cesbio.cnes.fr
2 Japan Society for the Promotion of Science, Chiyoda, Tokyo 102-0083, Japan; hiro.arai360@gmail.com
3 Institute of Industrial Science, The University of Tokyo, Meguro, Tokyo 153-8505, Japan; wataru@iis.u-tokyo.ac.jp
4 Earth Observation Research Center, Japan Aerospace Exploration Agency, Tsukuba, Ibaraki 305-8505, Japan; ohyoshi.kei@jaxa.jp
5 GlobEO, 31400, Toulouse, France; Stephane.Mermoz@cesbio.cnes.fr,
6 Ho Chi Minh City Space Technology Application Center, Vietnam National Space Center, Vietnam, 1 Mac Dinh Chi street, District 1, Ho Chi Minh City, Vietnam; ldnguyen@vnsc.org.vn

7 National Agriculture and Food Research Organization, Tsukuba, Ibaraki 305-8517, Japan; tamon@affrc.go.jp
8 Graduate School of Horticulture, Chiba University, Matsudo, Chiba 271-8510, Japan; inubushi@faculty.chiba-u.jp
* Correspondence: hiro.arai360@gmail.com; Tel.: +33-(0)5-61-556671


Approximately 60% of the global population is concentrated in the monsoonal Asian region, whose total land area is 20% of the global total land area; thus, there are limited land/water resources for further expansion of cultivation. In Asia, rice is the most important staple food, providing on average 32% of the total calorie uptake. Approximately 90% of the global paddy area and annual output of rice production are concentrated in monsoonal Asia. Approximately 75% of the global rice volume is produced in irrigated lowlands encompassing delta basins in Asia, such as the Mekong, Irrawaddy, Chao Phraya, and Bengal Rivers. To meet the increasing food demand derived from global population growth, the rice cropping in the region is becoming more intensive with double/triple rice cropping techniques. However, because rice production requires large amounts of water (3,000-5,000 L kg-1 rice), and its soil submersion due to the irrigation ended up causing the emission of methane which is 28 times stronger than carbon dioxide on mass basis. Consequently, about 11% of anthropogenic methane emissions come from rice paddy submerged soils, and among the major cereals, rice has the highest global warming potential due to the high methane emissions. Therefore, water-saving irrigation practices which has a potential to mitigate the methane emissions by oxidizing soil environment such as the alternate wetting and drying (AWD) are desirable to be disseminated in this region, to ensure sustainable water demand while lowering the greenhouse gas emission. To realize transparent carbon accounting, a robust method quantifying the dissemination status of the mitigation measures with high spatio-temporal resolution is required.
We developed a digital twin system quantifying greenhouse gas emission and rice productivity on every SAR data pixel basis by assimilating L/C band SAR observation (ALOS-2 and Sentinel-1) into GEORICE-pixel based DeNitrification-DeComposition-rice model (DNDC-rice) and validated it with ground observation data collected in the Mekong delta. To assimilate L/C band SAR signals with adequate observation error assessment, General-Purpose computing on Graphics Processing Units (GPGPU) based local particle filter which consists of rapid SAR data pre-processing was built. By optimizing parameters of DNDC-rice, which represents local farmers’ activities (e.g., irrigation frequency, standing water height) on every SAR data pixels adaptatively, we conducted an evaluation of irrigation status and conducted validation with ground observed field water levels (r=0.82, n=282, RMSE=3.77 cm). Relatively high field water level region which still has a potential to mitigate the emission by implementing intermittent irrigation was identified in relatively low elevation area ( 3m< ). The assimilation results also showed that about 22% of the total rice paddy areas in the delta were continuously flooded or irrigated before water level drops to -5 cm. The validation results showed the consistent spatio/temporal distribution of field water level with ground observation. In this presentation, details of the data assimilation system, the characteristics of back scattering signature and its performance of DNDC-rice model optimization will be introduced to discuss the future development with Sentinel-1/ALOS-2,4/NISAR/ROSE-L data integration.

Agriculture is at the crossroads of many challenges going from food security to biodiversity. Among those, climate change is one of the most pressing. Contributor to greenhouse gas emissions, agriculture has also a great potential for mitigating climate change, for instance through soil organic carbon (SOC) storage resulting in the 4/1000 initiative following the COP21 (Minasny et al., 2017). In order to guide the transition to sustainable agricultural practices and quantify their effects, comprehensive and scalable tools for regional monitoring of the local impacts of agronomic practices on soil organic carbon storage are needed.
In this context, assimilation of remote sensing data in agronomic models is a promising way to retrieve agro-environmental indicators (biomass production, water requirements, CO2 fluxes) meeting the needs of precision farming and large-scale monitoring of agroecosystems carbon budgets components. However, agronomic modelling exercises covering all of these processes are either limited to demonstrations over fairly small areas (hundreds of square km) at high resolution, or low to medium resolution over large areas, due to computational constraints. This curbs their potential to inform policy makers at regional/national scales of the country in the first case, and is inadequate when aiming to
monitor the impact of management at plot/farm scale in the second. Furthermore, most of these approaches when applied over small areas provide plot scale estimates, while an intra-plot estimate can be needed to represent process heterogeneity. This is especially true in regions with high soil variability or crops sensitive to environmental variations. Intra-plot scale estimates meet the needs of precision agriculture, while allowing validation of the cropland carbon budget components (measurements and estimates being performed at a consistent spatial scale).

In this study, we will address those challenges by providing an adequate scalable solution to estimate the daily CO2 fluxes and the comcesbiocesbiocesbiocesbiocesbioponents of the carbon budget at high spatial resolution. The SAFYE-CO2 agronomic model (Pique et al., 2020a, 2020b) is applied at 10x10 m² resolutions, over an entire Sentinel-2 tile that covers a 110x110 km² area. SAFYE-CO2 is a simple parsimonious agronomic model that relies on the Monteith method to represent photosynthesis, on Amthor, 2000 to estimate autotrophic respiration, and on the FAO-56 method for water budget and evapotranspiration estimates. Former applications of SAFYE-CO2 relied on iterative hard to scale least square error reduction optimization functions to assimilate remote sensing derived LAI time series, in order to calibrate light use efficiency and the phenological parameters in the model. Here, the previously inconceivable large scale exercise is made possible through a spatialised assimilation strategy, that combines the use of importance sampling and look up tables at the resolution of the weather forcing data. This strategy and the SAFYE-CO2 model have been implemented in the Agricarbon-EO agronomic modelling platform. First, it provides 10m resolution multitemporal LAI data derived from Sentinel-2 images through a bayesian inversion of the prosail radiative transfer model. Second, the LAI is assimilated at pixel level in the SAFE-CO2 model to provide dry biomass and daily CO2 flux estimates. The Agricarbon-EO modelling platform includes automated connections to the ERA5, Theia and Global soil map api’s to extract relevant data and to manage the data flow between the different modules in a parallel computation setup. This architecture allows an end-to-end comprehensive solution. End-to-end designates here the process that goes from the download of satellite reflectance products to production of maps such as the harvested yield map or the cumulated CO2 flux maps.

Agricarbon-EO is applied to an agricultural area located in the South-west of France. This zone is characterized by a hilly landscape with medium size plots(5 to 50 ha) and a high degree of soil variability due to post glacial erosion in the Garonne river bassin. It is also a highly monitored zone with 2 ICOS flux sites (Lamasquère, Auradé) managed by the Regional Spatial Observatory (ORS) that provides regular destructive biomass and LAI measurements on a network of plots near the ICOS towers.

Results show that we are able to process about 300 pixels/min compared to 0.2 pixels/min in the previous iterative approach, which gives a x1500 computational gain. Moreover this approach can be parallelised with a less than 5Gb RAM load per process which allows efficient deployment on modern HPC. Most importantly, the accuracy of the predictions for winter wheat are not hampered by this computational gain (RMSE of 200 g.m-2 R²=0.92 for dry aboveground biomass as in Pique et al., 2020). In addition, the new approach provides not only the estimates but also their associated uncertainties (we observe a coefficient of variation +/-0.2 for dry aboveground biomass for example) which are needed for the elaboration of new guidelines for improving SOC storage and for the carbon market.

This study illustrates the potential of pixel scale agronomical modelling based on the assimilation of LAI time series at full Sentinel-2 resolution and over large areas while 1) maintaining statistical performance obtained at plot level in previous modelling exercises on the components of the carbon budget (yield, biomass, CO2 fluxes) and 2) providing local estimates of their uncertainties. These improvements pave the way to more accurate and comprehensive monitoring of the components of the cropland carbon budgets and other agro-environmental indicators.

Références:

How to measure, report and verify soil carbon change to realize the potential of soil carbon sequestration for atmospheric greenhouse gas removal - Smith - 2020 - Global Change Biology - Wiley Online Library [WWW Document], n.d. URL https://onlinelibrary.wiley.com/doi/full/10.1111/gcb.14815 (accessed 12.9.21).https://doi.org/10.1016/j.geoderma.2020.114428
Minasny, B., Malone, B.P., McBratney, A.B., Angers, D.A., Arrouays, D., Chambers, A., Chaplot, V., Chen, Z.-S., Cheng, K., Das, B.S., Field, D.J., Gimona, A., Hedley, C.B., Hong, S.Y., Mandal, B., Marchant, B.P., Martin, M., McConkey, B.G., Mulder, V.L., O’Rourke, S., Richer-de-Forges, A.C., Odeh, I., Padarian, J., Paustian, K., Pan, G., Poggio, L., Savin, I., Stolbovoy, V., Stockmann, U., Sulaeman, Y., Tsui, C.-C., Vågen, T.-G., van Wesemael, B., Winowiecki, L., 2017. Soil carbon 4 per mille. Geoderma 292, 59–86. https://doi.org/10.1016/j.geoderma.2017.01.002
Pique, G., Fieuzal, R., Al Bitar, A., Veloso, A., Tallec, T., Brut, A., Ferlicoq, M., Zawilski, B., Dejoux, J.-F., Gibrin, H., Ceschia, E., 2020a. Estimation of daily CO2 fluxes and of the components of the carbon budget for winter wheat by the assimilation of Sentinel 2-like remote sensing data into a crop model. Geoderma 376, 114428. https://doi.org/10.1016/j.geoderma.2020.114428
Pique, G., Fieuzal, R., Debaeke, P., Al Bitar, A., Tallec, T., Ceschia, E., 2020b. Combining High-Resolution Remote Sensing Products with a Crop Model to Estimate Carbon and Water Budget Components: Application to Sunflower. Remote Sens. 12, 2967. https://doi.org/10.3390/rs12182967

The adoption of regenerative agricultural practices is a key component to enhance soil health and in the sequestration of carbon to mitigate the impacts of climate change. Tillage is a crucial agricultural practice that can have adverse environmental impacts, but where more sustainable approaches are possible and gaining traction. Significant adverse environmental impacts caused by conventional tillage practices include acting as a net carbon source, contributing to higher rates of nutrient runoff, and soil erosion [1]. Switching to reduced tillage or no tillage approaches can both significantly enhance soil organic carbon and have a positive impact on productivity with improved soil health [2]. Today, many governments and third party organisations have introduced incentive schemes to encourage these practices. These evolving programs have created an opportunity to build Measurement, Reporting, and Verification platforms that are largely driven by satellite observations to provide cost effective, transparent, and robust metrics as a solution to enable monitoring of impacts at scale.

To help scale these incentive schemes, we have developed a deep learning approach that utilises multi-source imagery from the Sentinel-1 and Sentinel-2 constellations to verify field level tillage practices. Using a dense network of training data, we produced a classification model which determines field level tillage practice using three categories; conventional, reduced or no tillage, keying off residue levels. Our work focuses on two large regions: the UK and the corn belt region of the USA. In each of these regions we have collected over 10,000 unique ground truth data points, providing detailed measurements and observations related to the state of the field following a tillage event. This data and the corresponding satellite imagery has been used to train and validate our models to a level of accuracy suitable for large scale monitoring. Interacting factors such as soil type, moisture and weather are known as having impacts on reflectance values in imagery and therefore may impact negatively on model performance. Furthermore, practices can vary according to crop type, and tillage types can be classified differently according to regional standards; thus a system must be flexible to address many end users. We will present our results from these regions, discuss how we have addressed some of these challenges, and suggest future developments to aid in the wide scale adoption of these regenerative practices.


References
[1] Conant, R.T., Easter, M., Paustian, K., Swan, A. and Williams, S., 2007. Impacts of periodic tillage on soil C stocks: A synthesis. Soil and Tillage Research, 95(1-2), pp.1-10.
[2] Holland, J.M., 2004. The environmental consequences of adopting conservation tillage in Europe: reviewing the evidence. Agriculture, ecosystems & environment, 103(1), pp.1-25.

Improvement of agricultural forecasts and decision support for agricultural management are some of the main goals of ESA’s Sentinel-2 satellite constellation. With the launch of Sentinel-2A and Sentinel-2B very high-quality data with a high temporal return rate are now available to scientists and service providers. Therefore, in recent years, service providers developed various satellite-based tools to assist farmers in their management decisions. The developed services differ greatly in terms of the complexity of the algorithms used and the transferability to different regions.
The TalkingFields® services (www.talkingfields.de) were originally developed in an ESA ARTES IAP Demonstration Project, with the first product being introduced to the market in 2011. Since then, TalkingFields has been continuously available on the market, adapted to Sentinel-2 and the service offer has been significantly expanded. In 2021, products offer a range of services that support farmers worldwide in their decision making with high-quality, well-validated approaches. We use the radiative transfer model Soil-Leaf-Canopy (SLC) in combination with the crop growth model PROMET to make the best use of the wide range of information available by Sentinel-2, combined with data from the Landsat missions and ancillary information. Particular attention is paid to providing the farmers with concrete instructions for action. The statements made go beyond qualitative information and instead give quantitative site-specific recommendation how to supply crop stands with water and nutrients by calculating the crops’ concrete requirements. For example, the exact amount of water in mm needed for the coming week or the required amount of fertilizer in kg N/ha for the upcoming application can be provided.
While the potential benefits for the individual farmer through cost savings and optimized crop production are already a very well-established argument for the use of SmartFarming services, the environmental aspect has recently also come to the fore. Site-specific fertilization not only helps to maintain high yields in a sustainable way and within the legal limits of fertilizer application by making optimal use of the fertilizer through precise distribution. Site-specific fertilization also helps to avoid N-leaching into the groundwater by supplying plants in a targeted manner. The optimized plant supply also protects the soils and helps to avoid a depletion of humus by under-fertilized plants. Humus is very valuable both as a water and CO2 reservoir.
We will present the TalkingFields® Nitrogen fertilization service NManager Pro, in specific results of N-fertilization trials on rapeseed fields in Germany that demonstrate the high economic and ecological value of site-specific fertilization of rapeseed. In the year 2020 NManager Pro was used on four test sites of different farms in southern Germany. The total yield of the test-fields could be increased on average by 10% in comparison to constant application of fertilizer while maintaining high crop qualities. The increase in yield could be achieved with the same amount of management input and fertilizer as with conventional management. The trial was repeated in 2021 on one field at one of the test farms. Results of both trials, of 2020 and 2021 will be presented.

Figure 1: N-demand delivered by TalkingFields® NManager Pro for the 2nd spring application of one of the test fields in 2020

Additionally, using the TalkingFields® service palette, yield maps were simulated based on crop growth modelling and leaf area observations over the whole growing cycle derived from Sentinel-2. Leaf area retrieval in rapeseed is a special challenge due to the very bright yellow flowers of the crop that for example have a very strong impact on the NDVI signal. This impact is also critical for the radiative transfer model inversion technique that is applied in TalkingFields® for crop parameter retrieval. To compensate this effect the SLC model was specifically adapted to the rapeseed case by changing the leaf optical properties.
With the availability of the yield map, the impact of the site-specific fertilization can be assessed spatially distributed. It allows the calculation of maps for the nitrogen efficiency of the fertilization, an important indicator of sustainable agricultural production.
In TalkingFields® NManager Pro satellite data are used to monitor nitrogen uptake of crops. Site-specific farming is however also possible using remote sensing sensors mounted on a tractor. This tractor-based method was established more than 10 years ago. With the availability of new services using Sentinel-2 the service became more efficient and easier to use by the farmer. Further, no investment in sensor equipment is required. For rapeseed fertilization the specific situation exists that the monitoring period for the nitrogen uptake is mainly in the winter months when no other farming practices are required. To evaluate nitrogen uptake over the winter with a tractor-based system, the farmer must travel to the fields to check the condition of the crops without any other tasks to perform on the land at that time. This has negative impact on the CO2 footprint of the crop production due to the additional fuel consumption of the tractor as well as on soil quality due to compaction.
With the improved nitrogen efficiency also a lower carbon footprint per ton of yield is achieved. Thus, smart fertilization techniques also serve our common goal for CO2 emission reduction.

In India, the world’s 2nd biggest cotton producer, 30-50% of the total chemical pesticide and 30% of the water for irrigation is used for cotton cultivation only. Traditional flood irrigation systems result in significant loss of water, with pesticides and fertilizers draining into water resources. Male workers are contaminated while spraying, and female ones while preparing pesticides or picking cotton. In India alone a total of about 60 million people, including 4.5 million farmers, depend on cotton for their livelihood. These circumstances led in recent years to initiatives for more sustainable cotton production to improve the livelihood and the income of farmers, as well as to minimize environmental impacts by land producing greener cotton, as defined in the Better Cotton Initiative (BCI) criteria. Here, agricultural monitoring through Earth Observation (EO), meteorological and soil data can provide important insights on the compliance with these criteria. This has been carried out in the recent feasibility study, funded by the European Space Agency, called „Greener Cotton“, aiming for the integration of EO and meteorological data for tracability and supply chain platforms. Especially the multispectral and multitemporal sensing capabilities of the Copernicus satellite constellation enables near-real time monitoring and agricultural information extraction for cotton parcels.
The core methodology is targeting to integrate operational data science and machine learning algorithms based on EO (Sentinel-1 and Sentinel-2) and meteorological data with six different service blocks: a) extraction of phenological markers from vegetation index time series, b) verification of cotton growth patterns based on machine learning classification algorithms (e.g. Random Forest) to quantify crop rotations for a more sustainble soil management, c) drought and water use efficiency monitoring from MODIS data, in order to support BCI water stewardship criteria, d) cotton yield and harvest date estimation based on supervised regression of multitemporal SAR and optical data, e) cultivated area delineation via clustering of Sentinel-2 data at Peak of Season and f) GIS-based compliance checks with BCI geospatial constraints. The latter service utilizes open geodatasets such as Key Biodiversity Areas, Intake Forest Landscapes, Water Body Data and IUCN Protected Areas, to enable checks and controls for BCI land conversion. This feasbility study was carried out successfully in Maharashtra and Gujarat, India.

Monitoring of land-use and vegetation cover is essential for estimating the effects of farming practices on the environment. The great diversity in winter cropland management and the variability in climatic conditions result in a wide variety of land-use patterns with specific
environmental impacts. For instance, cover crops, often composed of a mix of species and grown between the harvest of a cash crop and the sowing of the following one (Schulz et al., 2021), can limit soil degradation (e.g. erosion) runoff, nitrates leaching and increase the soil
organic carbon content (Kaye and Quemada 2017; Pellerin et al. 2019) but also the biodiversity or the climate cooling by increasing surface albedo (Lugato et al. 2020). Their development is a key issue to achieve sustainable agriculture (Fasona et al., 2005).

In this context, remote sensing data and more specifically satellite images from the Sentinel Copernicus program have emerged as an opportunity to identify, monitor and characterize phenological status and development of vegetation over large extents (Veloso et al., 2017). Despite this, only few studies have illustrated the potential of Sentinel data for mapping cropland status during the fallow period. Schulz et al., (2021) showed the potential of Sentinel-2 NDVI time-series to map two cover crop classes (catch crop and non-catch crop) with a mean accuracy of 84%. Also, the potential of Sentinel-1 SAR data , for winter land-use classification and winter vegetation mapping in five classes over small areas was illustrated by Denize et al., (2019) and Minh et al., (2018). They respectively achieved a F1-score of 0.70 and an overall accuracy of 98%. Finally, the European project SEN4CAP (The Sentinels for Common Agricultural Policy), illustrated the potential of Sentinel’s data for monitoring vegetation cover over large extent (http://esa-sen4cap.org/).

Still, to date, there is a lack of data and mapping of land-use typologies for cropland in the fallow period overlarge areas and the potential of Sentinel’s data for overcoming this issue should be investigated. The current lack of broad-scale research on this topic is due to (i) the winter meteorological conditions which limit remote sensing acquisition and (ii) the absence of reliable and exhaustive data (training and validation samples) for the development of classification models.

Thus, the objective of this study was to assess the ability of Sentinel-1 and Sentinel-2 time-series to map land-use typologies for cropland in the fallow period over the French metropolitan territory. To meet this objective, a five-steps methodological approach was implemented:

● Collection of ~ 4 000 fields data collected in several agricultural regions in France.

● Definition of a typology in 9 classes to represent the diversity of the cropland land-uses in the fallow period between two main crops.

● Pre-processing and processing of Sentinel’s data to create time-series parameters and field data to create training and validation samples.

● Evaluation of the best classification approach to map cropland land-uses in the fallow period (algorithms and satellite data).

● Application of the best classification procedure over the whole French metropolitan territory. Post-processing was performed using LPIS field’s borders to maintain only UAA areas.

The results show that the use of Sentinel-2 time-series alone or in combination with Sentinel-1 time-series, can produce accurate land-use typologies map for cropland in fallow period using Random Forest with a Kappa index of 0.82 and 0.83 respectively and a F1-score of 85% for both. For the national map, only Sentinel-2 datas were used due to the ratio of accuracy improvement / processing time.