Tropical Moist Forests product - Data Access

List of updates integrated in the TMF 2024 products

• Integration of all 2024 Landsat 8 and 9 imagery over the pan-tropical belt
• Delineation of newly established agricultural tree plantations, primarily rubber and oil palm, based on visual interpretation of 2024-2025 very high-resolution imagery and the spatial distribution of recent deforestation detected by TMF.
• Minor corrections applied to the integration of visually delineated agricultural tree plantations into TMF Transition Map -sub type.

List of updates that are integrated in the TMF 2023 products

• Full recomputation of TMF products based on the full dataset of reprocessed Landsat Collection 2 imagery. The Landsat archive from 1982 to 2022 has been improved as Collection 2 in terms of geometric and radiometric accuracies of the imagery (see: https://www.usgs.gov/landsat-missions/landsat-collection-2). The full Landsat Collection 2 has been used to recreate the new 'TMF 2023' dataset. The use of Landsat Collection 2 results not only in better quality input data, but also in an increase in the overall number of valid observations (i.e. observations free of cloud/cloud shadow/haze coverage or sensor issue) of 11% , 5% and 20 % in Latin-America, Africa and Asia respectively for the period 1990-2022 compared to Landsat Collection 1. The average number of valid observations per pixel for Landsat Collections 1 and 2 and is presented at continental level in the figure below:

  <img class="center" src="/static/tmf/valid_landsat_observations.png" alt="Average number of valid observations per pixel for Landsat Collections 1 and 2"/>

• Addition of year 2023 from the processing of Landsat 7, 8 and 9 imagery of year 2023 available in collection 2

• Addition of newly delineated tree plantations (that existed before 2023 and plantations established in 2023) mainly in Central and West Africa, Indonesia and Malaysia

• Application of improved spatial filters to remove noise or false positive in the detection of short-duration disturbances

• Improvements in the detection of the first year of forest degradation and deforestation events given the enhanced input Landsat C2 data

  • Degraded Forests are defined as pixels with a maximum occurence of 3 short-duration disturbance events observed between 1990 and 2022. These short-term events have a maximum duration of 900 days (when disruptions, i.e. absence of tree foliage cover within a Landsat pixel, are observed) and need to be separated by at least two years with no disturbance observation. In previous TMF versions, the duration of the first disturbance event was recorded in number of days whereas the following events were only characterized by their duration in number of years. In this TMF-v2023 version, the duration in days of the first three disturbance events are recorded. Beyond an occurrence of three short-duration disturbances events, the pixel is classified as deforestation from the starting date of the first observed disturbance event. Note that the distinction between forest degradation and deforestation in last three years of the analysis (i.e. including year 2023) is always based on the ratio between the number of valid observations and the number of observed disruptions.
  • Deforestation refers to conversion of an undisturbed or degraded forest to another land cover type which is characterised in the TMF approach as long-duration disturbance event (>900 days). The year of deforestation is attributed to the starting year of a disturbance event of more than 900 days duration or to the starting year of the first disturbance event when more than three consecutive short term disturbance events are detected. In the case of forest conversion to agricultural plantations (e.g. oil palm, coconut or rubber plantations), the year of deforestation corresponds to the year of the first disturbance event if longer than 900 days or to the year of the second disturbance event if the first disturbance event's duration was less than 900 days.

• Together, the improvement of the classification rules (for first year detection) and the reprocessing of the full Landsat archive to collection 2 led to updates in the historical dynamic of forest degradation and deforestation. It is important to note that change events that were detected in previous TMF version are not removed but can be potentially reassigned to a previous year or converted to another class of change (e.g. indirect into direct deforestation). The following figures show a comparison between the area estimates of deforestation (direct and after degradation) and forest degradation (not followed by deforestation) from TMF version 2023 (TMFv2023) with version 2022 (TMFv2022) for five years periods from 1990 to 2019 at continental level.

  Deforestation in TMF v2023 is higher across the three continents from 1990 to 2004 compared to the previous version v2022. This is due to several combined factors: (1) increase in the overall number of valid observations in particular for historical periods, (2) improvement in the distinction between degradation and deforestation through the more accurate recording of each disturbance event's duration, (3) earlier attribution of the deforestation year in the case of large number of disturbance events, and (4) earlier attribution of the deforestation year in the case of forest conversions to agricultural plantations. We quantify an increase of 15% in total global deforestation in TMF v2023 compared to TMF v2022 for the period 1990-2022 (2%, 12% and 39% increase in LAC, Africa and Asia, respectively).

  From 2010, we observe a decrease of deforestation area estimates in TMF v2023 compared to TMF v2022. This can be explained by a decrease in deforestation after degradation after 2010 which either occurred earlier in the time series or was reclassified into direct deforestation. Overall, deforestation of previously degraded forest has decreased by 22% in v2023 compared to the previous version (62.3 Mha of forest being degraded and then deforested in TMF v2023 against 79.7 Mha in TMF v2022). In other words, 27% of forest degradation is a precursor of deforestation in TMF v2023 against 41% in TMF v2022 (and 45% in TMF v2019).

  Overall, there is a decrease of 10% in total global degradation in TMF v2023 compared to TMF v2022 for the period 1990-2022 (8%, 12% and 11% decrease in LAC, Africa and Asia, respectively) but the trends using 5-years reporting periods remain similar between the two versions. Around 25% of degraded forest with a single degradation period in TMF v2022 are now classified as degradation with 2 or 3 degradation periods in TMF v2023. Moreover, 22% of long duration degradation in TMF v2022 is now classified as deforestation followed by forest regrowth as a result of improved classification rules and increased number of valid observations. Finally, 25% of degradation with 2 or 3 degradation periods in TMF v2022 is now classified as forest regrowth after deforestation and 11% as single degradation period.

  

New hybrid transition map at 10m (beta version)

• A hybrid transition map is now available at a resolution of 10m for the period from 1990 to 2022. This beta version combines the recently updated TMF map for the year 2022, derived from Landsat images at 30m, with the detections of forest disturbances by Sentinel 2 for the year 2022 at 10m. The integration of Sentinel-2 data allows a better identification of degraded forests, and a refinement of the disturbance edges and linear disturbances, such as logging roads and small rivers within the forest. The hybrid transition map is available for visualisation and download via ftp or GEE, along with a map detailing the data source of each disturbance detection. This feature allows users to identify whether disturbances were detected using Landsat, Sentinel 2, or both sensors in query mode by clicking on the map. Historical Sentinel 2 data for the previous years (2016-2021) will be used to produce an enhanced beta version of this hybrid transition map when the back processing of the Sentinel 2 archive, notably the improvement of the geometric accuracy, will be completed by ESA. Part of the small-scale forest disturbances that have been detected in the new beta map through the use of Sentinel 2 data in 2022 may have occurred in previous years but were not detected by 30m Landsat imagery. The use of the full Sentinel 2 archive for the final planned version of the hybrid map is expected to enhance the overall quality of the hybrid map and mitigate the issue of timing of the disturbance event. The final product combining the full archive of Landsat and Sentinel 2 data up to year 2022 is intended to be delivered with an accuracy assessment.

List of updates integrated in the TMF 2022 products

• Integration for year 2021 and year 2022 of Landsat Collection 2 imagery which is designed, amongst other issues, to improve the geometric and radiometric accuracy of the imagery (https://www.usgs.gov/landsat-missions/landsat-collection-2). The finalisation of the whole Landsat archive's (1990-2020) reprocessing and integration to the TMF dataset is foreseen for the end of 2023. This will not only result in better quality satellite image data, but increase the overall number of available imagery for analysis. For year 2021 and at the pantropical level, we quantified an increase of 13% in the detection of forest disturbances (new degradation or direct deforestation) from Collection-2 (C2) compared to Collection-1 data (C1). 66% of new forest disturbances from C2 are detected by C1, while 18% are within 60m (~ 2 Landsat sized pixels) from forest disturbances detected by C1 (most probably linked to the increased geometric accuracy in C2 data) and 16% of C2- forest disturbances are located beyond this buffer areas (most probably linked to the increased number of Landsat scenes and valid observations). You can find below a country-level comparison of 2021 detection of forest disturbances between C1 and C2 data: NewForestDisturbances_2021_C1_C2_byCountry.xlsx
• Integration for year 2022 of Landsat 9 (L9) imagery (https://www.usgs.gov/landsat-missions/landsat-9) along with Landsat 7 and 8 (resp. L7, L8). For the pantropical region, this resulted in a 35% increase of forest disturbances detection (new degradation or direct deforestation) in 2022. 43% of forest disturbances are detected by both L7/L8 and L9, while 31% are only detected by L7/L8 and 26% are only detected by L9. You can find below a country-level comparison of 2022 detection of forest disturbances between L7/L8 and L9 data: NewForestDisturbances_2022_L7_L8_L9_byCountry.xlsx
• Integration of the global closed-canopy coconut map from Descals et al. 2023 (https://essd.copernicus.org/preprints/essd-2022-463/) which helps refine the dynamics of forest conversion to plantations.
• Integration of the global mangrove watch 2020 map (https://data.unep-wcmc.org/datasets/45) to refine the maximum extent mask of mangroves which was originally defined from the 1996-2016 global mangrove watch data.
• Delineation of new tree plantation mainly in Central, West Africa, Indonesia and Malaysia.
• Improvements and corrections of errors in the Annual Change collection in the sequence of values for deforestation of old regrowth forest (areas of deforestation after regrowth can now be identified using a dedicated GEE asset).
• Afforestation can now be monitored in the Annual Change collection as the change of values 6 (other land cover) to 4 (regrowth).

List of updates integrated in the TMF 2021 products

• We improved the identification of forest conversion to commodities (classes 81-86 in the Transition map- Sub types) as follows:

  • We visually delineated missing plantation areas (mainly industrial plantations of oil palm, rubber, tea…) in Indonesia, Malaysia, Cambodia, Nigeria, Ghana and Brazil on the basis of the transition map 2020 for all areas with specific geometric shapes corresponding to plantations and very high resolution images available from Google Earth Engine (GEE) and Planet monthly composites. We also used the latest version of the Spatial Database of Planted Trees from the Global Forest Watch (accessed in February 2022) to complement missing large-scale industrial plantation from the TMF dataset.
  • We fully incorporated the smallholder and industrial closed-canopy oil palm plantations dataset from Descals et al. 2021 (https://doi.org/10.5194/essd-13-1211-2021) within the commodities mask.
  • We excluded tree plantations from the commodities mask and reclassified them into other land cover or afforestation (classes 91-93). To do so, we used the Spatial Database of Planted Trees database and more specifically using the plantation classes of acacia, eucalyptus, pine, teak or araucaria but also the class forest plantation identified in the Level-2 of the MapBiomass dataset 1985-2018 (https://www.mdpi.com/2072-4292/12/17/2735/htm). This reclassification was visually checked and completed using high-resolution images available in GEE and Planet monthly composites.
  • We used a buffer of 60m to ensure that the commodities mask did not overlap with the Open Street Map dataset (accessed in 2019) and the built-up areas (from 1975-2014 epochs) from the Global Human Settlement Layers P2016 (https://ghsl.jrc.ec.europa.eu/ghs_bu.php) (where TMF products are classified as non-forest).

• We improved the distinction between deforestation without prior degradation, deforestation occurring after degradation and multiple events of degradation. Up to TMF version 2020, we applied two conditions to consider that deforestation occurred after degradation: "a recurrence value lower than 58% or a recurrence value lower than 70% with at least 6 years without any disruption events between the degradation and the deforestation disturbances". In this new version, we now analyze the full sequence of disruptions (1982-2021) to better characterize complex trajectories such as direct deforestation, several events degradation (up to 4 events of degradation) or deforestation after degradation (single or multiple short events). Degradation is still characterized by a 2.5 years maximum duration while deforestation can be observed for a longer time. We required a minimum period of two years between two disturbance events (with no disruption detection).

  The figure below displays the values of Transition Map (2021), Annual Change (1982-2021) and disruption observation (1982-2021) for a given pixel (long:-58.57, lat: -11.40) The sequence of disruptions shows three distinct disturbance events of less than 2.5 years duration and separated by at least 2 years with no disruption observation. Previously this pixel corresponded to a tropical moist forest that has been degraded in 2005 and deforested in 2008. After corrections, this pixel is now classified as degraded forest with three stages of short duration degradation in 2005, 2008 and 2019.

  

  The update described above consequently improved the detection of the year of deforestation occurring after degradation which also concerns the date of forest conversion to plantation when a prior degradation occurred. We now provide this information as a GEE asset (see the Google Earth Engine below).

  The figure below shows the information of a pixel (long:-47.76, lat: -3.49) of tropical moist forest that has been degraded and deforested. The deforestation year in TMF version 2020 was 2001 and corresponded to the second time a disruption was observed. After corrections, the year of deforestation after degradation has been modified to 2012 as we detect a disturbance event of more than 3 years duration. We also improved the detection of the several events of short duration degradation prior deforestation (in 1998, 2001 and 2008). This improvement modifies the Transition Map but also the values of the annual change collection.

  

• We improved the separation between degradation and deforestation that started the last year (2021). Up to TMF version 2020, a threshold of 10 disruptions was used to define a deforested land. We now calculate a ratio between the annual number of disruption and the annual number of valid observation. We defined a threshold of 45% (or 23% if the disturbance started after the second half of the year) on the basis of visual identification over new logging, burning or deforestation events in 2021.

• We now integrate deforestation of regrowth forest of at least 10 years old and reclassify this trajectory in the Transition Map as a deforested land. To distinguish it from deforestation of undisturbed or degraded forest, we provide a mask of deforestation of regrowth as a GEE asset (see the Google Earth Engine below).