Land with Synergy of Proba-V and Sentinel-3

The New Zealand dairy industry is crucial to the economic welfare of the country. In the 2013-14 season, 4.92 million cows produced 20.6 billion litres of milk worth $15.5 billion to the farmers. While supplementary feeds are becoming more widely used, the majority of herds are fed on reygrass and clover pastures, and hay or silage cut from these pastures. Monitoring of pasture growth is an essential tool in the management of dairy herds whose average size has now reached over 400 cows.

In the field, pasture production is commonly measured using either a hand-held rising plate meter or a C-Dax Pasture Meter towed behind a 4WD motorbike; both are calibrated against pasture samples that have been cut and dried to deliver values in kgDM/ha (kilograms of dry matter per hectare). DairyNZ provides weekly measurements made this way for selected farms from all over the country.

NDVI and EVI, calculated from Proba-V and SPOT Vegetation imagery acquired over the past four years, were analysed to determine the relationship between these values and the pasture production data. Two very dry summers (2013 and 2015) and a 'normal' summer (2014) provided a wide range of conditions for pasture growth, or lack thereof.

The use of time-series and near-real-time satellite data provides an opportunity to supply information to the farmers that is not only more immediate than the present 1-2 week old measurements, but also provides complete coverage of the farmlands rather than just a few monitored farms in the district. Development of a system that could monitor individual farms is also progressing.

Knowledge about wildfire occurrence is fundamental to plan prevention and control. Therefore, monitoring burned areas is essential to assess and quantify fire impacts on ecosystems. However, it is often difficult to delimit burned areas in the field, due to limited accessibility. Proper application of remote sensing to map burned areas requires a thorough understanding of different kinds of vegetation, land use, climatic characteristics, as well as spatial and temporal distribution of fires. In this context, it is important to adapt image-processing techniques to the regional peculiarities of the problem. The aim of this study was to analyze the ability of discriminating between burned and unburned areas using PROBA-V channels BLUE, RED, Near Infrared (NIR), Sort Wave Infrared (SWIR), and the Normalized Difference Vegetation Index (NDVI), in the Minas Gerais state, Brazil. We analyzed the discriminant ability provided by these channels and NDVI to distinguish burned areas across all land cover types, as well as specifically for Cerrado savanna, Atlantic Forest, and for agricultural crops, and pastures. The hypothesis tested in this study is that discriminant ability for burned areas provided by the channels and NDVI differs by land cover type. We used images from the PROBA-V 300M SYNTHESES S1, Top Of Canopy (TOC), X13Y09, acquired 10/10/2014 (T1) and 10/17/2014 (T2) to test the discriminant ability for 1-week old fire scars. In general, the fire season starts in July, peaks in October, and end in November. The study area encompasses the Cerrado savanna, and the Atlantic Forest biomes. Sample pixels were collected in savanna, grassland, forest, cropland, and pasture areas. The BLUE, RED, NIR, SWIR channels and NDVI values were extract for T1 and T2 images. Based on these samples, statistical parameters (mean, variance, and standard deviation) and a separability index (M) were calculated. The M index was calculated based on sample means and standard deviation for T1 and T2 image pixels. The results showed that the BLUE channel provides the best discrimination between burned and unburned pixels in general (m=1.34), as well as for the different natural vegetation types, croplands, and pastures. The NDVI displayed better separability (m=0.83) than the RED, NIR, SWIR channels (respectively, 0.45; 0.56; 0.23). For all land cover types pooled together the best discriminator was the BLUE channel. Ability to separate between burned and unburned surfaces was highestin croplands, followed by savanna, grassland, pasture and forest (respectively, 3.53; 2.29; 2.25; 1.32; 1.18). After the channel BLUE, the NDVI provided more separability for savannah, agriculture, grassland, forest and pasture (respectively, 1.85; 1.42; 1.39; 1.13; 0.97). The SWIR channel showed worst discriminant ability for burns in for pasture, grassland and savannah (respectively, 0.23; 0.35, 0.42). For agriculture the RED channel was the worst (0.27) and for forest the worst was NIR channel (0.41). Thus, we conclude that there is variation between indices for discriminating burned areas depending on the type of vegetation under investigation. But, for the study area and time frame of the present analysis, the BLUE channel derived from PROBA-V showed the best ability to discriminate between burned and unburned surface, across all land cover types.

Compositing methods are commonly used in optical imagery to generate multi-days cloud-free images, which are input to terrestrial land cover and phenology studies. Since composite images are a mosaic of measurements obtained at different dates and with various viewing geometries, the algorithm used to select each pixel is of critical importance and artefacts due to directional effects should be minimized. The most popular approach used for compositing is the Maximum Value Composite (MVC) technique [1]. This method consists in the selection of those pixels, which presents the maximum value of Normalised Difference Vegetation Index (NDVI) for the considered compositing period. The rationale of the method is that high values of NDVI typically correspond to photosynthetically active cover types, i.e. green plants, while cloudy pixels correspond to low NDVI values. The MVC approach is currently used for the generation of Proba-V daily and decade synthesis products (S1 and S10). This choice was driven by the requirement of continuity with the SPOT-VGT 15 years’ dataset. It is however clear that the MVC method has some limitations, namely the tendency to select off-nadir observations in the forward scattering direction [2]. Owing to this, residuals directional effects can still be present in MVC composite images, introducing a systematic uncertainty for the vegetation data users. It is, therefore, essential to review and reassess the current Proba-V compositing approach and investigate potential evolutions in particular in the frame of Sentinel-3 Synergy VGT products definition.

Several other methods have been proposed that reduce the artefacts observed with the MVC technique. On the one hand, physically sound approaches, based on the inversion of a BRDF (Bi-Directional Reflectance Distribution Function) model using the available cloud-free pixels. On the other hand, statistically sound approaches based on averaging over the compositing period all pixels that satisfy stringent quality control criteria. The physically based approaches are more accurate, but they are computationally intensive and require a sufficient number of cloud-free observations in order to adjust the BRDF model’s parameters; owing to this, the compositing period is usually extended with the problem of losing information on surface parameters’ changes and vegetation dynamic. Statistically based approaches, such as the Mean Compositing (MC) method, are much easier to implement in an operational context and allow to effectively use the information from all cloud-free pixels acquired during the synthesis period [3].

Critical analysis of the current approach used for Proba-V synthesis products will be carried out in this paper and the systematic errors introduced by this method will be investigated in details. The improvements obtained applying the MC method will be quantified and discussed for different land cover types. Visual inspections of the composite image as well as statistical tests will be used to compare the compositing methods. The outcomes of this study will contribute to the improvement of Proba-V products algorithm definition and will be beneficial in the development of the Sentinel-3 Synergy VGT algorithm.

[1] Holben, B.N. (1986). Characteristics of maximum-value composite images from temporal AVHRR data. International Journal of Remote Sensing, 7, 1417-1434.

[2] Cihlar, J., Manak, D. and Voisin, N. (1994). AVHRR bi-directional reflectance effects and compositing, Remote Sensing of Environment, 48, 77–88.

[3] C. Vancutsem, J.‐F. Pekel, P. Bogaert & P. Defourny (2007). Mean Compositing, an alternative strategy for producing temporal syntheses. Concepts and performance assessment for SPOT VEGETATION time series, International Journal of Remote Sensing, 28:22, 5123-5141.

Mapping the global cropland extent is of critical importance for food security. Indeed, accurate and reliable information on cropland and the location of major crop types is required to make future policy, investment, and logistical decisions, as well as production monitoring. In both agriculture monitoring and climate modelling, cropland maps serve as mask to isolate agricultural land for (i) time-series analysis for crop condition monitoring and (ii) to investigate how cropland area may change with climate change. For all early warning systems such as GIEWS, EU-MARS and FEWS NET, precise and up-to-date cropland mask is a prerequisite to use satellite time series for crop growth monitoring. A systematic analysis of existing information provides insights on priority regions for cropland mapping and highlighted a major need for the Sahelian region. Indeed, the cropland classes in all global land cover maps hardly depicts the annual cropland mixing fallow, natural grasslands and cropland and typically have a poor accuracy. More specific efforts such as the Global map of rainfed cropland areas (GMRCA) and the Global irrigated area map (GIAM) are both at 10 km while the new 1 km global IIASA-IFPRI cropland percentage map is interesting baseline but for year 2005. National land cover maps completed by local or international programs (e.g. Africover, GLCN) provide detailed information but are usually not updated on regular basis. In the Sahelian region, the inter-annual variability of meteorological conditions, the diversity of cropping systems, and the agricultural landscape combining fallows and cultivated lands make cropland classification challenging. Taking advantage of the hectometric resolution of the PROBA-V central camera, this paper focuses on developing a new method for mapping the annual cropland extent in the Sahelian countries. This 100-m imagery is getting much closer to the agricultural field size allowing to better discriminate cultivated lands and fallows. The agro-ecological knowledge of the vegetation cycle and of the agriculture practices support the data selection from the full time series of PROBA. A data-driven approach then identifies relevant temporal features for cropland discrimination to be used in the framework of a 2014 cropland mapping at regional scale. The classification results are compared to existing products based on MERIS and SPOT-Vegetation time series and validated from reference data set. Finally, the expected performances from Sentinel-2 and Sentinel-3 time series are discussed in the perspective of an operational system for this region.   

The Copernicus land monitoring service provides information on land cover and on variables related, for instance, to the vegetation state or the water cycle. It supports applications in a variety of domains such as spatial planning, forest management, water management, agriculture and food security. The global component produces data across a wide range of biophysical variables at a global scale. One of the products is burned area and associated characterisation of the fire season. Burned area is detected at the native resolution of 1km (using SPOT VGT and PROBA-V) and 333m (PROBA-V) and is available at daily reporting resolutions since 1999. Fire season is described over a 1x1 degree grid cell. Accuracy metrics show that burned area algorithms developed for global scale mapping tend to perform better when the detection of a burn scar is combined with the detection of a flaming fire through the use of a sensor in the thermal domain. In this study, we show initial results on the integration of active fire data into the burned area detection algorithm. It is hoped that the method can be developed for inclusion of the Sentinel-3 SLSTR - Active Fire: Fire Detection product. Further characterisation of fire may be possible using Fire Radiative Power data in combination with the aforementioned products.