Agriculture - National 1

Remote Sensing Applications Consultants Ltd (RSAC) and the Centre for Ecology and Hydrology (CEH) have developed Land Cover Map Plus, an annual crop map of the UK based on Sentinel‐1 SAR data, as an add‐on layer to the National Land Cover Map. Land Cover Map 2007 has just a single class for all arable and horticultural crops.

Previous research work carried out by RSAC within the ESA AgriSAR programme demonstrated the potential of multi‐temporal C‐band SAR data for crop mapping using a test site in The Netherlands. In a project co-funded by Innovate UK, methods were refined and potential accuracies demonstrated for crop mapping in the UK using a time series of Radarsat‐2 images obtained through the CSA-ESA SOAR-EU 2 initiative. In 2015, crop mapping was implemented for the whole of the UK using the first routinely acquired Sentinel-1 data and the Land Cover Map 2007 land parcel framework. Production of Land Cover Map Plus 2016 for the UK is currently in progress.

Land Cover Map Plus will provide more detail for agricultural land cover than is provided by Copernicus Land Monitoring Services, and is amongst the first products to be routinely generated from Sentinel data without Copernicus support.

Annual mapping of agricultural cropping over the whole country, as provided by Land Cover Map Plus, will have a wide range of potential applications, including the medium and long term detailed analysis of crop rotations and changing cropping patterns. Up-to-date crop maps are a vital resource for stakeholders who have an obligation to safeguard our environment. Agriculture impacts directly on biogeochemical and water cycles, nutrient balances and hydrological and biological systems: through better understanding of crop distribution, crop rotations and their downstream consequences, changes in practice might be devised and implemented to bring about reduced pollution, wildlife conservation and improved control over the spread of crop diseases. Knowledge of crop distribution will also provide policy makers and regulatory agencies with better evidence to inform planning and decision making in areas as diverse as rural development, biofuels and food security.

An overview of the crop classification methodology will be presented, together with the 2015 crop map of the UK.

During the last years satellite data with high enough spatial and temporal resolution have become available under free and open licenses. The very large volumes of these data allow providing classification maps at global, national and regional scale in operational procedures. Crop mapping and classification of agricultural crops is extremely valuable source of information for many applied problems in agricultural monitoring and food security.

We propose a classification method that consists of two parts. Clouds and shadows effect are present in optical images, particularly Landsat-8 and Sentinel-2. Therefore, at first self-organizing Kohonen maps (SOMs) are used for missing pixel restoration in a time series of optical data. In the subsequent classification step, an ensemble of neural networks, in particular multi-layer perceptrons (MLPs), are used for time series classification [1], [2]. For Sentinel-1 SAR series, only pre-processing to produce geocoded imagery is required before classification, for which we use the Sentinel-1 Toolbox. Ground truth data were collected within along the road surveys and were randomly divided for training and test samples in equal proportions. Test set was using for independent result validation.

In 2015 dual polarization (VV/VH) Sentinel-1A coverage of Ukraine with a 12 day repeat frequency (descending mode) is available for the crop growth season. The weather-independent synthetic-aperture radar (SAR) images provided by Sentinel-1A constitute a series of 15 images over the Kyiv oblast, which is one of the JECAM sites. That is much more data than Landsat-8, for which only 4 images with permissible level of clouds coverage during the same period of time (March - August) in 2015 were acquired. Using time series of optical and SAR images (Landsat-8 and Sentinel-1A) we have explored classification accuracy for the Kyiv region, for each of the data sets separately and for their combinations. Overall classification map accuracy for 11 classes, including 7 crop classes: winter wheat, winter rapeseed, soybeans, sunflower, sugar beet, maize and other cereals spring crops, based on Landsat-8 time series was 85.4%, Sentinel-1A – 91.4%, at the same time, data fusion of Sentinel-1A and Landsat-8 provides overall accuracy 92.7% that was higher by +7.3% and +1.3% for Landsat-8 and Sentinel-1 based classifications, respectively. We explain the results with the higher temporal resolution of Sentinel-1 data which can be consistently acquired due to cloud independence and the complementarity of the optical and SAR signal response from the crop types. Detailed experimental results and crop classification maps for Sentinel-1A and Landsat-8 will be presented. We will also discuss initial results for the 2015-2016 early crop season in which we integrate Sentinel-2 data.

[1] N. Kussul, S. Skakun, A. Shelestov, M. Lavreniuk, B. Yailymov, and O. Kussul, “Regional Scale Crop Mapping Using Multi-Temporal Satellite Imagery,” Int. Arch. Photogramm. Remote Sens. Spatial Inf. Sci., XL-7/W3, pp. 45–52, 2015.

[2] S. Skakun, N. Kussul, A. Y. Shelestov, M. Lavreniuk, O. Kussul, “Efficiency Assessment of Multitemporal C-Band Radarsat-2 Intensity and Landsat-8 Surface Reflectance Satellite Imagery for Crop Classification in Ukraine,” IEEE Journal of Selected Topics in Applied Earth Observations and Remote Sensing, 2015, DOI: 10.1109/JSTARS.2015.2454297.

Grassland yearly maintenance by cutting or grazing is among the many requirements in European Union (EU) Common Agricultural Policy (CAP) for being eligible for subsidies. Moreover grasslands are one of the most common agricultural land cover types in Europe, e.g. in Estonia alone about 50% of all agricultural land and 400 000 ha is grasslands. Accurate and efficient grassland maintenance checks are therefore important to assure honest and effective use of EU taxpayer money.

In April 2014 the first Copernicus programme satellite Sentinel-1A was successfully launched followed by Sentinel-2A in June 2015. In 2016-2017 the Copernicus Sentinel-1 and -2 systems will become fully operation after the launch of the B unit satellites. This free and open 10-20 m resolution optical and radar satellite data with unprecedented temporal density and excellent technical parameters has high potential for being used to improve the management of CAP. This study explores the potential to use Sentinel-1 radar data time series for grassland cutting checks. In 2015 a careful field survey supported campaign was organised and the first results are very promising. A 28 image time series stack with three different acquisition geometries covering period from May 1 to August 17 was used for the study. For support data field border vector layers were provided by Estonian Agricultural Registers and Information Board and meteorological data by Estonian Weather Service. Based on the results on our 23 field survey supported grasslands, it seems that it is not only possible to tell if the cutting took place, but also provide the likely cutting dates.

The presentation gives an overview of the 2015 study preliminary results and discusses the potential of Sentinel-1 compared with other radar sensors like TanDEM-X, RADARSAT-2 and COSMO SkyMED for the grass cutting checks application. In addition the vision for future operational service development is presented comparing alternative hardware and software solutions for the processing realisation. Even though it might be currently possible to implement the processing with desktop/workstation/ conventional server systems, the requirements and complexity as well the amount of data is increasing. Future solutions will likely exploit cloud computing with processing next to Sentinel-1 and Sentinel-2 database, GPU-based computing and other novel IT technologies.

An agricultural subsidy is a governmental subsidy paid to farmers to supplement their income, manage the supply of agricultural commodities, and influence the cost and supply of such commodities. In the European Union, the agricultural policy is called the Common Agricultural Policy (CAP) and it implements a system of agricultural subsidies and other programmes. In 2014, the total CAP budget was approximately €58 billion, of which appr. €41 billion was used for direct aids to farmers. The CAP budget forms approximately 40% of the total EU budget.

 On member state level, the CAP funds are managed by national paying agencies. These are responsible for both making the payments to the final beneficiaries and controlling that no undue payments are made. In Finland this duty is carried out by the Agency for Rural Affairts (MAVI). The controls in Finland are currently carried out as various cross checks between farmers’ applications and MAVI register data, but also as on-the-spot-checks (OTSC) in the field.

 This current system for controls is in need of development. Mavi´s resources are steadily decreasing and more efficient methods are therefore needed. The system should not only be more cost efficient but also faster, in order to enable Mavi to make payments on time. In practice this means less time spent in the field (OTSC), as this is the most costly part of the controls. This study will aid these objectives by providing information based on remote sensing for the control process. The more specific aims are to perform a crop classification on a coarse level (autumn cereal, spring cereal, grass cover), to study which is the appropriate number of remote sensing data captures required in order to be able to perform the classification reliably, crop classification at the end of the growing season and tillage of agricultural parcels. The work is carried out using Sentinel-1 SAR imagery.

So far, the fields of Viikki and Haltiala, Helsinki, have been used as test sites for the initial ideas and processing of Sentinel-1 images. 15 agricultural parcels have been visited 6 times during the growing season and plant height and species have been recorded on each visit. The Agency for Rural Affairs has also provided more extensive data (e.g. vector delineations of crop parcels, their crop species, VHR- and HR-imagery) for 1652 parcels from two areas around Loimaa and Nakkila in the South-West of Finland. Crops in the agricultural parcels were classified into 11 different classes, out of which the most common was cereals (69% of the total parcel area) consisting of barley, spring wheat, oat and autumn wheat. The proportion of grass covered parcels is 10%. The average size of parcels is 3.7 ha.

The mode of used Sentinel-1 images is Interferometric Wide swath. In order to decrease the effect of viewing angle, images only from those days with both ascending and descending orbit imagery were used. So far, images from 14 dates have been downloaded. Images have been georeferenced using ESA’s Sentinel-1 Toolbox function Range-Doppler Terrain Correction to backscatter Sigma0, using projected local incidence angle from DEM. Then, the Sigma0 images have been transformed to ETRS TM35FIN coordinate system with 20 m pixel. Finally, the average image of descending and ascending orbit images is computed, Sigma0 values multiplied by 100 and saved as integer numbers in order to save disk space.

The image in attached file presents an example of Sentinel-1 time series, VV- and VH-polarization from three different dates, from the Haltiala test site. The red areas in the VV-polarization image are parcels which were ploughed, bright green are autumn rye, dark green are grass, dark reddish blue are autumn cereals. In VH-polarization image there are some differences between grass (green) and autumn cereals (red). Another way to look at the data is to plot the parcel average backscatter as a function of the acquisition date with information about the parcel (see sencond image from attached file).

It can be seen that due to ploughing, the average backscatter is about two times higher than for non-ploughed parcels at the start of the growing season, but as plants grow, the backscatter becomes about the same. Snow melt causes considerable drop of backscatter. We also have weather information for all test sites, but have not checked the temperature and rain data with against the backscatter data.

The next steps will be to start processing data from Nakkila and Loimaa test areas, study the behavior of plant species average backscatter as a function of time and weather conditions, and make classifications.

The combined availability of free and open high resolution satellite imagery and open access to precise parcel reference data and annual crop declarations enables the establishment of a robust test framework for crop classification products. In the European Union, several Member States are releasing annual crop area declaration data, which are collected for the verification of support measures under the Common Agricultural Policy, into the public domain. A key advantage of these data sets is that they are based on the use of very detailed topographical base data, typically at scale 1:10,000 or better, and therefore of the required quality to serve as very large ground reference data sets for validation of high resolution (5-20 m) image classification products from which regional and national crop area estimates can be derived.
For example, the Netherlands national database contains around 770,000 vectors covering more than 95% of the cultivated land area. A drawback of these data sets is their release at a late stage in the growing season, usually around October.
Thus, their use is somewhat restricted to the a posteriori validation of classification results derived from the in-season imagery. Still, the sheer size of the data sets, and their geo-statistical robustness make them a desirable alternative to the usual practice of ad-hoc, limited ground data collection campaigns that are only poorly representative of national agricultural landscapes.
We use the Netherlands reference data in a series of tests with 1000s of parcels to verify classification results and derived crop area estimates of major arable crop areas, using Sentinel-1, Landsat-8 and (late season) Sentinel-2 image time series for 2015. We show that in areas with a large variety of arable crops, several image classification approaches still have considerable difficulties in achieving the accuracies that are required for estimation of inner-annual variability. We enumerate the regional variability in these results, which is largely explained by local variation in soil types and (related) management practices. Finally, we extrapolate our findings to the 2016 seasonr, for which the timely and more frequent availability of Sentinel-2 imagery should results significant improvements.