REDD+ and greenhouse gas reporting for the agriculture, forestry and other land use (AFOLU) sector requires land use changes to be characterized to estimate the associated greenhouse gas emissions or absorptions. It is becoming increasingly common to generate these estimates using, sample-based approaches. Known as sample-based area estimation, the technique has been widely used in recent years in the generation of activity data - particularly for estimating areas of deforestation - for REDD+ MRV. However, implementing countries and agencies have repeatedly highlighted the lack of guidance on how to address certain frequently encountered issues with this approach. This paper responds to this need by trying to address the most urgent technical issues faced by countries relating to sample based area estimation, such as how to best monitor forest changes beyond deforestation, how to account for variability between interpreters looking at the same sample, how to define the sample unit to use and how many assessments are needed per sample unit, among others. The issues were identified and prioritized based on a review of country experiences and expert consultations beginning in March 2020. For each issue, a description and recommendations are provided. Recommendations are collected and consolidated; the paper also indicates areas for future research, which should be pursued to answer the remaining questions surrounding area estimation. This paper seeks to enable donors, academia, and countries that currently use or that want to use sample based area estimation for generating activity data for REDD+ or for other purposes to delve into current good practice and existing literature and gain an understanding of the most pressing research needs in the area. It will give non-experts an overview of area estimation, its applications and limitations.
The issues are addressed in full in the white paper “Issues and good practices in sample-based area estimation” (FAO, 2021 (in press)).

Reference

FAO, 2021 (in press). White Paper: Issues and good practices in sample-based area estimation. Rome, Italy. Forestry Division.

Tropical forests store roughly half of the world's forest carbon stocks, acting as carbon sinks through a positive balance of tree growth, recruitment and mortality (Pan et al., 2011). While forest carbon stocks and sinks in the tropics have mainly been studied in structurally intact, undisturbed forests in humid ecosystems (Brienen et al., 2015; Hubau et al., 2020; Qie et al., 2017), disturbed and recovering forests are both increasing in extent (Aide et al., 2013; Lewis et al., 2015). In this respect, understanding and quantifying forest carbon stocks and sinks in disturbed or recovering forests is essential to realistically represent the current role of tropical forests within the global carbon cycle.

Here, we will showcase two examples in which remote sensing of forest dynamics using the Landsat archive and national forest inventory data are integrated to contribute towards the understanding and estimation of aboveground forest carbon stocks and sinks in tropical forests and woodlands. The first example, published in 2021 (Requena Suarez et al., 2021), integrates Tanzania’s national forest inventory data on forest carbon stocks with information on time since establishment obtained from satellite time series of forest cover. The objectives of this study are twofold: (1) explore the carbon sink capacity of Tanzania’s recovering forest and woodlands and (2) identify the environmental and anthropogenic drivers of carbon stocks in these forest ecosystems. The second example, currently under preparation, integrates Peru’s national forest inventory data on forest carbon stocks and biodiversity with information on time and intensity of disturbance obtained from satellite time series of a vegetation index (NDMI; Normalized Difference Moisture Index) in Peruvian Amazonia.
This study seeks to (1) evaluate the degree of forest disturbance in Peruvian Amazonia; (2) assess the effects of disturbance intensity and time since disturbance on carbon stocks, biodiversity and their recovery; and (3) identify the main drivers of biomass and biodiversity in disturbed forests.

Until now, largescale estimates had been lacking for both recovering dry forests and woodlands in south-eastern Africa, as well as disturbed humid forests in Peruvian Amazonia. Both examples produce estimates of carbon stock and sinks which can be applied in national or regional (Tier 2) levels of IPCC reporting. In addition, the use of existing spatial data on climate, soil fertility, human accessibility and surrounding forest cover were used to test the effects of environmental gradients and variations in human use. In this respect, large-scale assessments of recovering and disturbed forests that combine space-based forest dynamics data with national forest inventories and spatial datasets can aid in the derivation of estimates better reflect the reality of natural forests across large areas in terms of carbon stocks and sinks, as well as address knowledge gaps regarding the main drivers behind forest carbon stocks. In this respect, we expect insights derived from both examples to be used in the characterisation of disturbed and recovering forests, mainly in connection to landscape conservation and restoration planning.

References
Aide, T. M., Clark, M. L., Grau, H. R., López-Carr, D., Levy, M. A., Redo, D., Bonilla-Moheno, M., Riner, G., Andrade-Núñez, M. J., & Muñiz, M. (2013). Deforestation and Reforestation of Latin America and the Caribbean (2001-2010). Biotropica, 45(2), 262–271. https://doi.org/10.1111/j.1744-7429.2012.00908.x

Brienen, R. J. W., Phillips, O. L., Feldpausch, T. R., Gloor, E., Baker, T. R., Lloyd, J., Lopez-Gonzalez, G., Monteagudo-Mendoza, A., Malhi, Y., Lewis, S. L., Vasquez Martinez, R., Alexiades, M., Alvarez Davila, E., Alvarez-Loayza, P., Andrade, A., Aragao, L. E. O. C., Araujo-Murakami, A., Arets, E. J. M. M., Arroyo, L., … Zagt, R. J. (2015). Long-term decline of the Amazon carbon sink. Nature, 519(7543), 344–348. https://doi.org/10.1038/nature14283

Hubau, W., Lewis, S. L., Phillips, O. L., Affum-Baffoe, K., Beeckman, H., Cuní-Sanchez, A., Daniels, A. K., Ewango, C. E. N., Fauset, S., Mukinzi, J. M., Sheil, D., Sonké, B., Sullivan, M. J. P., Sunderland, T. C. H., Taedoumg, H., Thomas, S. C., White, L. J. T., Abernethy, K. A., Adu-Bredu, S., … Hemisphere, N. (2020). Asynchronous carbon sink saturation in African and Amazonian tropical forests. Nature, 579(7797), 80–87. https://doi.org/10.1038/s41586-020-2035-0

Lewis, S. L., Edwards, D. P., & Galbraith, D. (2015). Increasing human dominance of tropical forests. Science (New York, N.Y.), 349(6250), 827–832. https://doi.org/10.1126/science.aaa9932

Pan, Y., Birdsey, R. A., Fang, J., Houghton, R., Kauppi, P. E., Kurz, W. A., Phillips, O. L., Shvidenko, A., Lewis, S. L., Canadell, J. G., Ciais, P., Jackson, R. B., Pacala, S. W., McGuire, A. D., Piao, S., Rautiainen, A., Sitch, S., & Hayes, D. (2011). A Large and Persistent Carbon Sink in the World’s Forests. Science, 333(6045), 988–993. https://doi.org/10.1126/science.1204588

Qie, L., Lewis, S. L., Sullivan, M. J. P., Lopez-Gonzalez, G., Pickavance, G. C., Sunderland, T., Ashton, P., Hubau, W., Abu Salim, K., Aiba, S., Banin, L. F., Berry, N., Brearley, F. Q., Burslem, D. F. R. P., Dančák, M., Davies, S. J., Fredriksson, G., Hamer, K. C., Hédl, R., … Phillips, O. L. (2017). Long-term carbon sink in Borneo’s forests halted by drought and vulnerable to edge effects. Nature Communications, 8(1), 1966. https://doi.org/10.1038/s41467-017-01997-0

Requena Suarez, D., Rozendaal, D. M. A., De Sy, V., Gibbs, D. A., Harris, N. L., Sexton, J. O., Feng, M., Channan, S., Zahabu, E., Silayo, D. S., Pekkarinen, A., Martius, C., & Herold, M. (2021). Variation in aboveground biomass in forests and woodlands in Tanzania along gradients in environmental conditions and human use. Environmental Research Letters, 16(4), 44014. https://doi.org/10.1088/1748-9326/abe960

Emissions from tropical forest deforestation contribute to an estimated 20% of global emissions, and if these forests are restored, this could cut Greenhouse Gas (GHG) emissions by a factor of a third. During the COP26 in November 2021, ambitious commitments were made by over 100 nations to increase investments and activities to achieve the objective of halting and even reversing deforestation by 2030. To achieve these goals, financial incentives are being developed, giving rise to a host of programmes funded by International Financial Institutes (IFIs) to key Client States (CS) stakeholder institutions. Key to achieving these objectives lies in the provision of national forest monitoring data that are geographically complete, consistent, reliable, repeatable, and easy to use. It is widely recognised that Earth Observation (EO) technologies offer the opportunity to provide seamless, repeatable, accurate, and quantitatively precise information, however, adoption uptake and integration of these technologies is inconsistent between key institutions for a wide variety of reasons. Barriers can include, but are not limited to; access to suitable technology, IT-infrastructure, and expertise within the EO domain. Therefore, as part of an ESA funded project a Consortium of experts led by GAF AG, have engaged closely with IFI’s and selected CS with the aim to develop and showcase useful EO-based solutions that monitor tropical dry and humid forests, developed specifically with the aim to be integrated via knowledge transfer programmes, promoting accessibility of the geo-spatial data

From engagement with key stakeholders in the tropical countries, it was clear that spatial extent and density of tree crowns is essential information that formed a fundamental component of detecting forest disturbances and health, for the purpose of managing forest activities at the local to national scale. Reporting of accurate national forest metrics benefit from the use of a Tree Cover Density (TCD) product; it underpins the quantification of deforestation according to national forest definitions, and it is also a crucial variable in determining existing and impending forest degradation.
TCD maps are provided within the Copernicus Land Monitoring Services, which up to now have been limited to the European Economic Area (EEA). Technological and resource challenges are commonly faced when characterising typically heterogeneous tropical dry forest types. A possible expansion of the Copernicus product portfolio by new services to pan-tropical regions began under the H2020 supported REDDCopernicus project, the success of which led to the uptake of TCD products within the current ESA funded Earth Observation for Sustainable Development (EO4SD) Forest Management (EO4SD-FM) project.

Focussing here on Mozambique as an example, this paper focusses on contributions to the Mozambique Forest Investment Project (MozFIP), Mozambique Conservation Areas for Biodiversity and Development Project (MozBIO), and the Zambézia Integrated Landscape Management Program (ZILMP). The total Above Ground Biomass (AGB) in Mozambique is estimated at more than 5.2 billion tons of Carbon demonstrating the magnitude and importance of this region for Carbon sequestration. With a historically high rate of recorded deforestation - approximately 0.58 % per annum (estimated at 220,000 ha) in 2015, it is estimated that the practice contributes to 69 % of Mozambique’s entire GHG emissions, translating to approximately 46 million tonnes of CO2 each year. These combined factors of sequestration potential, and high anthropogenic disturbance, make Mozambique a natural focus for engagement with the EO4SD-FM.

Based on an example of mapping TCD in the Cheringoma district of Mozambique this paper presents the results for three components; a sampling strategy based on Very High Resolution (VHR) Web Map Services (WMS) combined with a semi-automated image segmentation workflow; the comparison of three sampling techniques for training data generation from manual digitising, point based and segmentation ; and the analyses performed on training data amount versus accuracy by using CatBoost, Random Forest (RF) and Support Vector Machine (SVM) regression.

In this study the results of using a self-contained operational system for TCD mapping on a regional to national scale are presented. The system includes components for generating training samples directly from WMS sources, tuning and running regression models and downloading and predicting on Copernicus Sentinel-2 multitemporal satellite image stacks. These components are all amalgamated within three QGIS plugins. The system was designed to be driven by domain experts. The only external requirement after the freely available QGIS software and the plugins is an internet connection. The solution provides results that are accompanied by spatially explicit metrics of uncertainty, which are presented for the Mozambique site. Further, the workflow provides users with an end-to-end, expandable solution with seamless integration into commonly used applications. The ease of use, speed, and flexibility offered by this novel EO solution, allows users from a wide range of backgrounds to utilise the tool to apply the same methodology in tropical dry and humid forests, to confidently quantify localised and national estimations of forest degradation and monitor changes in a comparable way. The solution will be used in the second Phase of EO4SD-FM to promote better the integration of EO technologies into IFI programmes and to help Forest experts and institutions to realise and expand upon the value these technologies can bring without the need for heavyweight technical knowledge or resources.

In addition to the highlighted IFI programmes, the TCD can potentially contribute to a host of additional international programmes such as; Sustainable Forest Management (SFM), REDD+ Measurement, Reporting, and Verification (MRV) activities, Forest Landscape Restoration (FLR), and the identification of High Carbon Stock (HCS) forests, especially when combined with other products. With these data underpinning evidence-based decision making, the information can be used to increase the effectiveness of targeted interventions aimed at mitigating deforestation whilst providing accurate reporting mechanisms. Using the EO based services developed and demonstrated here, IFIs and CS can demonstrate tangible results over time, using the evidence to promote and monitor implementation of performance schemes, ultimately securing more funding and achieving results-based payments under their respective commitments.

In this study, we demonstrate the ability of a new operational system to detect forest loss at a large scale accurately and in a timely manner. To do so, we produced forest loss maps every week over Southeast Asia (Vietnam, Cambodia, Laos), Gabon, French Guiana, Suriname and Guyana using Sentinel-1 data. Drivers of forest loss vary by regions, with for example massive tree plantations in Southeast Asia and gold mining in French Guiana, Suriname and Guyana.

Through the SOFT project funded by ESA, we first adapted the method developed by Bouvet et al. (2018), to Southeast Asia and produced the maps over Vietnam, Laos and Cambodia for the years 2018 to 2021. Optical-based alerts were produced over these three countries by Hansen et al. (2016), but the presence of clouds persisting during the wet season causes substantial temporal detection delays. The processing was achieved using the CNES High Performance Computing cluster. The forest loss maps produced were thoroughly validated according to the recommendations made by Olofsson et al. (2020). The estimated user accuracy was 0.95 for forest disturbances and 0.99 for intact forest, and the estimated producer’s accuracy was 0.90 for forest disturbances and 0.99 for intact forest, with a minimum mapping unit of 0.1 ha, in Southeast Asia.

Cambodia has 69 protected areas covering a large proportion of the country, i.e., approximately 40%. However, we estimated the part of forest disturbances in protected areas in Cambodia in 2018, 2019 and 2020 to 46%, 57% and 52%. Approximately half of forest disturbances in Cambodia occurs in protected areas, which emphasizes the lack of efficiency in the protection and conservation of natural resources and biodiversity in protected areas.

The map also evidences the striking difference between low forest loss rates in Northern Vietnam versus high forest losses currently happening in Northern Laos, which is explained as follows. The impacts of the transboundary displacement of forest pressures from Vietnam to Laos and Cambodia are substantial and have accelerated rapidly over the past decade, with globalization and the acceleration of resource flows. The fragmentary adoption of REDD+ across forest nations leaves room for the displacement of deforestation from early-adopters to late-adopters of REDD+. Vietnam, an early adopter of REDD+ that has experienced significant reforestation despite exponential growth in exports of key forest-risk commodities, sourced in large part from Laos and Cambodia (Ingalls et al., 2018).

On an annual basis, the forest loss areas detected using our method are found to be rather similar to the estimations from Global Forest Watch (Hansen et al., 2013), with the largest difference (26%) being found for Laos in 2020. These results highlight the fact that this method provides not only quick alerts but also reliable detections that can be used to calculate weekly, monthly, or annual forest loss statistics at a national scale.

We automatized the method and expanded the forest loss mapping to Gabon, French Guiana, Guyana, and Suriname from the year 2018 to present, through the TropiSCO project funded by CNES. The resulting maps, updated daily, are browsable and downloadable via a webGIS that will be presented here.


References:

Bouvet, A., Mermoz, S., Ballère, M., Koleck, T., & Le Toan, T. (2018). Use of the SAR shadowing effect for deforestation detection with Sentinel-1 time series. Remote Sensing, 10(8), 1250.

Hansen, M. C., Potapov, P. V., Moore, R., Hancher, M., Turubanova, S. A., Tyukavina, A., ... & Townshend, J. (2013). High-resolution global maps of 21st-century forest cover change. science, 342(6160), 850-853.

Hansen, M. C., Krylov, A., Tyukavina, A., Potapov, P. V., Turubanova, S., Zutta, B., ... & Moore, R. (2016). Humid tropical forest disturbance alerts using Landsat data. Environmental Research Letters, 11(3), 034008.

Olofsson, P., Arévalo, P., Espejo, A. B., Green, C., Lindquist, E., McRoberts, R. E., & Sanz, M. J. (2020). Mitigating the effects of omission errors on area and area change estimates. Remote Sensing of Environment, 236, 111492.

Forests store carbon, contribute significantly to the formation of new groundwater, provide cooling, and are a unique habitat to a wide variety of species. However, the past drought years in large parts of Central Europe have demonstrated how vulnerable many forests are. In 2018, 2019 and 2020, periods of droughts and pests such as the bark beetle have particularly hit conifer monocultures hard in Germany. Additionally, forest fires and the expansion of open-cast mining (coal, gravel, sand, etc.) have contributed to forest damage and forest loss. Since the droughts of recent years, the abundance of bark beetles and other calamity factors highly differed between regions, the extent of forest damage varies greatly. A yearly in-field forest status assessment at fixed sampling plots is realized to identify how severe this forest damage is in Germany (BMEL 2020). Since this assessment cannot be done at full coverage, a satellite-based monitoring system is needed to map the forest damage and forest loss at high resolution and country-scale.

In order to address this need, the “Forest Monitor Germany”, which uses a long time series of Sentinel-2 data with 10m spatial resolution to assess changes in the forest status, was developed (https://map3d.remote-sensing-solutions.de/waldmonitor-deutschland/#). All Sentinel-2 data of the main growing seasons from 2016–2020 were used, while data from 2016 and 2017 were used as base periods to compare the impact of the drought years 2018–2020 on forest stands. Established vegetation indices were calculated on these time series and following trend analyses were applied in order to identify positive or negative changes in forest stands in regard to forest biomass and foliage water balance. A classification of the forest damage showed that 48,165 hectares of the deciduous forest area and 201,741 hectares of the coniferous forest experienced total tree loss since 2018, with very large regional differences. North Rhine-Westphalia (10.3%), Saxony-Anhalt (8.2%) and Hessen (8.1%) have suffered the greatest losses of coniferous forest in recent years. The forest areas with partial damage, as can be seen in the biomass trend map, are much larger. The forest water balance trend particularly indicates deciduous forest stands with crown damages caused by drought stress. This effect is especially visible at forest edges where an adjacent conifer forest experienced total tree loss and the deciduous trees were suddenly exposed to more direct sunlight. A qualitative validation of the forest status assessment was done via field surveys, drone data acquisition and aerial surveys.

Besides the trend analysis for the assessment of the forest status, a machine learning based classification of the dominant tree types was done. Different forest communities with various tree types react differently to climate change and require different management and protection measures. A country-wide map of the dominant tree species can help to better characterize the latest forest damage and also to identify areas at risk. The dominant tree species and the distribution of tree communities have been recorded at sampling points in Germany (Riedel et al. 2017). These data were used here as reference data for a machine learning approach to generate a country-wide mapping of the dominant tree types. In order to retain a high representativeness of the reference data for the different tree species, only forest plots that are mostly composed of a single tree species (coverage of that species >= 80%, i.e. almost pure stands) were used as training data. The tree types to be considered were limited the sample sizes in the reference data (appropriate sample size per tree type is required). To account for regional differences in forest ecosystems, the reference data was spatially split and regional machine learning models were established. By linking these data to time series of Sentinel-2 data from 2017 (pre-drought scenario), the different phenologies and spectral characteristics of tree species could be derived and the dominant tree species being determined. XGBoost was used as machine learning algorithm to classify Sentinel-2 time series image stacks (1.6 TB in total) and the regional dominant tree type classifications showed F1-scores between 0.6 and 9.96 depending on the region and tree types (calculated based on 30% test data).

The information layers in the Copernicus-based Forest Monitor on the dominant tree types and forest status assessment, allow for a country-wide and continuous forest characterization, in order to support the planning of reforestation, promote establishing climate-change resilient forests and to provide geographic information relevant to forestry in practice and science. The range of geodata is openly provided online to be used by a wide various interest groups such as forest owners, municipalities, forestry authorities and scientists. The information of the Forest Monitor will be further expanded also through cooperation with interested institutions that would like to contribute. The aim of the forest monitor is to regularly provide country-wide information about the forest status and will also be used to monitor the forest development in areas with previous total tree loss. All used data and applied methods allow to establish a Sentinel-based forest monitoring for other European countries.

References:
BMEL (2020): https://www.bmel.de/SharedDocs/Downloads/DE/Broschueren/ergebnisse-waldzustandserhebung-2020.pdf?__blob=publicationFile&v=11
Riedel T., Hennig P., Kroiher F., Polley H., Schmitz F., Schwitzgebel F. (2017): Die dritte
Bundeswaldinventur (BWI 2012). Inventur- und Auswertemethoden, 124 S

It is critical for forest management and sustainable forest ecosystems to have access to detailed and spatially explicit information on the state of our forests, including tree species composition. Such information becomes increasingly relevant on regional and national levels as changing climate conditions exert stress on the forest ecosystems. Typically, species composition is monitored through detailed, sample-based forest inventories. Remote sensing is a promising tool to augment statistical estimates with high-resolution species composition maps. Such maps are not yet available for entire Germany and methodological approaches for national-scale tree species mapping need further research to provide reliable results.

Our goal is to develop an area-wide tree species map from combined Sentinel-2 and Sentinel-1 time series for entire Germany. For achieving this goal, we addressed a number of challenges such as 1) building an automated workflow to process, clean, and analyze large amounts of satellite data, and 2) learning how to use the National Forest Inventory (NFI) to train classification models for satellite data. While NFI is designed to statistically estimate forest attributes like tree diameter, tree height, age, and species distribution with high accuracy at national level, the inventory is not optimized for use in spatially explicit, wall-to-wall data analyses based on satellite data. Among others, geolocation uncertainties of the Global Navigation Satellite System (GNSS) preclude a precise matching of inventory plots and satellite pixels.

We processed all available Sentinel-2 imagery of 2017 and 2018 in the Framework for Operational Radiometric Correction for Environmental monitoring (FORCE) to generate cloud free, radiometrically corrected and co-registered time series. Using a radial basis convolution filter ensemble, the Sentinel-2 observations were processed to a cloud free time series at 5-day intervals. This time series was combined with terrain-corrected, monthly Sentinel-1 backscatter composites and environmental data on topography, meteorology and climate. The Sentinel-2 time series capture the spectral and phenological characteristics of tree species, whereas Sentinel-1 data provide information on the physical and structural properties of tree species. We incorporated information on topography, and data on annual and long-term temperature, precipitation and soil moisture to account for environmental gradients over Germany, which increase the variation in reflectance characteristics for each species.

By using NFI data as ground truth, we were for the first time able to develop and train machine learning (random forest, support vector machine) and self-learning algorithms capable of generating a wall-to-wall tree species map for Germany. One approach to overcome the impact of geolocation errors in the NFI, is to include only homogeneous forest stands as training samples, i.e. information from plots where the deviation of NFI-plot location is less problematic. However, such an approach will create a data set biased towards species frequently occurring in pure stands, while species mostly present in mixed stand settings will be underrepresented. Alternatively, we implemented a self-learning approach to address the challenges related to the NFI surveying method and the imbalance of ground truth data in the process of training machine learning classification models. An initial set of supervised classification models is trained on samples from homogeneous forest stands. Those classifiers are used to label samples from mixed forest stands for which the general species composition is known through the NFI, but the inference of spatial species distribution within the stand is impeded by the geolocation error of the NFI sample. The resulting newly labeled samples can then augment the initial training data set to build a final, more balanced tree species classification model.

We then estimate resulting map accuracies to assess the capability of the proposed data and methods for a wall-to-wall tree species mapping at national scale. First results show user´s and producer´s accuracies between 71% and 95% for dominant species in Germany (including pine, spruce, beech, oak, and larch), while species occurring less frequently or which are present mostly in mixed forest stands (e.g. lime, maple, ash, fir), remain challenging. To address challenges related to the variation in environmental conditions over Germany, we identify key variables within the environmental data which are most beneficial for performing large-area tree species classifications. Based on the results presented, we discuss remaining challenges in using NFI data for mapping tree species at the national level and potential approaches to exploit NFI data more efficiently in future tree species mapping efforts. Our analyses and findings not only allow to further improve national-level tree species mapping with medium to high resolution data for Germany, but also provides guidance for similar approaches in other countries where ground-based inventory data is available.