S2 and L8 Exploitation Synergy 2

Land surface phenology, including not only the timing of phenophase transitions but also the seasonal cycle of surface reflectance and vegetation indices, is important for a wide variety of applications including ecosystem and agro-ecosystem modeling, monitoring the response of terrestrial ecosystems to climate variability and extreme events, crop-type discrimination, and mapping land cover, land use, and land cover change. While methods to monitor and map phenology from coarse spatial resolution instruments such as MODIS and SPOT-VEGETATION are relatively mature, the spatial resolution of these instruments is inadequate for many applications, especially where land use and land cover vary at scales of 10’s of meters.  To address this need, algorithms to map phenology at moderate spatial resolution (~10-30 m ground resolution) using data from Landsat have recently been developed.  However, the 16-day repeat cycle of Landsat presents significant challenges for monitoring seasonal variation in land surface properties in regions where changes are rapid or where cloud cover reduces the frequency of clear-sky views.  The ESA/EU Sentinel-2 satellites, which will provide moderate spatial resolution data at 5-day revisit frequency near the equator and 2-3 day revisit frequency in the mid-latitudes, will alleviate this constraint in many parts of the world. Further, by combining data from Sentinel-2 and Landsat, it should be possible to monitor large areas of the Earth’s land surface at frequencies that were previously not possible.  In this paper, we describe efforts to exploit the combined observational capabilities of Landsat and Sentinel-2 to develop the algorithmic and computational basis for moderate spatial resolution monitoring of land surface phenology.  Specifically, we describe early results from algorithms that use a combination of Landsat and Sentinel-2 data to: (1) quantify the timing and magnitude of land surface phenology events (“phenometrics”), and (2) generate gap-filled time series of spectral vegetation indices that characterize the seasonal cycle of land surface phenology at fixed time steps.  Results from this research are designed to provide the foundation for operational production of multi-sensor land surface phenology data products at moderate spatial resolution.  Further, we plan to implement our algorithms in TIMESAT, thereby providing a flexible tool that can be exploited by the user community for location- and application-specific needs.

Fire products derived from coarse spatial resolution satellite data have become an important source of information for the multiple user communities involved in fire science and applications. There is, however, an unequivocal demand for systematically generated higher spatial resolution burned area products. Moderate spatial resolution contemporaneous satellite data from the Sentinel-2 and Landsat-8 sensors provide the opportunity for detailed mapping of burned areas. Combined, these polar-orbiting systems provide 10m to 60m multi-spectral global coverage up to every 3 days.  This NASA funded research presents early results to prototype a combined Landsat-8 Sentinel-2 burned area product.The Sentinel-2 data are reprojected into alignment with Landsat-8 data Web Enabled Landsat Data (WELD) tiles defined in the global sinusoidal projection. Two burned area mapping algorithms that are applied on a multi-temporal per-pixel basis are considered, one based on a non-parametric statistical classification approach and another that uses a more physically based inversion. Preliminary results for Southern Africa, showing time series mosaics of Sentinel-2 and Landsat-8 reflectance, and the burned area products are presented. Future research and the challenges for global implementation based on our global WELD (http://globalweld.cr.usgs.gov/) and global MODIS burned area (http://modis-fire.umd.edu/pages/BurnedArea.php) product generation experiences are discussed.

Related references:

Boschetti, L., Roy, D.P., Justice, C.O., Humber, M., 2015, MODIS-Landsat fusion for large area 30m burned area mapping, Remote Sensing of Environment, 161, 27-42.

Roy, D.P., Wulder, M.A., Loveland, T.R., Woodcock, C.E., Allen, R.G., Anderson, M.C., Helder, D., Irons, J.R., Johnson, D.M., Kennedy, R., Scambos, T.A., Schaaf, C. B., Schott, J.R., Sheng, Y., Vermote, E.F., Belward, A.S., Bindschadler, R., Cohen, W.B., Gao, F., Hipple, J.D., Hostert, P., Huntington, J., Justice, C.O., Kilic, A., Kovalskyy, V., Lee, Z. P., Lymburner, L., Masek, J.G., McCorkel, J., Shuai, Y., Trezza, R., Vogelmann, J., Wynne, R.H., Zhu, Z., 2014, Landsat-8: science and product vision for terrestrial global change research, Remote Sensing of Environment, 145, 154–172.

Roy, D.P., Ju, J., Kline, K., Scaramuzza, P.L., Kovalskyy, V., Hansen, M.C., Loveland, T.R., Vermote, E.F., Zhang, C., 2010, Web-enabled Landsat Data (WELD): Landsat ETM+ Composited Mosaics of the Conterminous United States, Remote Sensing of Environment, 114: 35-49.

Roy, D.P., Boschetti, L., Justice C.O., Ju, J., 2008, The Collection 5 MODIS burned area product -Global evaluation by comparison with the MODIS active fire product, Remote Sensing of Environment,112: 3690-3707.

Roy, D.P., Jin, Y., Lewis, P.E., Justice, C.O., 2005, Prototyping a global algorithm for systematic fire-affected area mapping using MODIS time series data, Remote Sensing of Environment,97: 137-162.

Over the years, time series break detection principles have been applied to different problems, ranging from deforestation and regrowth monitoring to detecting shifts in global vegetation trends. Since their publication in 2010, the BFAST algorithm and software package have evolved over the years, going from the change detection within time series (trend analysis) to the near real-time monitoring and follow-up of changes in space and time (e.g. forest disturbance and regrowth). Different satellite image time series, including those derived from MODIS, Landsat and PALSAR sensors, have been analysed separately or synergistically for change monitoring. However, an overview of what is possible today with break detection and monitoring approaches is lacking. Here, we provide an overview of what is possible today with break detection and monitoring approaches, together with practical recommendations and an outlook on further challenges (space-time monitoring) while entering a new era of satellite sensors (Sentinel 1 and 2, Landsat DCM). The objective is to clarify concepts, demonstrate current possibilities and highlight upcoming developments (e.g. open-source modular toolbox development for Sentinel time series and processing in the cloud via the Amazon cloud or Google Earth Engine). We will illustrate based on results from multiple studies across the tropics (DeVries et al. 2015a, b, Reiche et al. 2015, Dutrieux et al. 2015, Hamunyela et al., in review, Lu et al, in review) the need to move towards a sensor-independent change monitoring approach that is able to ingest multiple data streams for the detection and monitoring of changes in satellite image time series.

DeVries, B., Verbesselt, J., Kooistra, L., & Herold, M. (2015a). Robust monitoring of small-scale forest disturbances in a tropical montane forest using Landsat time series. Remote Sensing of Environment, 161, 107–121. doi:10.1016/j.rse.2015.02.012

Devries, B., Decuyper, M., Verbesselt, J., Herold, M., & Joseph, S. (2015b). Tracking disturbance-regrowth dynamics in tropical forests using structural change detection and Landsat time series. Remote Sensing of Environment, 169, 320–334. doi:10.1016/j.rse.2015.08.020

Dutrieux, L. P., Verbesselt, J., Kooistra, L., & Herold, M. (2015). Monitoring forest cover loss using multiple data streams, a case study of a tropical dry forest in Bolivia. ISPRS Journal of Photogrammetry and Remote Sensing, 107, 112–125. doi:10.1016/j.isprsjprs.2015.03.015

Reiche, J., Verbesselt, J., Hoekman, D., & Herold, M. (2015). Fusing Landsat and SAR time series to detect deforestation in the tropics. Remote Sensing of Environment, 156, 276–293. doi:10.1016/j.rse.2014.10.001

Hamunyela, E., Verbesselt, J., & Herold, M. Using spatial context to improve early detection of deforestation from Landsat time series. Remote Sensing of Environment, in review.

The need for quantitative and accurate information to characterize the state and evolution of forest types at the regional and national scales is widely recognized, particularly to assess deforestation and forest degradation processes, and related carbon emissions (Achard et al 2014).

Recently Hansen et al (2013) produced a global ‘forest’ cover change product (GFC) at 30 m resolution which includes tree cover changes between 2000 and 2012. However, a number of limitations can be observed: (i) GFC does not contain any thematic class; (ii) GFC does not relate strictly to a forest definition; (iii) GFC presents some confusions between forest and other land cover types such as forest-cropland mosaics; (iv) dense dry tropical forests are often mapped with a low tree cover percentage (e.g. in Northeast Brazil) ; (v) sparse dry tropical forests (e.g. in Tanzania) are not depicted (Hojas-Gascon et al 2015). Mapping forest types and monitoring forest conditions (e.g. intact versus degraded) at regional to national scale with fine spatial resolution (from 5 m to 30 m resolution) is still a challenge.

The objective of the study is to assess initial capacities of Sentinel-2 MSI sensor for monitoring degradation processes – burned areas, logging, fuel collection, grazing, all of which require high temporal and spatial resolution data for monitoring (Miettinen et al 2015) - in combination with Landsat-8 imagery and to assess the benefits of Sentinel-2 MSI data in comparison to Landsat-8 data, and to demonstrate the potential for synergy between the two satellites. The wide swath, high revisit frequency, choice of bands (SWIR) and spatial resolution (10m) of Sentinel-2 MSI sensor means that if rapid access to Sentinel-2 data is assured it will be a major tool in monitoring tropical forest cover and support tropical countries in enforcing forest law. In this study, we use Sentinel-2 data acquired from September 2015 over in test areas distributed in the following tropical countries:

South America : Colombia

Africa : Cameroon, Democratic Republic of Congo, Congo, Ivory Coast and Tanzania

Southeast Asia : Cambodia , Laos and Vietnam

Data quality issues (data format and acquisition) will be first addressed. Rationale for high level products (e.g. temporal mosaics at monthly or seasonal frequency, burned areas) will be looked for application to tropical forest degradation monitoring (Miettinen et al 2015; Shimabukuro et al 2015). New techniques for data processing and mosaicking, hereto only applied to lower spatial resolution sensors, such as MODIS, will need to be to be applied to derive the most from the higher temporal frequency of Sentinel-2 image acquisitions as compared to Landsat-8 imagery.

This preliminary assessment shows that Sentinel 2 data are suitable for the control of the location and extent of logging activities and that when combined with Landsat 8 data, may provide monthly data on burnt area extent.

References

Achard F et al 2014. Determination of tropical deforestation rates and related carbon losses from 1990 to 2010. Global Change Biology 20:2540–2554

Hansen M C et al 2013. High-resolution global maps of 21st century forest cover change. Science 342:850–853

Hojas-Gascon L et al 2015 Monitoring deforestation and forest degradation in the context of REDD+ Lessons from Tanzania. CIFOR Info Brief 124 DOI:10.17528/cifor/005642

Miettinen J et al 2015. First assessment on the potential of Sentinel-2 data for land area monitoring in Southeast Asian conditions. Asian Journal of Geoinformatics 15:23-30.

Shimabukuro YE et al in press. Estimating burned area in Mato Grosso, Brazil, using an object-based classification method on a systematic sample of medium resolution satellite images. IEEE JSTARS accepted

Earth Observation (EO) is a critical tool to develop better agricultural monitoring capabilities. With its 10-20 meter resolution, its 5-day revisit frequency and its global coverage, the new Sentinel-2 mission offers unprecedented performances for agriculture applications. Further, its compatibility to the Landsat missions allows building on an historic time series of observations.

In this context, the European Space Agency (ESA) launched in 2014 the Sentinel-2 for Agriculture (Sen2-Agri) project which aims at developing open source algorithms and software to process Sentinel-2 data in an operational manner for major worldwide representative agriculture systems.

Three agricultural products are targeted by the project [1]: (i) a cropland mask which is updated on a monthly basis along the season, (ii) a crop type map which identifies the main regional crop types or crop groups of the area (including the distinction between rain-fed and irrigated crops) and which is produced as soon as possible after the end of the season, with an early delivery after the first half of the season, (iii) vegetation status indicators (NDVI and LAI) describing the vegetative development of crops on a 7 to 10 day basis.

The processing methods required to derive these products were selected through a benchmarking study, which tested concurrent algorithms described by a state of the art review. A unique “Sentinel-2 like” time series including SPOT4 (Take5) and Landsat-8 data was acquired and processed over 12 sites spread over the world and representative from a large variety of agricultural practices and climate conditions. Special emphasis was placed on sites with crops of high global importance for food security surveillance (wheat, maize and soya), thus contributing to the GEOGLAM initiative [2]. Linear interpolation of cloudy pixels produced a gap-filled temporal image series. For each site, in situ crop measurements were shared by JECAM network members and other teams in the field.

For the dynamic cropland mask, both supervised and unsupervised approaches were considered. The supervised approach was assessed using a Random Forest classifier based on ground truth as prior information and temporal and spectral features derived from NDVI, NDWI and Brightness indices [3]. The unsupervised approach relies on prior information corresponding to a globally available baseline to train a maximum likelihood or a K-means classifier with spectral-temporal features. These features consists in selected bands corresponding to temporal features based on the NDVI time series [4]. The impact of spatial unit, i.e. the pixel or the field, was tested in order to improve the results accuracy thanks to the spatial information. The field spatial unit performances was assessed as a pre-processing step, providing averaged information to the classifier, as well as a post-processing step with a majority voting decision rule. The supervised and unsupervised approaches for cropland mapping enable to handle various situations (cropping systems but also availability or not of in situ data) with an overall accuracy over 85% and reaching more than 95% on many sites.

For the crop type product, a large set of combinations of existing supervised classification algorithms were analyzed including approaches related to feature selection, selection and parameterization of the classifier. The finally selected classification algorithm is a supervised Random Forest classifier. Temporal and spectral features used as input to the classifier were selected based on spectral indices time series (NDVI, NDWI and Brightness) [5].

The vegetation status product provides an LAI estimation based on the inversion a vegetation radiative transfer model [6]. This approach was selected after a comparison with linear and nonlinear regressions applied on ground truth data. Two improvements were applied to the single date estimation of the method presented in [6] in order to exploit the time series provided by Sentinel-2. The first approach is a real-time reprocessing of the LAI retrieval by using a weighted combination of previous values. The second approach is applied at the end of the crop season and consists in fitting a phenological model over the LAI time series [7]. These approaches were validated over 4 test sites (Belgium, France, Morocco and Ukraine) and showed consisted results across sites.

Such composite EO dataset allowed assessing both the added value of high spatial and temporal resolutions that will characterize Sentinel-2 time series and the comprehensive synergy with Landsat-8 data for agriculture monitoring applications. As a result, the algorithms selected from the benchmarking have been implemented in the Sen2Agri tool box and be directly usable to exploit Sentinel-2/Landsat-8 time series.

[1] Bontemps, S.; Arias, M.; Cara, C.; Dedieu, G.; Guzzonato, E.; Hagolle, O.; Inglada, J.; Matton, N.; Morin, D.; Popescu, R.; et al. Sentinel-2 for Agriculture: Towards the exploitation of Sentinel-2 for local to global operational agriculture monitoring. Remote Sens. 2015, under review.

[2] Becker-Reshef, I.; Justice, C.; Sullivan, M.; Vermote, E.; Tucker, C.; Anyamba, A.; Small, J.; Pak, E.; Masuoka, E.; Schmaltz, J.; et al. Monitoring global croplands with coarse resolution Earth observations: The Global Agriculture Monitoring (GLAM) project. Remote Sens. 2010, 2, 1589–1609.

[3] Valero, S.; David, M.; Jordi, I.; Guadalupe, S.; Hagolle, O.; Arias, M.; Dedieu, G.; Bontemps, S.; Defourny, P.; Koetz, B. Production of a dynamic cropland mask by processing remote sensing image series at high temporal and spatial resolutions. Remote Sens. 2015, under review.

[4] Matton, N.; Canto, G.S.; Waldner, F.; Valero, S.; Morin, D.; Inglada, J.; Arias, M.; Bontemps, S.; Koetz, B.; Defourny, P. An Automated Method for Annual Cropland Mapping along the Season for Various Globally-Distributed Agrosystems Using High Spatial and Temporal Resolution Time Series. Remote Sens. 2015, 7, 13208-13232.

[5] Jordi Inglada, Marcela Arias, Benjamin Tardy, Olivier Hagolle, Silvia Valero, David Morin, Gérard Dedieu, Guadalupe Sepulcre, Sophie Bontemps, Pierre Defourny and Benjamin Koetz, "Assessment of an Operational System for Crop Type Map Production Using High Temporal and Spatial Resolution Satellite Optical Imagery", Remote Sensing, 7:12356-12379 (2015)

[6] Marie Weiss, Frédéric Baret, Marc Leroy, Olivier Hautecœur, C´edric Bacour, Laurent Pr´evot, and Nadine Bruguier. Validation of neural net techniques to estimate canopy biophysical variables from remote sensing data. Agronomie, 22(6):547–553, 2002.

[7] Koetz, B., Baret, F., Poilvé, H., & Hill, J. (2005). Use of coupled canopy structure dynamic and radiative transfer models to estimate biophysical canopy characteristics. Remote Sensing of Environment, 95(1), 115-124.

The use of time series data at spatial resolutions that capture human activities is revolutionizing the use of optical remote sensing to monitor, among others, land use and land cover change, ecosystem health and condition and forest species composition.  The opening of the Landsat archive has led to exciting new methods and new findings with respect to the kinds of change that can be monitored and how rapidly it can be detected. I propose to show example results from the following: land cover change monitoring in the US, Colombia, Brazil and Viet Nam; forest species compositon mapping from time series in the US; monitoring of forest health and condition in the US.  The work to date is based on Landsat, which serves as a pathfinder for the more frequent observations being collected by Sentinel-2.  While access to Sentinel-2 data has been limited to date, the goal is to present results that integrate Landsat and Sentinel-2 data at the Symposium.  Challenges to be addressed include atmospheric correction and cloud/shadow detection, geometric and spectral differences between sensors and the integration of both sensor in time series analysis.