Tropical Forest and REDD+ 1

Current methods for monitoring deforestation from satellite data at sub-annual scales only use pixel time series to statistically differentiate deforestation from normal forest dynamics. However, a pixel time series may have too few observations for such approach to work. This problem of insufficient temporal observations at pixel level could be addressed by including spatial information in the analysis. In this work, we investigated how spatial and temporal information can be combined to detect deforestation from satellite image time series. We developed a data-driven space-time change detection method that exploits both spatial and temporal information, and detects deforestation from satellite image time series as an extreme event. We applied the method to Landsat normalised difference vegetation index (NDVI) time series to detect deforestation at a dry tropical forest site in Bolivia in a “near real-time” monitoring scenario. We achieved a median temporal detection delay of one observation and overall spatial accuracy of 96.3% when using a data cube with spatial size of 7.29ha. Our method only requires a pixel time series to have a minimum of three observations, of which one observation must be in the reference data cube for such pixel to be assessed for deforestation. The reference data cube contains historical spatio-temporal observations- observations acquired prior to the start of deforestation monitoring. The requirement of only three observations makes our method suitable for tracking deforestation events from high spatial resolution data e.g. RapidEye whose time series are highly irregular and temporally short. In a next step, we will apply the our method to RapidEye time series to detect forest disturbances in order to demonstrate the usefulness of combining spatial and temporal information when monitoring forest disturbances from short image time series. The method we proposed here can also be used to immediately exploit satellite data from newly launched satellite sensors (e.g. Landsat 8 and sentinel-2) to detect deforestation events even when the image time series from such sensors are still short.

While REDD+ implementation is progressing, the question how development in technologies and evolving needs from multiple stakeholders for REDD+ monitoring and reporting are influencing global, national and local engagement mechanisms. These includes the broadening of the scope how REDD+ interacts with sustainable development agenda (i.e. sustainable development goals), and the need to include many more stakeholder than just governments in the process (i.e. local communities, private sector/sustainable supply chains, citizens, research); including the increase in transparency and legitimacy of observations, datasets and estimates.

The EU-funded project on “Strengthening Independent Monitoring of GHG Emissions from Land Activities for Publishing, Comparing and Reconciling Estimates” aims at developing analysis, case studies and other research that ultimately lead to proof of concept towards an operational, independent, publicly available, comprehensive and global spatial information system on land cover, land emissions, land use, their dynamics and the associated carbon stocks and flows. Several global monitoring systems are now being started by several initiatives, and the project takes the opportunity to analyze these initiatives comparatively and to scrutinize them in light of user needs, to derive recommendations for more efficient and effective monitoring systems that cover land use and land use change beyond forests, address different user needs, particularly those of users with limited capacities of data handling and interpretation, identify research gaps, and bridge the increasingly widening gap between advanced and less advanced countries. Such global information systems on land cover, land emissions, and land use serve multiple objectives.

First we analysed 23 key datasets and 31 web portals. The datasets and portals were evaluated based on the criteria and indicators using an assessment template. In addition we conducted a comprehensive stakeholder analysis with more than 600 respondents from a large array of users. Overall, the outcomes of the stakeholder survey indicate that in general all stakeholders emphasize: data with increased transparency and documentation of data sources, methods, definitions and assumptions, data on biomass with high opportunities and eventually including data on soil organic matter, higher Tier (2 or 3) emission factors and emission estimates, harmonized GHG AFOLU emissions data, especially for the tropics and region-specific data and data and uncertainty targeted for users. These findings have resulted in a series of case studies, where the results are presented, in particular considering the role of Earth Observation and the role of the new ESA and EU initiatives related to Copernicus and BIOMASS:

Global contribution of AFOLU GHG emissions (2000-2005): patterns, uncertainties and drivers

Reconciling data conflicts: forest change, deforestation and degradation datasets in Indonesia and Borneo

Global forest biomass uncertainties and their integration with national and regional estimation and reporting

Contributing to improved emission factors for forest and agriculture - Using biophysical soil models: challenges and opportunities for Tier 3 approaches

The results highlight the need to fill key data gaps through better observation and data and how new scientific datasets may provide an opportunity are important to continue investments for continuation of programmes and improvements of datasets.

The need for quantitative and accurate information to characterize the state and evolution of forest types at the regional and continental scales is widely recognized, particularly to analyze the forest diversity and dynamics, assess the degradation and deforestation and better manage the natural resources.

Various global land cover maps derived from satellite imagery depict several forest classes such as the Boston University’s MODIS-derived land cover product [1], and the Climate Change Initiative Land Cover (CCI-LC) maps [2].  Such maps contain up-to-date and detailed thematic information but at coarse resolution (from 300m to 500m).  Moreover large discrepancies have been observed between the various global products [3]. Other land cover maps are at a much better resolution (30m) such as the land cover map of South America of Giri and Long [4] or the global land cover map of Chen et al. [5], however these maps do not distinguish the evergreen from the deciduous forest and they not consider the forest dynamic over a long period. Other maps are more dedicated to the forest monitoring such as the annual maps of Xiao et al. [6], but still derived from coarse resolution data. Recently Hansen et al. [7] produced a global forest change product (GFC) at 30 m resolution which includes (i) the tree cover percentage for year 2000 and (ii) tree cover changes between 2000 and 2014. This product presents a great improvement in term of spatial resolution and offer a finer delineation of tree cover. However, two main limitations can be observed:

(i)                 GFC does not contain thematic classes. Moreover the tree cover percentage does not relate strictly to a forest definition [8] and there is no defined continental or regional tree cover threshold that would allow discriminating forest areas or forest types [9];

(ii)                GFC presents some confusions between forest and other land cover types. When considering a tree cover percentage above 50%, confusions appear mainly with shifting cultivation, irrigated cropland and tree plantations. When decreasing the tree cover percentage below 50%, more confusions appear between flooded vegetation, savannahs and agriculture.

Mapping the remaining dense humid forest over the tropical belt at fine spatial resolution is still a challenge.  

In this paper, we present a new methodology that exploits the 32-years database of Landsat imagery (3 sensors from 1984 to 2015) which allows to produce a pan-tropical map of dense humid forests at 30m resolution with discrimination of evergreen forests, secondary forests and vegetation regrowths and identification of recent deforestation and degradation patterns.  In addition, we characterize these later classes by providing the timing and occurrence of the deforestation or degradation events (start date, end date, number of repetition). Such timing indicators are particular interest for assessing the impact of a disturbance on the tropical forest.  

Our pixel-based fully automatic methodology includes four steps: (i) pre-processing of the Landsat time series with cloud masking  adapted to tropical areas, (ii) single image classification (using a threshold approach) into forest and non-forest classes while keeping the date of the non-forest event, (iii) creating forest/non-forest maps for three epochs, i.e. 1984-2004, 2005-2012 and 2013-2015, based on the occurrence of non-forest events, and (iv) creating a final map of forest types based on the succession of classes from 1984 to 2015.

Due to the large volume of data to process, the Google Earth Engine platform is used for the testing phase, the selection of the parameters, the processing of the Landsat imagery and the validation.

A first version of the forest map has been produced at 30m resolution including five classes: evergreen forests, secondary forests, vegetation regrowths, recent (last 3 years) degraded forests and non-forest areas. The map provides a much finer spatial resolution compared to existing global land cover products what is particularly interesting for delineating the linear features such as the gallery forest (Figure 1a) and the small degradation events like the skid trails and logging decks (Figure 1b, 1c and 1d). The use of a 32-year time series allows (i) identifying the majority of the deforestation and degradation events that occurred during this period (Figure 1b) and to classify according to the date of the disturbance (Figure 1c, in orange when it started in 2014 and in red when it started in 2015), and (ii) considerably reducing the contamination errors with the agriculture (Figure 1d) (shifting cultivation, tree plantation and irrigated crops…). Finally the methodology allows assessing the impact of a deforestation or a degradation event by providing the start and end dates (and consequently the period), and the occurrence of the event (Figure 1e). 

The accuracy of the forest map will be assessed using an independent pan-tropical sample of reference data: this dataset will be created through visual interpretation by expert from Landsat and very high resolution images available during the three epochs used in the classification.  The sample will be designed using a two-stage selection approach: first a systematic scheme based on the latitude / longitude geographical grid is used [9], then applying a randomly stratified scheme based on the GFC tree cover percentage with an equal number of sample points in the three following tree cover ranges: (1) 75-100%, (2) 50-75%, and (3) below 50% because the strata (2) and (3) are considered to include most of the disturbed areas.

Further work will apply the method on Sentinel-2 data as soon as a time series dataset will be available (e.g. on Congo Basin). 

The estimation of deforestation area in tropical countries often relies on satellite remote sensing in the absence of National Forest Inventories (NFI). The assessment of areas of forest cover and forest cover change is essential to estimate activity data, defined as areas of various categories of land use change by the IPCC guidelines.

Wall-to-wall mapping is often required to provide a comprehensive assessment of forest resources and as input to land use plans for management purposes, but wall-to-wall approaches requires specialized equipment and staff.that are often not available. The recent release of the University of Maryland (UMD) Global Forest Change (GFC) map products or the global (2007-2010) Forest/Non Forest map based on ALOS PALSAR sensor data provide an alternative for tropical countries wishing to develop their own wall-to-wall forest monitoring map products but without the resources to do so. However, these map were produced at global scale and cannot be used directly for national/regional mapping because they need to comply with a selected forest definition.

A purposely designed probability sample was developed to calibrate global maps based on the visual estimation of percent tree cover from Very High Spatial Resolution satellite imagery for the southwest forest massif in Central African Republic covering an area of about 70,000km². Optimal thresholds were determined based on the comparison of tree cover percentage calibration data with map data to match the selected forest definition.

Local forest cover maps were produced based on the selected thresholds and assessed with reference data for forest cover obtained from an additional separate probability sample. A model assisted regression (MAR) estimator was applied using the combination of reference and map data to produce estimates of forest cover and their associated uncertainties as required in the IPCC 2006 guidelines and for reporting to the United Nations Framework Convention on Climate Change (UNFCCC).

French Guiana’s tropical evergreen forests constitute a significant portion of France’s overall forest cover. Nevertheless, in various global vegetation mapping initiatives – including the European Space Agency’s GlobCover project – French Guiana’s forests have been depicted as an extensive “green carpet.” Those global efforts have not differentiated the range of forest types present across the French overseas territory, even as the mapping of forest types constitutes a key input to efforts ranging from biodiversity conservation to forest management to REDD+. Taking that into context, an effort was made to produce a detailed map of French Guiana’s forest types, taking advantage of both the spatial and spectral resolution of the Landsat series of satellites. As the territory spans 3 separate Landsat orbital tracks, a dry season, wall-to-wall mosaic of 8 images was assembled using data acquired in September (albeit images from different years, due to different patterns of cloud cover). Those images were acquired from both Landsat-5’s Thematic Mapper (TM) and Landsat-7’s Enhanced Thematic Mapper+ (ETM+). Anisotropic surface reflectance – a consequence of the bi-directional reflectance distribution function, BRDF – became an issue, leading to image overlap areas where reflectance values not only did not match but differed significantly. A solution was found by using MODIS data already corrected for BRDF (i.e. the MCD43A4 product) as a reference reflectance dataset to calibrate the Landsat imagery to, especially given the similarities between the spectral bands of the TM and the ETM+ sensors and MODIS. Where a few earlier studies used regressions involving both the Landsat-MODIS bias and the Landsat scan angle to generate band-by-band correction factors for the Landsat data, in this study, global polynomial interpolation of the Landsat-MODIS bias data was used to generate the correction factors. The consequence was that while the across-scene anisotropy was corrected, relative differences in reflectance were maintained. The resulting BRDF-corrected wall-to-wall Landsat mosaic was in turn used as an input to a supervised classification of 12 forest types of French Guiana, and 4 additional land cover classes. In terms of the importance of the BRDF correction of the Landsat input data, when the same training data was used on Landsat imagery not corrected for surface anisotropy, only ~33% of the output classification coincided with the classification derived from the BRDF-corrected data. Much of the areas classified differently between the two maps were in areas near the scene overlaps, where scan angles were high, and BRDF effects can be expected to be greater, a problem resolved through BRDF-adjustment of the input Landsat imagery. While a full validation of the forest type map is pending, a smaller scale assessment indicates that the map has, with reasonable accuracy, identified the distributions of a few classes of forest (e.g. liana-covered forest and bamboo-dominated forest) which are otherwise too small to be mapped with coarser-scale sensors. This also sets the stage for eventual extension of the study using data from Sentinel-2 when it becomes available.