Crop Mapping along the Season at 10m: Methodological Developments Towards an Operational Exploitation of S-1, -2, and -3
Waldner, François; d'Andrimont, Raphaël; Defourny, Pierre
Université catholique de Louvain - Earth and Life Institute, BELGIUM
In the context of increasing pressure on land areas for food and fodder production increases and the competition with energy and fiber production, the G20 has recognized the importance of timely, accurate and transparent information to address food price volatility and quality of data, such as production, on agricultural market. Crop specific masks produced early in the season are required to improve crop acreage and productivity estimates. Therefore, the FP-7 project IMAGINES proposes to produce 10-m crop specific maps updated along the growing season integrating images from Sentinel 1, 2 and 3. The overarching objective is to develop a methodology that combines the advantages of each satellite: the temporal coverage of Sentinel-3, the high resolution of Sentinel-2 and the weather independent acquisitions of Sentinel-1. The method that integrates time-series and object-based analysis as well as machine learning techniques, is currently tested over two large contrasted sites that participate in the Joint Experiment of Crop Assessment and Monitoring (JECAM): one in Russia and the other in South Africa. The South African site, covering the Free State, is characterized by a sub-humid to semi-arid climate. Typical field size ranges from 0.5 ha to 40 ha planted with a continuous cropping season that include both summer crops (maize, sunflower, soy beans, sorghum, groundnuts) as well as winter crops (wheat, oat). On the other hand, the Russian site is located in the Tula region. In this continental climate region, the average field size is 100-ha. Fields are mainly planted with winter wheat, barley maize and potato. Contrarily to South African irrigation practices, crops are only rainfed.
For the growing season of 2013, Sentinel data are simulated by currently available satellite images. Daily MODIS and PROBA-V data are acquired as well as the available Landsat-8 images. At two-week intervals, fifteen RADARSAT-2 and RapidEye or SPOT-4 images are acquired from February to August and April to November for South Africa and Russia, respectively. Combining those proxy data, three maps will be produced along the season taking advantage of the accumulation of the information along the growing season to progressively further discriminate the different crop species. Such multi]scale and multi]spectral information should improve the temporal density of observations essential to capture the crops phenology, which is the main criterion of the discrimination process. The joint use of high spatial resolution of both optical and radar images is expected to resolved most field. The methodological development targets three successive outputs:
(i) At the beginning of cropping season, a pre-season cropland extent map is produced using a dedicated land cover algorithm based on the previous year time-series at moderate resolution. Agricultural calendar, crop rotation systems and the best local land cover map available (i.e. LC Corine, AFRICOVER, GLOBCOVER, CCI LAND COVER) helps this early diagnostic.
(ii) At the end of the winter, crop group recognition (winter and summer crops) is allowed thanks to a multi-scale and multi-temporal segmentation taking both advantage of time consistency of medium resolution data and spatial consistency of high resolution data.
(iii) At the mid-growing season, corresponding to the peak of the vegetation cycle, a crop specific classification is finally achieved.
All over the growing season, this challenging combination of 3 sensors (Sentinel 1, 2 and 3) will thus allow a continuous diagnostic at regional scale.
In this way, crop maps will be updated along the season with an increased precision (increased number of classes and decreased omission and commission errors) that, in chain, will support the retrieval of bio-physical variables for crop growth models and early acreage estimation.