Spatially estimating the right amount of required nitrogen to manage crops in a sustainable manner is one of the major concern in precision agriculture. SENTINEL-2 observations are particularly well suited for this kind of applications due to its short revisiting time and a spatial resolution allowing to assess the intra-field heterogeneity. In this context, Green Area Index (GAI) and Canopy Chlorophyll Content (CCC) are variables of particular interest both relate to the canopy photosynthetic potential and to the amount of electromagnetic radiation reflected into space which make them accessible from S2.
Regardless the type of inversion method (Optimization, Look-Up-Table, Machine Learning), physically based algorithms generally require a significant number of simulations. Therefore, 1D radiative transfer models have been mainly used since they are computationally efficient and characterized by a low number of inputs. This eases the setting of numerical experiments and constrains the possible ambiguities between variables during the inversion process. However, these models are based on the assumption that the canopy is a horizontally homogeneous and infinite turbid medium with leaves randomly distributed. With the advances of computational performances, it is now possible to use more realistic and accurate radiative transfer models.
We propose here to demonstrate the gain in accuracy when using and inverting crop specific 4D radiative transfer models as compared to a generic algorithm based on 1D turbid medium as implemented in the SENTINEL2 toolbox. We emphasize on the distinction between the several definitions of the green/leaf/plant area index depending if all the green elements are taken into account or not, or if only the effective or true GAI is considered.
We present the setting up of the 3D architectural wheat and maize development models and the LuxCoreRender physically based and unbiased rendering engine. We first evaluate our model performances using the RAMI On-line Model Checker (ROMC). We then present the inversion algorithm that relies on the use of neural networks and was designed in an equivalent way as the generic 1D model inversion.
Results shows that 3D model inversion allows getting the most accurate as well as unbiased estimates of both the effective and true GAI as compared to 1D, especially for high LAI values. The gain in accuracy is obviously larger when considering maize which departs more from the 1D model assumption. Estimation performances are always significantly better when considering the effective GAI as compared to the true GAI. These results are confirmed when applying the inversion algorithm to actual satellite images of both wheat and maize crops.

The introduction of high-resolution Sentinels combined with the use of high-quality digital agricultural parcel registration systems is driving the move towards at-parcel agricultural monitoring. The European Union’s Common Agricultural Policy (CAP) has introduced the concept of CAP monitoring to help simplify the management and control of farmers’ parcel declarations for area support measures. This study proposes a proof of concept of this monitoring approach introducing and applying the concept of ‘markers’. Using Sentinel-1- and -2-derived (S1 and S2) markers, we evaluate parcels declared as grassland in the Gelderse Vallei in the Netherlands covering more than 15,000 parcels. The satellite markers—respectively based on crop-type deep learning classification using S1 backscattering and coherence data and on detecting bare soil with S2 during the growing season—aim to identify grassland-declared parcels for which (1) the marker suggests another crop type or (2) which appear to have been ploughed during the year. Subsequently, a field-survey was carried out in October 2017 to target the parcels identified and to build a relevant ground-truth sample of the area. For the latter purpose, we used a high-definition camera mounted on the roof of a car to continuously sample geo-tagged digital imagery, as well as an app-based approach to identify the targeted fields. Depending on which satellite-based marker or combination of markers is used, the number of parcels identified ranged from 2.57% (marked by both the S1 and S2 markers) to 17.12% of the total of 11,773 parcels declared as grassland. After confirming with the ground-truth, parcels flagged by the combined S1 and S2 marker were robustly detected as non-grassland parcels (F-score = 0.9). In addition, the study demonstrated that street-level imagery collection could improve collection efficiency by a factor seven compared to field visits (1411 parcels/day vs. 217 parcels/day) while keeping an overall accuracy of about 90% compared to the ground-truth. This proposed way of collecting in situ data is suitable for the training and validating of high resolution remote sensing approaches for agricultural monitoring. Timely country-wide wall-to-wall parcel-level monitoring and targeted in-season parcel surveying will increase the efficiency and effectiveness of monitoring and implementing agricultural policies.

In Europe, conventional tillage involves approximately 66% of the arable land [https://ec.europa.eu/eurostat]. This practice has negative impact on soil surfaces as, for instance, it enhances losses of water, nitrogenous compounds and phosphorus, and it also increases surface runoff and soil erosion (e.g. Foley et al., 2011). For these reasons, the Common Agricultural Policy (CAP) promotes the Conservative Agriculture (CA), e.g. zero tillage practices, that can prevent and mitigate soil degradation processes [https://ec.europa.eu/agriculture/envir/soil_en]. It is, however, also important to set up appropriate procedures to periodically assess the level of adoption of CA at European scale. Copernicus Sentinel-1 (S-1) and Sentinel-2 (S-2) data, systematically acquired at high spatial and temporal resolution, can enable cost effective tools to monitor tillage/no-tillage practices at large scale. Indeed, S-1 SAR data are particularly sensitive to surface roughness and its change due to, for instance, ploughing, arrowing, etc. [e.g. Mattia et al., 1997; Davidson et al., 2000]; while S-2 optical data are suitable to identify bare or scarcely vegetated soils [e.g. Serbin et al., 2009], which are exposed to tillage operations.

In the context of the SENSAGRI H2020 project [SENSAGRI.eu], an algorithm to retrieve from S-1 and S-2 data time series of tillage change maps at high resolution (e.g. 100m) is proposed. The algorithm is based on a multiscale approach applied to the temporal change of S-1 cross-polarized (VH) backscatter in order to single out changes due to agricultural practices only. The method is applied to bare or scarcely vegetated fields, which are identified through S-2 Normalized Difference Vegetation Index (NDVI) [Satalino et al., 2018]. In addition, the S-1 VH/VV ratio is used as proxy of NDVI in case of cloud cover [e.g. Veloso et al., 2017].

The paper presents algorithm and its validation over the SENSAGRI European sites in Italy, Spain and France. Ground data collected in 2017 and 2018 are used to assess the performance of the monitoring tool. Examples of tillage change maps at regional scale are presented and discussed.

References:
Davidson, M. W., Le Toan, T., Mattia, F., Satalino, G., Manninen, T., & Borgeaud, M. (2000). On the characterization of agricultural soil roughness for radar remote sensing studies. IEEE Transactions on Geoscience and Remote Sensing, 38(2), 630-640.
Foley, J.A., Ramankutty, N., Brauman, K.A., Cassidy, E.S., Gerber, J.S., Johnston, M., Mueller, N.D., O’Connell, C., Ray, D.K., West, P.C., Balzer, C., Bennett, E.M., Carpenter, S.R., Hill, J., Monfreda, C., Polasky, S., Rockstro¨m, J., Sheehan, J., Siebert, S., Tilman, D., Zaks, D.P.M., “Solutions for a cultivated planet”, Nature 478, 337–342, 2011.
Mattia, F., Le Toan, T., Souyris, J. C., De Carolis, C., Floury, N., Posa, F., & Pasquariello, N. G. (1997). The effect of surface roughness on multifrequency polarimetric SAR data. IEEE transactions on geoscience and remote sensing, 35(4), 954-966.
Satalino, G. A. Balenzano, F. Mattia, and M. W. J. Davidson, “C-band SAR Data for Mapping Crops Dominated by Surface or Volume Scattering”, Geosci. and Remote Sensing Letters, Vol. 11, Issue 2, pp. 384-388, Feb. 2014.
Serbin, G., Daughtry, C.S.T., Hunt Jr., E.R., Brown, D.J., McCarty, G.W., “Effect of soil spectral properties on remote sensing of crop residue cover”, Soil Science Society of America Journal 73, 1545–1558, 2009.
Veloso, A., Mermoz, S., Bouvet, A., Le Toan, T., Planells, M., Dejoux, J. F., & Ceschia, E. (2017). Understanding the temporal behavior of crops using Sentinel-1 and Sentinel-2-like data for agricultural applications. Remote Sensing of Environment, 199, 415-426.

The applications towards the use of precision agriculture provide solutions for problems like land degradation, over fertilization, suboptimal yield and soil carbon sequestration. In the last decade, the use of satellite images has become an indispensable source of gathering information on agricultural areas. In this aspect, Copernicus mission of the European Space Agency opened a new era in agricultural monitoring with both high-spatial resolution Synthetic Aperture Radar (SAR) and Multi-Spectral (MS) satellites called Sentinels. When SAR and MS data are used together, they provide information on physical, dielectric and spectral properties.

In agricultural monitoring with remote sensing, there are two main subjects: crop and the underlying soil. The research that focuses on crop monitoring uses satellite images to mainly obtain information on vegetative density, growth stage and canopy health. The research on soil monitoring, on the other hand, has been focusing on moisture content and roughness. However, it is important to emphasize that the crop health is closely related to the properties of the underlying soil.

This research investigates the distribution of different soil zones in fields, to further improve the precision agriculture applications on the crop and to create a foundation to measure soil carbon sequestration. Soil zoning is particularly crucial in understanding the crop dynamics in the field and the generation of optimized tasks including application maps for every field activity starting from crop to be grown, variety to be planted, soil tillage, crop drilling, fertilization, crop protection and other measures to get the most sellable calories produced from the farmland. The foundation of all optimization is soil and soil management.

To get a better understanding of soil, we used the temporal data from SAR and MS sensors to assess the temporal similarities of information carried in each pixel within a field. During the analysis, we take the threshold of 0.3 on the NDVI parameter from MS data to focus on the time points which are most likely soil. Later, we reduced the dimensionality of the dataset and applied clustering to obtain the similar zones. Specific to this study, we compared the commonly applied soil zoning algorithms with the developed state of the art approach.
In this work, we analyzed data from Sentinel-1 (SAR) and Sentinel-2 (MS) satellites on 737 fields, that are in Germany and Switzerland. The chosen Sentinel-1 data has the polarization of VV-VH and has consistent incidence angle. Both datasets are downloaded from the Google Earth Engine platform using the Python API. Before downloading the Sentinel-2 MS data, we have filtered the dataset using the Hollstein algorithm for 95% cloudiness conditions. Before the analysis, we applied the SMAC atmospheric correction algorithm, developed by CESBIO, to calculate the bottom of the atmosphere reflectance values from the top of the atmosphere reflectances.

The research and novel soil zoning conducted within the SolumScire demonstrator project is validated qualitatively with farmers. The validation made is a comparison based on the available data sets of all known and competing methods that are soil scans with electromagnetic induction, electrical resistivity, biomass-based soil zoning, yield maps, geological maps, drone maps, and farmer’s expertise. The results have been validated with soil profiles taken and show consistency to the soil profiles and soil profile-changes as observed on 80 fields. The initial results show high visual similarity to maps that are previously obtained through field surveys and meeting with farmers, In the next steps, we will investigate the relationship between the soil properties, that are being collected from more than 80 fields with the information that we have collected from the satellite data.

Rice management practices which make the sustainable use of resources more efficient are important interventions towards food security. Monitoring rice crop establishment methods (transplanting or direct seeding) using remote sensing data, can provide vital information on the type of practices as well as their spread and change over time. Direct seeding is a water and labour saving practice that is being promoted across Asia, yet most existing rice crop monitoring methods assume that the rice is transplanted and hence they may not perform as well as direct seeding becomes more popular. To improve the ability of rice crop monitoring methods, it is important to understand the differences in the multi-temporal information due to different crop establishment methods and then incorporate that into improved monitoring systems.

Crop establishment is a rapid event that cannot be easily seen with remote sensing. However, we can infer which establishment method is used from resulting observable differences in land surface characteristics such as field condition and crop development stage in terms of delayed or prolonged stages that occur over a longer time. In this study, temporal information from Sentinel-1 Synthetic Aperture Radar (SAR) backscatter was used to first detect alterations in field condition and rice growth, and then link those to crop establishment practices. Farmer surveys and field observations were conducted in the province of Nueva Ecija (Philippines) in four selected municipalities across the province in 2017, to obtain information on field boundaries and crop management practices for 61 fields. Multi-temporal, dual-polarised, C-band backscatter data at 20m spatial resolution was acquired from Sentinel-1A every 12 days over the study area during the dry season, from November 2016 to May 2017. Mean backscatter values were calculated for each rice field and SAR acquisition date. The SAR acquisition dates were selected based on the reported dates for land management activities and the estimated dates of the crop growth stages. We used a Mann-Whitney U test to study whether there are significant differences in backscatter between the two practices during the land management activities and crop growth stages.

Significant differences were observed in the early growing season, particularly during land preparation, crop establishment, rice tillering and stem elongation in cross-polarised, co-polarised and band ratio backscatter values. Our findings demonstrated that crop establishment methods could be clearly discriminated by SAR at these stages and that there is more opportunity for their discrimination than has been presented in the earlier literature.

The increased practice of dry and wet direct seeding has implications for many remote sensing-based rice monitoring methods that rely on a strong water signal during the early season to determine if a field has been (trans)planted with rice. This signal is weakened or is even non existent when direct seeding is practised. Rice monitoring systems will need to adapt so that they can still accurately monitor the cultivated rice area as new resource conserving crop management practices become more common. As well as SAR intensity information, these algorithms should also incorporate coherence, spectral reflectance and vegetation indices in a multi-sensor approach to rice crop monitoring.

Satellite Image Time Series (SITS), such as the ones acquired by Sentinel-2 (S2), combine a large amount of information, compared to previous satellite generations, since a better trade-off in terms of spatial/spectral/temporal resolutions is guaranteed. The high revisit period and the high spatial resolutions of S2 result in: i) increase of the probability to acquire cloud free images and ii) availability of detailed information for analyzing small objects (e.g., single crop fields). These characteristics become relevant in the agricultural analysis, where availability of dense SITS is required to map and analyze fast evolving crop behaviors. So far, agriculture applications have been limited by the spatial and temporal resolutions to: i) daily analysis at low spatial resolution or ii) analysis at high spatial resolution with low temporal resolution, and has been applied over large regions and mixed crops. Instead, the high spatial resolution offered by S2 allows for a better compromise, even when the crop field size is relatively small (lower than 9.4ha, as it frequently happens in Europe w.r.t. 19.6ha in US [1], [2]), with a high temporal resolution. In the literature, several methods exist that analyze the evolution of crop fields in an aggregated fashion, but none is able to work: i) at single field level; and ii) with irregularly sampled data.
This paper presents a system for the analysis of spatio-temporal evolution of crop fields in SITS that is able to deal with the spatial/spectral/temporal characteristics of S2 data. The system was developed in the framework of the S2-4Sci Land and Water – Multitemporal Analysis project [3] in the context of the Scientific Exploitation of Operational Missions (SEOM) program. This project aimed at exploiting the time variable together with the spectral and spatial resolutions offered by the S2 satellite in three different application areas: i) change detection and attribution, ii) land cover maps updating; and iii) time series analysis. Here, we present the results achieved in the last one, where the main goal was to design a system for the use of long time series of S2 data for precision agriculture. The system had to be completely unsupervised and adaptive to accurately detect fields/crops and extract relevant temporal information for precision agriculture. The main assumptions of the developed system are as follows: i) an updated multitemporal crop field map does not exist for all the areas in the world, ii) at least one cloud-free acquisition exist over the area of interest per month; and iii) the crop fields in the area of interest are cultivated with different types of crops (thus they show different phenological cycles).
The system is based on four main steps [4]: (i) time series pre-processing. It guarantees that S2 images in the time series have similar spectral and spatial characteristics among each other and that cloudy images are removed, (ii) spatio-temporal fusion generates a multitemporal vegetation map by fusing spectral, spatial and temporal evolution of Normalized Difference Vegetation Index (NDVI) time series. (iii) time series reconstruction uses a novel non-parametric regression model based on data augmentation by a dropout strategy to build regular and continuous NDVI time series at the level of single objects/crop fields, and (iv) spatio-temporal information extraction analyzes different radiometric indices to derive products such as phenological parameters and cumulative indices for statistical analysis, unsupervised and/or supervised classification of land use classification, rainfed and irrigated crop fields detection, etc.
Experimental results were carried out over two test areas that show intensive agricultural activities and different weather and orographic conditions: i) Barrax, Spain (~70 km2); and ii) Prostějov-Přerov, Czech Republic (~500 km2). The results confirmed the effectiveness of the proposed system and were further validated by interaction with experts directly working on the areas.

References

[1] C. J. Weissteiner, C. García-Feced, and M. L. Paracchini, “A new view on EU agricultural landscapes: Quantifying patchiness to assess farmland heterogeneity,” Ecol. Indic., vol. 61, pp. 317–327, Feb. 2016.
[2] L. Yan and D. P. Roy, “Conterminous United States crop field size quantification from multi-temporal Landsat data,” Remote Sens. Environ., vol. 172, pp. 67–86, Jan. 2016.
[3] “ESA SEOM-S2-4Sci Land and Water Multitemporal Analysis (MTA) | RSLab,” 29-Oct-2018. .
[4] Y. T. Solano-Correa, F. Bovolo, L. Bruzzone, and D. Fernández-Prieto, “Spatio-temporal evolution of crop fields in Sentinel-2 Satellite Image Time Series,” in 2017 9th International Workshop on the Analysis of Multitemporal Remote Sensing Images (MultiTemp), 2017, pp. 1–4.