There are now more than 4000 field cover transect sites collected across the Australian continent that provide estimates of cover type (green, non-green and bare) by stratum (canopy, midstory and ground) over a nominal one hectare area. These have been collected by multiple government agencies and NGOs since 1999 for various purposes including the local assessment of ground cover, the measurement of tree cover and for calibration and validation of satellite imagery, however, as they all have the same methodological approach underpinning their collection, they are largely inter-comparable and free of observer bias.

These field data have driven the development, cal-val and broad uptake of Landsat-based products that estimate the per-pixel biophysical proportion of bare, green and non-green vegetation for every image in the time series. These these Landsat-based cover data sets are now operationally used for quantifying and monitoring grazing land condition, assessing management initiatives and have been integrated with tools to understand land condition and degradation.

However, the spatial arrangement and temporal dynamics of cover are additional important indicators of grazing land condition for both productivity and water quality outcomes. In cloudy locations, there are also significant benefits in having a 5 day revisit as provided by Sentinel 2. And the higher spatial resolution provided by the Sentinel 2 MSI affords us the opportunity to detect and report on the spatial arrangement of cover and its links with land condition.

This paper reports on the development of a native Sentinel 2 based fractional cover model, and compares this to a Landsat based model applied to Sentinel 2 using spectral band transforms. We then use these Sentinel fractional cover products to determine the cover-based spatial autocorrelation to derive metrics of land condition at a statewide scale.

The inclusion of the additional Sentinel 2 red-edge bands in the fractional cover model provided a modest improvement in the discrimination of senesced cover on bright soils, and helped in the separation of wet dark soils over the current Landsat based model (Figure 1 - Sentinel 2 Bands 11-8-4 false colour image (left) and Bare-Green-NonGreen fractional cover image (right) output for a selected rangeland property in NSW, Australia). The spatial based land condition metrics captured new indicators of landscape clumpiness at the paddock scale and further work in underway to link these with established grazing land management frameworks.

While the full stack remote sensing process to build, validate and operationally deliver nationally consistent biophysical products based on Sentinel 2 is undeniably complex, the larger issues often come with working with users to understand the products, and working with developers across all sectors to maximise value from the information and delivery systems. The enhanced spectral, spatial and temporal resolution of Sentinel 2, when well calibrated, offers a significant advantage over Landsat for both grassland biophysical product generation, and also has driven uptake by end users who appreciate the spatial scale and accuracy of the products along with their timely delivery.

The grassland areas of Azerbaijan are under heavy anthropogenic pressure, including overgrazing, land use conversion to crop land, infrastructure projects, and climate change. To put a hold on further degradation of grasslands, to conserve and sustainably use winter, summer and common pastures as well as hay meadows, and to increase the productivity of those areas, consistent and up-to-date data on grassland extent and status is a prerequisite. In the GRAZE project of GIZ, DLR and E.C.O, we therefore aim at mapping the extent, productivity, and degradation status of all grassland areas in Azerbaijan based on current high spatial resolution remote sensing data. The derived nationwide maps should support decision-making processes as well as the development of pastures management policies, and improve the understanding of the condition of mountain and lowland grasslands.
Two field campaigns were conducted in August and October 2018 in 22 Azerbaijani districts, following two north-south and east-west transects through the country and covering a wide range of landscape types as well as a height gradient from -50 m to 2280 m above sea level. In total, 296 field points in grassland, arable land, and shrub land areas were visited, for which land cover, vegetation type, coverage, management and erosion intensity were recorded. These field observations are used together with samples of urban areas, bare soil, water, cropland, and forested areas (about 100 samples per class) collected manually on the screen from high spatial resolution Sentinel-2 scenes. The classification is conducted then using a Sentinel-2 time series including all 28 tiles that cover Azerbaijan and the entire year of 2018 (about 2900 scenes in total) in a random forest classifier. For the random forest models, several statistical metrics (e.g. 10, 25, 50, 75, 90 percentiles) were derived from Sentinel-2 imagery based on the band reflectances and selected vegetation indices, such as NDVI. Furthermore differences between percentiles and seasonal metrics were estimated. Hence, a total of 500 independent models where built and combined based on over 70 temporal and spectral input features. While 50 % of the data points are used as training data, the other half is used for validation purposes. A special focus of the land cover mapping is the proper discrimination of meadows, winter pastures, summer pastures, and common pastures occurring in a large variety of grassland types (steppes, semi-deserts, mountainous areas etc.), as well as of perennial crops such as alfalfa, which are cultivated and used in a similar manner as grasslands.
Relative grassland productivity is assessed in a next step for all identified grasslands areas based on seasonal intensity measures such as maximum and cumulative vegetation index (VI) metrics. The estimation of grassland condition is performed last through a multi-temporal comparison of the current grassland productivity measures with the same seasonal VI metrics derived from historical multi-temporal satellite imagery (Landsat). As reference periods, the years 1985, 1995, 2000, 2005, 2010, and 2015 have been selected in order to capture all grassland management and condition changes since the collapse of the soviet management regime. Through the assessment of relative changes in vegetation abundance and productivity, in combination with experiences from the field campaigns and from pilot projects on erosion control conducted in the Greater Caucasus, we aim at deriving nationwide indicator maps for the detection of degraded grassland areas in Azerbaijan.

Above ground biomass (AGB) is an important measure of terrestrial ecosystem productivity, and it is used in quantifying the role of vegetation in the carbon cycle, the potential for energy production, and the carbon stock estimation for climate change modelling. Dryland AGB, also recommended as the indicator of land productivity by UNCCD in desertification assessment and monitoring, need to be quantitative assessed and evaluated. With the development of remote sensing technology, vegetation parameters estimation at regional scale has become available. Earth observation satellites in the optical domain enable the retrieval and monitoring of vegetation. The in-orbit and forthcoming optical satellite scheduled by Europe and China, such as Sentinel 2A/B, BJ-2, GF-1/2/4/5/6 etc., will produce large data streams for vegetation parameters estimation. This vast data volume requires enhanced processing techniques that are accurate, robust and fast. However, owing to the characteristics of arid and semi-arid areas, vegetation coverage was sparse with relatively complex surface conditions, and the vegetation parameter inversion was significantly influenced by the soil or sand matrix, making it difficult to achieve high precision results.
Otindag Sandy Land, located in the centre of Xilin Gol grassland, Inner Mongolia Autonomous Region of China, is one of the four largest semi-arid desert areas in China. The sandy land stretches for approximately 400 km from east to west and 200 km from north to south, covering approximately 53,000 km2. The Otindag Sandy Land is in a typical semi-arid area of the grassland and the vegetation coverage is characterized by typical grassland plants and semi-arid desert vegetation. In this study, the Otindag Sandy Land and its surroundings (OSLAIS) was set as the study area and a useful method for sparse vegetation aboveground biomass inversion in dryland was promoted. Firstly, the Sentinel-2 remote sensing data coved the whole area in growing season (May to September) during 2015 to 2018 and synchronous field survey data was collected and processed. Then, the estimation model was constructed by linear regression model, power function model, exponential model and machine learning model by taking band information, texture information and different vegetation index into consideration. In addition, total 2/3 field survey sampled AGB data were used for modelling, and the remaining 1/3 measured AGB data was set as the testing to evaluate the AGB estimation models. Finally, the AGB distribution of the OSLAIS was mapped and analysed based on the optimal model. This research is expected to provide technical support and scientific reference data for vegetation assessment and monitoring in the study area, and even across the entire dryland of northern China.

Grassland can be managed in various ways – as meadow, pasture, or mowing pasture. The frequency of use or the stock densities defines the intensity of use of the respective grassland area. This intensity is an important factor determining the ecosystem functions related to biodiversity, water quality, greenhouse gas emissions and carbon storage as well as forage yield and economic value. Depending on site-dependent capabilities and objectives of use, different management strategies exist throughout Germany. However, there is no national data set on the character of grassland utilization.

Remote sensing provides information on land use and land use change for the whole country. As SAR signals are affected by the physical properties of vegetation, changes due to mowing or grazing lead to detectable changes in the signal. The utilization intensity of grassland can be determined by identifying mowing events and frequencies for meadows, or certain patterns arising from grazing for pastures. For this, a dense time-series of high-resolution satellite data is needed, which can be provided by Sentinel-1 SAR data.

In this study, parcel and management information from different test-sites in different regions of Germany with different management strategies are used to determine manner and intensity of grassland utilisation. For this purpose, the pixel-wise backscatter values of Sentinel-1 satellite images within the boundaries of each parcel are combined and different indices like median, variance, standard deviation or quartiles are calculated. The time-series of these indices per parcel are then examined regarding recurring patterns due to management events. Additional precipitation data is used to filter its impact on the SAR signal.

Auxiliary information from other geographic data sources or land use statistics allow assumptions to be made about the expected intensity of use, thereby supporting the development of a classification model. F. e. grassland on steep slopes, on lean or very wet soils is most likely extensively managed, due to management restrictions. Agricultural statistics with stocking rates on county level or agri-environmental and climate measures may also give clues on utilisation intensity.
In this ongoing study, we test the ability of SAR backscatter to discriminate management strategy classes for grassland in different regions in Germany. Moreover, we test how fuzzy auxiliary data from different sources can support and improve classification results. First results indicate that a model including Sentinel-1 data along with additional information allow a higher model fit than a model operating only on backscatter information.

The increased availability of systematically acquired high spatial and temporal resolution optical imagery offers great potential for improved characterization of dynamic land surface processes. Utilization of time series-based vegetation phenology can hereby help advancing applications to go beyond what is possible using conventional classification-based mapping. Especially agricultural applications can benefit greatly from the exploitation of the temporal domain. While the majority of agricultural time series applications targets commodity crops, the grassland domain has received comparatively less attention. This is unfortunate as grasslands are integral elements of land systems around the world, as they provide fodder and often primary support for livelihoods, can maintain biodiversity and provide many other ecosystem services such as carbon sequestration, water filtration and flood water retention. In Europe, permanent grassland accounts for more than one third of all agricultural land and a considerable share of the EU Common Agricultural Policy (CAP) payments is devoted to maintaining grasslands. These payments depend to a large extent on the intensity of grassland use which can vary considerably according to land-use and management practices, the local bio-geophysical environment or physiognomic varieties of different grass species. The frequency and timing of mowing events is an important proxy for grassland use intensity and of direct policy relevance in the frame of the CAP. Therefore, there is a strong requirement for methods that allow characterization of grassland use intensity at the parcel level and over large areas. Here we present a novel algorithm that allows detecting and quantifying the number and timing of mowing events in central European grasslands. The algorithm utilizes all imagery acquired by Sentinel-2 MSI and Landsat-8 OLI for one entire year as available from the NASA Harmonized Landsat-Sentinel dataset. Cloud-free observations from both sensors are first synthesized through compositing at a narrow temporal interval. Machine learning algorithms are then used to derive a grassland stratum. The intra-annual growing season profiles of vegetation index values are subsequently assessed and compared to an idealized profile of an unmanaged grassland phenology. Residuals between the idealized trajectory and a time series fit to the observed annual values are then evaluated to detect potential mowing events. We demonstrate and evaluate the performance of our algorithm and utilize its large area analysis capabilities by mapping the frequency and timing of grassland mowing events in 2016 on the national-scale across Germany. The ranges of detected mowing events compare overall well to a sample of interpreted time series points and to farm level reports on mowing dates. The mapped mowing patterns depict typical management regimes across Germany. Overall, our results exemplify the value of multi-sensor, high resolution intra annual time series applications for characterizing land use intensity across large areas.

Grasslands cover about one third of the agricultural area in Germany. In comparison to arable fields grasslands support a higher number of species and grassland soils are an important carbon sink. Due to economic reasons the total grassland area in Germany has been strongly reduced during the last decades. From 2003 to 2012 about 5% of the total grassland area in Germany was converted into other land use classes, mainly into arable fields. To stop this negative trend various actions were taken by the German government in order to reduce the turnover of grasslands and to promote a higher appreciation of grasslands. In order to increase the amount of high-quality fodder produced in Germany and to reduce international imports intensely managed grasslands are gaining increasing attention as forage production sites.
While satellite information is are already frequently used to optimize management and enhance yields on arable fields, there are hardly any satellite-based information services available to support grassland farmers in their management decisions. The goal of this work is to transfer already existing techniques for arable fields to intensely used grassland systems and to adapt those services to the needs of grassland farmers. The services will be easily accessible for the customers through integration into their farm management information system. Knowledge about the actual biomass on the site can serve as a first information for farmers to better determine the best moment for grassland cutting and to optimize harvest logistics. The challenge lies in deriving meaningful results for highly heterogeneous sites with varying species composition and multiple harvests during the year.
In this study we worked on parameterizing the radiative transfer model SLC, hitherto mainly used to analyze the growth of field crops, for the derivation of canopy parameters of grassland sites such as green leaf area and leaf chlorophyll. These variables are derived from optical satellite data acquired by the ESA-sensors Sentinel 2A and 2B. We compared the green leaf area derived by the radiative transfer model SLC to in-situ LAI-measurements conducted on a meadow in northern Germany. First tests were conducted how well cutting dates of meadows can be determined from Sentinel 1 and Sentinel 2 data on several grassland sites on northern Germany. Both, green leaf area derived out of optical satellite imagery as well as cutting dates can subsequently be assimilated into the crop growth model PROMET in order to derive the actual biomass on the sites and to determine potential harvest amounts. First results of this approach will be shown. The calculated harvest amounts were compared with yields measured by a forage harvester as well as to values calculated from empirical models to assess the quality of the entire processing chain.
This work was conducted as part of the SattGrün project, funded by the German Federal Office for Agriculture and Food (Grant agreement no: 28-1-83.013-16). We thank Lars Behrendt of the Agrargenossenschaft Paplitz eG, Germany, for providing us with harvest data of his grassland sites; Peter Kaim of Havellandhof Ribbeck, Germany, for the possibility to take LAI measurements on his fields and everyone who helped acquiring the in-situ data.