Spatiotemporally consistent data on forest height support carbon accounting, forest resource management, and monitoring of ecosystem functions at continental to global scales. Forest conversion, degradation, and restoration monitoring in the context of REDD+, the Paris Agreement, and the Glasgow Declaration requires annual forest height data time series. Forest height has traditionally been directly measured in the field or using airborne laser scanning (ALS). Lately, a sample of near-global forest height data was collected by the NASA Global Ecosystem Dynamics Investigation (GEDI) spaceborne lidar operating onboard the International Space Station (ISS). Recent advantages of Landsat data archive processing and application of the machine-learning tools enabled operational annual forest height monitoring through the integration of the Landsat observation time series with GEDI and ALS-based forest height measurements.
The Landsat data archive is the only tool that enables global multidecadal forest monitoring at medium (30 m) spatial resolution. Our team processed the entire global archive of the Landsat data from 1997 to 2020 to an analysis-ready dataset (GLAD ARD) that supports global and annual forest structure modeling. The GLAD ARD product represents a 16-day time series of normalized clear-sky surface reflectance and brightness temperature. We transformed the annual Landsat reflectance time series into a set of multitemporal metrics (reflectance distribution and phenology statistics) that support spatiotemporal consistency of the forest height modeling.
The Landsat optical data time series do not allow direct measurements of forest height. Instead, the forest structure variables are modeled by relating reference lidar-based forest height measurements to multitemporal spectral data. Non-parametric machine learning tools, such as regression tree ensembles, enable empirical model implementation. The geographic extent of the model calibration is one of the parameters that control model accuracy; locally calibrated models have higher accuracy compared to continental or global models. A near-global sample of forest height collected by the GEDI instrument supports the calibration of local models to create a single global product. We mapped global forest height by calibrating a separate regression tree ensemble model for each GLAD ARD 1x1 geographic degree tile using training data collected from the target and neighboring tiles. We generated 11,860 individual ensemble models within the GEDI data range; the neighboring tiles were defined using a 12-degree radius. Implementation of the overlapping locally trained models ensured spatial consistency of the output product and high accuracy of the map.
To select the GEDI percentile of waveform energy relative to the ground elevation (RH) metric as training data, we compared the RH metrics with the ALS-based forest height metrics at 30 m spatial resolution. The comparison was done for the regions where both ALS and GEDI data were available. For the global forest height model, we used the 90th percentile of the ALS-based canopy height as the reference. The GEDI RH95 metric had the highest correlation with the ALS-based forest height and was selected for the global forest height modeling. The ALS-based 90th percentile height, however, performed poorly in the temperate and northern managed forests, where tall trees are often left within the clearcut area. Using the 75th percentile ensured the correct representation of forest loss within the logging sites. The best matching GEDI metric for the 75th percentile ALS height is RH85, and this metric was used as calibration data to model forest height in Europe.
The GEDI data extent is limited by the ISS orbit, and no data are available north of the 52nd parallel. The ALS data are not universally available, prohibiting local model calibration in boreal forests. We employed two approaches for forest heigh modeling in northern forests. For the global map, we implemented the regional models to predict forest height in the North. Three continental models (North America, Europe, and Asia) were calibrated separately using GEDI RH90 data between 52nd and 40 parallels and manually added non-forest training over tundra and wetlands. For the forest height mapping in Europe, where the ALS data were available at national and sub-national extent for Norway, Sweden, Finland, and Estonia, we have implemented the locally trained models using both ASL 75th forest heigh percentile and GEDI RH85 metrics as calibration data.
Both global and European continental models were implemented to map forest height change from the year 2000 to 2020. To map the global forest height change, we applied the model calibrated for the year 2019 to selected years and calculated the forest height value for the years 2000 and 2020 as the median of 2000, 2001, 2003, and 2017, 2019, 2020 products, respectively. We implemented extensive filtering of the year 2000 and 2020 maps to reduce errors in the model outputs due to remaining atmospheric contamination and differences in the radiometric resolution of Landsat sensors. For the European continental model, we extracted calibration spectral reflectance data from multiple years over stable forests to ensure multitemporal model consistency. The continental model was applied annually to create a pan-European 21-years forest height time series.
The year 2019 global forest height map was compared to the GEDI set-aside validation data. The comparison yielded RMSE of 6.6 m and MAE of 4.45 m, confirming model suitability for forest monitoring and carbon accounting applications. We expect that the year 2000 and 2020 products that are using the same regression model and similar Landsat metrics have similar model uncertainties.
We also validated the forest extent and change using a statistical sample of reference data collected through visual interpretation of the Landsat and high spatial resolution data time series. Forest extent was defined using a 5 m forest height threshold. For the global time-series product validation, we used a reference sample of 1,000 units (individual Landsat 30-m pixels). The sample was allocated using a stratified random design. The strata represent stable forest, non-forest land cover, forest loss, gain, and the possible change omission. The result shows the high accuracy of the year 2000 and 2020 products, with user’s and producer’s accuracies above 94%. The forest loss and gain accuracies are lower, which illustrates the uncertainty of forest change detection, specifically, within open canopy dry seasonal forests. The accuracy of the year 2020 product for Europe was estimated using a stratified sample of 600 pixels and showed the high quality of the continental dataset; user’s and producer’s accuracies are above 87%. The forest change product validation in Europe is ongoing.
The global forest height maps for years 2000 and 2020 confirm the net forest extent loss reported by the FAO Forest Resource Assessment 2020. Both the global map (using the 5 m height threshold to define forest extent) and the FAO data show that the forest area declined by 2.4% during the past 20 years. The FAO forest extent estimate is 0.9% higher compared to our map-based estimate for both years 2000 and 2020. The national-scale year 2020 forest area estimates are also comparable between FAO and our map, yielding r^2 of 0.98 for countries with at least 10,000 ha of forests. Spatiotemporal forest height data allows us to quantify not only the total forest extent change but also the change of the extent of tall, high carbon forests that are frequently converted to shorter tree plantations or secondary forests in the tropics. Our global data shows that the area of tall forests (defined using 20 m height threshold) declined by 4.1%, nearly twice as fast as the total forest extent decline. Our findings, methods, and global and continental products provide tools supporting the implementation of international agreements toward sustainable forest use and climate change mitigation.

Radar satellite imagery from the Sentinel-1 missions is routinely used to map new disturbances in the primary humid tropical forest at 10 m spatial scale and in near real-time (weekly updates). Sentinel-1’s cloud-penetrating radar provides gap-free observations for the tropics, enabling the rapid detection of small-scale forest disturbances, such as subsistence agriculture and selective logging across large forest regions. The RADD (Radar for Detecting Deforestation) alerts were developed in cooperation with Google and the World Resources Institute Global Forest Watch program. The RADD alerts are currently operational for 45 countries across South America, Africa and Insular Southeast Asia, and are available openly via http://www.globalforestwatch.com and http://radd-alert.wur.nl. Accuracy assessment in the Congo Basin yielded high performance of the method with an estimated user’s and producer’s accuracy of ≥ 95% for events > 0.2 ha (Reiche et al., 2021).
We will provide an overview of new developments of the RADD alerts and lessons learned from expanding the RADD alerts to South America. Method advancements will be presented that include, e.g., Sentinel-1 pre-processing workflows (Mullissa et al., 2021), improving disturbance detection using radar texture, and expanding the RADD alerts to other forest ecosystems (incl. dry tropical forests). We will also provide an overview of research progress beyond the timely detection of the location of new forest disturbances into the characterization of different disturbance types. This includes the combination of multiple operational alerts, characterization of drivers, and the rapid monitoring of local carbon losses when combined with locally calibrated biomass estimates (Csillik et al., in review). We will further reflect on the use of the alerts in specific applications with a particular focus on tracking selective logging in the Congo Basin.


Reiche, J., Mullissa, A., Slagter, B., Gou, Y., Tsendbazar, N-E., Odongo-Braun, C., Vollrath, A., Weisse, M. J., Stolle, F., Pickens, A., Donchyts, G., Clinton, N., Gorelick, N., Herold, M. (2021) Forest disturbance alerts for the Congo Basin using Sentinel-1. Environmental Research Letters 16, 2, 024005. https://doi.org/10.1088/1748-9326/abd0a8.
Mullissa, A., Vollrath, A., Odongo-Braun, C., Slagter, B., Balling, J., Gou, Y., Gorelick, N., Reiche, J. (2021) Sentinel-1 SAR Backscatter Analysis Ready Data Preparation in Google Earth Engine. Remote Sensing 13, 10, 1954; https://doi.org/10.3390/rs13101954.
Araza, A. B.; Castillo, G. B.; Buduan, E. D.; H., Lars; Herold, M.; Reiche, J.; Gou, Y.; Villaluz, M. Gabriela Q.; Razal, R. A. (2021): Intra-Annual Identification of Local Deforestation Hotspots in the Philippines Using Earth Observation Products. Forests 2021, 12, 1008. https://doi.org/ 10.3390/f12081008.
Csillik, O., Reiche, J., De Sy, V., Araza, A., Herold, M. (under review) Rapid monitoring of local carbon losses in Africa’s rainforests.

Accurate information is needed to characterize tropical moist forest cover changes, to support conservation policies and to quantify their contribution to global carbon fluxes more effectively. Global and pan-tropical maps have been derived from Landsat time series to quantify global tree cover losses or tropical moist forest (TMF) changes (Hansen et al. 2013, Kim et al. 2014, Vancutsem et al. 2021). However, short-duration degradation and small disturbances (less than 0.09-ha size) are not depicted due to the limiting spatial resolution and frequency of observations of Landsat imagery. Sentinel 2 data bring finer spatial resolution and higher temporal frequency that can support more accurate forest monitoring, in particular for capturing degradation events.

To map the extent and changes of the TMF cover at 10 meter spatial resolution, we developed an expert system that exploits the multispectral and multitemporal attributes of the Sentinel 2 imagery in combination with historical information provided by Landsat to identify deforestation and degradation events since year 2015 (when S2 observations are available).
The mapping method includes four main steps: (i) single-date multispectral classification of Sentinel 2 and Landsat (7 and 8) scenes into four classes (potential moist forest cover, potential disruption, water and invalid observation), (ii) analysis of trajectory of changes from 1990 to 2021 combining the temporal information of S2 and Landsat and production of a “transition” map, (iii) identification of subclasses (tree plantations and mangroves) based on ancillary information and visual interpretation, and (iv) production of a change map at 10m resolution from year 2020 to year 2021.

For the single-date classification, multispectral clusters are defined first by establishing a spectral library capturing the distribution of spectral signatures of a few mainland cover types (moist and dry forests, savanna, bare soils, urban areas, irrigated and non-irrigated cropland, flooded vegetation, snow, ice and water) and atmosphere perturbations (clouds, haze, and cloud shadows) over the pantropical belt. An initial set of 28000 pixels was selected and labeled through visual interpretation of Sentinel 2 data to represent these land cover and cloud classes and an additional complementary set of 25000 pixels was extracted automatically using the TMF map of year 2020.

In the second step of the mapping approach, the temporal sequence of single-date classifications is analyzed for each pixel to determine the extent of the TMF domain at 10m resolution for the year 2020 and then to identify the change trajectories over the period 1990-2021 with six main transition classes: (i) undisturbed forest, (ii) degraded forest, (iii) deforested land, (iv) forest regrowth, (v) other land cover.

The uncertainties of area estimates from this new S2-TMF map will be produced from an independent reference sample of 6000 plots that has been created through the interpretation of most recent high resolution image that is available from Google Earth platform and Planet imagery. The accuracy of the S2-TMF map will be compared to the accuracy of the Landsat-TMF disturbance map that is 91.4% (Vancutsem et al. 2021).


References
● M. C. Hansen, P. V. Potapov, R. Moore, M. Hancher, S. A. Turubanova, A. Tyukavina,
D. Thau, S. V. Stehman, S. J. Goetz, T. R. Loveland, A. Kommareddy, A. Egorov, L. Chini,
C. O. Justice, J. R. G. Townshend, High-resolution global maps of 21st-century forest cover
change. Science 342, 850–853 (2013).
● D.-H. Kim, J. O. Sexton, P. Noojipady, C. Huang, A. Anand, S. Channan, M. Feng, J. R. Townshend, Global, Landsat-based forest-cover change from 1990 to 2000. Remote Sens. Environ. 155, 178–193 (2014).
● C. Vancutsem, F. Achard, J.-F. Pekel, G. Vieilledent, S. Carboni, D. Simonetti, J. Gallego, L. E. O. C. Aragão, R. Nasi, Long-term (1990–2019) monitoring of forest cover changes in the humid tropics. Sci. Adv. 7 (2021), DOI: 10.1126/sciadv.abe1603

Tropical deforestation is the main driver of the biodiversity crisis, contributes to climate change, and results in the widespread degradation of ecosystem services that are critically important for local communities. A major cause of deforestation in the tropics is the expansion of various forms of agriculture, including smallholder agriculture, agroforestry, or agribusinesses cropping and ranching. These diverse forms of agriculture are themselves embedded in a range of social-ecological and institutional contexts. Together, this produces complex deforestation patterns, with major variation in the severity, speed, and spatial patterns of forest loss. Navigating and structuring this complexity remains a major challenge for sustainability science, and hinges on robust, satellite-based data describing how deforestation takes place and how deforestation frontiers advance.
We developed a novel methodology, based on time series of annual forest extent, forest loss, and post-deforestation land cover, to derive a new generation of satellite-based metrics to describe deforestation frontier processes. We showcase this approach for the world’s tropical dry forests, which harbor some of the most rampant deforestation frontiers, yet remain understudied and neglected by policy-making and conservation planning. Specifically, we derive a set of frontier metrics to characterize deforestation frontiers for (1) the entire South American Gran Chaco, based on the entire Landsat archives for 1985 to 2020, and (2) for all tropical dry forests globally, based on Global Forest Watch data from 2000 to 2020. Using time series analyses on stacks of annual forest maps, we derive frontier metrics capturing different aspects of advancing deforestation frontiers, including baseline forest, fractional forest loss, the speed of deforestation, forest fragmentation, the level of activeness of deforestation, current extent of forest, and the post-deforestation land-cover trajectory. Together, these metrics allow to identify key types of frontiers and to map these at high spatial resolution.
For the Gran Chaco, our results show that more than 19.3 million ha of forests were converted to agriculture between 1985 and 2020. Our frontier metrics revealed a heterogeneous pattern of slowly advancing frontiers, often associated with the expansion of ranching, and rampant frontiers (i.e., frontiers of high speed and severity), often associated with the expansion of cropping. Over much of the time period we studied, pansion for ranching dominated, yet cropland expansion into forests was a major deforestation process during the mid-2000s in the Gran Chaco. Importantly, we found widespread areas initially cleared for ranching that were eventually converted to cropping, highlighting the need for considering post-deforestation land uses for better linking frontier dynamics to the underlying processes and for attributing deforestation to commodities.
Expanding our approach to the global scale confirmed that the Gran Chaco contains some of the most rampant deforestation frontiers globally. However, several other areas, especially in South America and Asia, are also characterized by rampant deforestation frontiers. Clearly, such frontiers are associated with expanding commodity agriculture, highlighting the potential of supply-chain policy response steer deforestation, and the need to provide robust monitoring data for that purpose. Additionally, our analyses revealed many deforestation frontiers that are currently in their early stages, particularly in Africa, translating into an urgent need for forward-looking sustainability planning as frontier dynamics unfold. Finally, despite high diversity in frontier dynamics, our approach identified five high-level frontier archetypes that occur globally, paving the way for comparative research and for cross-regional learning.
Satellites map land cover and all satellite-based indicators therefore require some level of translations to infer knowledge on land use change. Our concept of frontier metrics, based on high-resolution forest-cover indicators, provide a robust, repeatable and transferable way for how this translation process can be implemented to move towards a deeper process-based understanding of land-use change. Our approach can identify high-level, recurring frontier types and can therefore be a step towards more context-specific monitoring and policy responses to deforestation.

Tropical forests compose only one third of the global forest estate, but are critical to maintaining biodiversity, abating climate change, and sustaining human livelihoods. Yet, tropical forests are among the most rapidly vanishing biomes on the planet. From the dry forests of the South American Chaco to the tropical moist lowland forests of South-East Asia, tropical forests experience intense pressure from deforestation and forest degradation, and are thus at the center of global conservation initiatives. Yet, tropical moist and dry forests are important in different ways, making their reliable distinction crucial. For example, whereas tropical moist forests harbor disproportionately large portions of species and living carbon stores, tropical dry forests sustain the livelihoods of hundreds of millions people worldwide, regulating freshwater water supplies and delivering other ecosystem services in the world’s drylands. Despite their important, but distinct roles for global sustainability and their exposure to extreme threats, there is a surprising scarcity of data enabling a reliable distinction of these biomes, or a reliable quantification of their respective long-term change rates. Global assessments largely rely on overlays between generic forest-change maps and coarse biome masks. Tropical dry forests dynamics have a particularly poor documentation, with strong disagreements between available maps of tropical dry-forest extents, and no reliable data on long-term global changes.
We will present the first global wall-to-wall data product distinguishing annual spatial extents and changes of tropical moist and dry forests at 300-m resolution back into the early 1990s – a period of adamant importance for tropical forest monitoring, as it marks peaks in both tropical and dry forest losses in several regions worldwide. We developed this product by hindcasting Hansen et al.’s Global Forest Change (GFC) time-series using a multi-scale, spatially and temporally explicit machine learning approach and fusing multiple remote sensing products derived with different satellite sensors. We built our predictive model with more than 600,000 samples, collected systematically in every country to reflect regional differences in forest-management practices and long-term change patterns and drivers. We performed a spatially explicit cross-validation, including a comparison with yearly, high-resolution forest cover data mapped by an independent dataset. Moreover, we collated over 500,000 field plots distinguishing moist and dry tropical forests to infer local boundaries between the two forest types. Our historical forest mapping approach achieved high performance accuracies (R=0.90 RMSE=9.78%), while our forest classification approach performed similarly well in distinguishing tropical moist forest (F1-score=0.91) and tropical dry forest (F1-score=0.92).

The EU Forest Strategy for 2030 calls for an EU-wide integrated forest monitoring framework using, among others, remote sensing technologies. As part of this, the Commission’s EU Observatory on deforestation, forest degradation, changes in the world’s forest cover, and associated drivers is developing Earth-Observation-based monitoring tools for forests. Here, we highlight recent advances from the Observatory, particularly regarding the monitoring of European forest disturbances using remote sensing. Various components of a broader observatory system are namely being developed and tested on pilot case studies before eventual deployment for the entire EU territory. We provide a technical overview of these components, illustrate the latest research developments and show new forest-related prototype products. Reporting forest dynamics on an annual basis requires accurate mapping, characterization and causal attribution of forest disturbance events. To this end, we developed spatio-temporal methods that exploit the full multi-dimensional nature and potential of Copernicus data, while maintaining scalability for use over large areas. These novel approaches rely on the combination of a supervised deep learning model for accurate detection of disturbance events and radiative transfer modeling for detailed disturbance characterization. The large volumes of highly specific training data required for such approach are generated following a semi-automatic iterative protocol. The model performances and accuracy resulting maps are then quantified using a new multi-purpose reference set that will be made publicly available, along with the training data. Radiative transfer models are used to quantify changes in biochemical and structural plant traits associated with forest disturbances. Another key component of the system will be its ability to continuously monitor and provide near real time alerts related to forest anomalies. To efficiently compare a set of existing near real time change detection algorithms (CCDC, EWMA, CuSum/MoSum (also known as BFASTmonitor) and IQR) in an unbiased way, we implemented them all in a single package optimized for fast computation and with a common standardized interface. Results of this comparative assessment which allow for informed decisions on how to best monitor forests for different EU regions are also presented. In the future, we expect the spectrum of this system to further widen around the core components presented here, acting as a catalyzer for research and development on forest monitoring and management by a variety of stakeholders.