Rapid remote monitoring reveals spatial and temporal hotspots of carbon loss in Africa’s rainforests
Continental, regional, and local spatiotemporal patterns of carbon lossFor Africa’s primary tropical humid forest, carbon losses due to forest disturbances reached 42.2 ± 5.1 MtC yr−1 (mean ± standard deviation, where MtC yr−1 is one million metric tons of carbon loss per year) in 2019 and 53.4 ± 6.5 MtC yr−1 in 2020. Just 9 countries out of the 23 analyzed accounted for 95.0% of total gross losses in 2019 and 94.3% in 2020. These countries contain about 95.7% of all primary tropical humid forests of Africa, with the DRC accounting for 52.8%, Gabon 11.8%, the Republic of the Congo 11.0%, and Cameroon 9.8%. Of these, DRC and Cameroon were responsible for 49.3% and 19.1% of losses in 2019 and 44.7% and 20.6% in 2020. DRC and Cameroon had an annual increase of 15.0% and 36.5% respectively, between 2019 and 2020. From countries with at least 1 MtC emitted in the two years analyzed, Madagascar had the highest annual increase in carbon loss (+153.9%), while Equatorial Guinea is the only country with a decrease in carbon loss (−20.1%). Extending the carbon loss analysis for both past and future will help to better understand these variations and whether the COVID-19 global pandemic had any influence on the general increase between 2019 and 202019. While the absolute numbers for carbon loss estimates should be treated carefully and a sample-based approach should be preferred for an unbiased estimate of absolute numbers20, we focused our analysis on the trends of carbon loss at the continental, country, and local scale (Fig. 1 and Supplementary Fig. 1).Fig. 1: Carbon loss across Africa’s rainforests.We analyzed 23 countries containing primary moist forest. The aboveground carbon stock (green palette) underlies the carbon loss estimations (red palette). Several hotspots can be seen across these regions. The uncertainties of the carbon loss estimations are expressed as standard deviations and shown in Supplementary Fig. 1.Full size imageThe high temporal detail of the analysis revealed various monthly patterns of carbon losses for countries, highly related to local rainfall patterns18 (Fig. 2). Countries like Cameroon, Liberia, Nigeria, Central African Republic (CAR), and Madagascar showed a clear dry-wet seasonal variation in carbon loss per year, while the Republic of the Congo and the DRC, due to their latitudinal extent, exhibited two dry-wet season variations per year with varying intensities (Fig. 2). The seasonal variation can be explained by higher accessibility to forests during the dry months when activities related to smallholder agriculture and logging are more feasible than in the wet season when many roads become inaccessible.Fig. 2: Temporal patterns of carbon loss for the top 10 countries.We show monthly statistics for 2019 and 2020 and the associated uncertainty (black lines). We separate between high (red bars) and low (yellow bars) confidence alerts, the latter showing up for the last 3 months of 2020.Full size imageOne of the highest differences between the months with the most and the least carbon losses was found for Madagascar (72 times more carbon loss in March compared to November 2019). In CAR, the three consecutive months with the highest cumulative carbon loss (January to March 2020) contributed to 75.7% of the total annual loss (between February and April 2020), in Nigeria 73.9% (January to March 2020), Liberia 73.1% (February to April 2020), Madagascar 70.7% (September to November 2020), and Cameroon 62.2% (January to March 2020). Lower percentages were found for countries with mixed seasonality and patterns, like DRC 36.7% (January to March 2020), and the Republic of the Congo 32.8% (January to March 2020) (Fig. 2). For the latter two countries, we expect better-defined peaks of carbon loss at local scales, where climatic conditions are not mixed. The annual cumulative carbon loss (%) per country (Fig. 3) showed that Liberia, Nigeria, CAR, and Cameroon reached between 70-90% of their annual carbon loss in April, while Madagascar reached 60% in October. The DRC, Gabon, Republic of the Congo, Equatorial Guinea, and Ghana have a more gradual monthly increase of cumulative carbon loss with less contrasting seasonality effects. Monthly patterns of carbon losses between the two years analyzed resulted in a correlation coefficient of 0.94 for the CAR, 0.92 for the DRC, 0.91 for Madagascar, 0.90 for Gabon, and 0.83 for Cameroon (Supplementary Fig. 2). For the Republic of the Congo, the two years correlated 0.51. Knowing the peak months of carbon loss for each country and that these patterns are repeatable from one year to another can contribute to better target and prioritize enforcement activities, as well as predicting future patterns and early reporting of annual forest carbon losses.Fig. 3: Annual cumulative carbon loss (%) for both years analyzed, 2019 and 2020.Africa’s total cumulative carbon loss is shown with a black line. The 10 topmost emitting countries out of 23 countries analyzed are shown and represented by distinct colored lines.Full size imageSeveral hotspots of carbon losses can be seen in Fig. 1. The high spatial and temporal details of our analysis are shown in Fig. 4, where several local examples with different drivers of forest disturbances are shown, like logging roads, selective logging, mining, oil palm plantations, urban expansion, and small-holder agriculture. This kind of information, coupled with auxiliary datasets (e.g., legal concessions, protected areas) can identify the legality of forest disturbance21.Fig. 4: Local examples of approx. 10 × 10 km in extent showing different spatiotemporal patterns and drivers of carbon loss.The first column shows the carbon loss, the second column the associated uncertainty, the third column the day-of-the-year when the loss occurred, and the last column shows the monthly distribution of carbon loss and associated uncertainty for each local example. The center coordinates of each location are shown in the third column as latitude and longitude. Exact locations are shown in Supplementary Fig. 3. a Logging roads and selective logging in the Central African Republic, b mining of gold and titanium in the Republic of the Congo, c development of an oil palm plantation in Cameroon, d forest disturbance related to building a new capital city in Equatorial Guinea, and e small-scale agriculture expansion at the edge of the forest in the DRC.Full size imageImplications of rapid monitoring of local carbon lossNear-real time alerts combined with biomass maps result in spatially explicit forest carbon loss, unlike global tabular statistics of national data22,23. We provide new insights into the spatiotemporal dynamics of carbon loss with consistent assessment of accuracy that could enable transparency and completeness for countries reporting on their REDD + progress to the UNFCCC24. We provide monthly carbon loss estimates that could play a key role in local, national, and international forest initiatives for global carbon policy goals25. Such a system can be implemented with minimal costs and is based on open-source datasets and Google Earth Engine cloud computing platform26, thus enabling cost-effective national monitoring of forest carbon loss7. Providing rapid reporting on the location, time, and amount of carbon lost across Africa’s primary humid forest will help undertake immediate action to protect and conserve carbon-rich threatened forests. Furthermore, countries will be able to predict and estimate their annual carbon loss before a reporting period ends, thus having the opportunity to adjust their practices to meet their country-specific commitments for climate change mitigation initiatives.Limitations and future improvementsWe used the RADD alerts (Radar for Detecting Deforestation)18 with a minimum mapping unit (MMU) of 0.2 ha as accuracy estimates were available for this MMU. Events smaller than 0.2 ha would add to the total carbon loss but are by nature associated with higher uncertainties18. The implications of the RADD alerts using a global humid tropical forest product as a forest baseline for 201816,27,28 are twofold. First, the global nature of this product might result in inconsistencies at the local level18. Second, because the forest cover loss information used to generate the forest baseline is based on optical Landsat data, persistent cloud cover in the second half of 2018 in some areas led to missed reporting of forest disturbances, thus being detected at the beginning of 2019 by the RADD alerts. This possible overestimation of carbon loss at the start of 2019 is not an issue for a near-real-time alerting system since later months are not affected. Furthermore, the alerts do not distinguish between human-induced disturbances and natural forest disturbances18. When a new forest disturbance alert is detected, it will be confirmed or rejected within 90 days by subsequent Sentinel-1 images18. That is why our carbon loss reporting separates between high and low confidence alerts for the last three months of 2020, which is common for most forest disturbance alerting products18,29. We separated all the alerts into core and boundary pixels. Core alerts represent complete tree cover removal and we assumed complete carbon loss within a pixel. For boundary alerts, we assumed a 50% carbon loss since these mainly represent forest disturbances with partial tree cover removal. Detecting and quantifying the level of degradation remains challenging and future developments will minimize this uncertainty by providing variable percentages of degraded forest30. The timeliness and spatial details of future forest disturbance alerting products will improve with the availability of open access long-wavelength radar data from near-future satellite missions (e.g., NISAR L-band SAR in 202331), by using a combination of optical and radar forest disturbance alert products, and integration with high-resolution satellite products.We relied on an aboveground biomass baseline map from 201832, prior to RADD alerts starting from 2019. Biomass estimation for the tropical moist forests is based on ALOS-2 PALSAR-2 L-band satellite and its usage needs to account for the local biases, especially underestimating AGB values higher than 250 Mg ha−1 (ref. 32). Although we reduced this underestimation by adjusting the AGB map based on ground field data, more research is needed on providing up-to-date high-resolution aboveground carbon estimates33 that could further increase the accuracy of local carbon loss estimation. Radar-based estimation of forest carbon stocks is challenging over mountainous terrain and is less accurate in complex canopies3 and future integration of radar and optical satellite data will provide more robust estimates33. Nevertheless, new spaceborne missions (e.g., GEDI34, BIOMASS35) will provide an unprecedented amount of forest structure samples that will improve the algorithms and thus the final accuracy of aboveground biomass estimates.We focused on exploring and analyzing local carbon losses and showing high temporal and spatial patterns of carbon losses. We showed the country statistics to emphasize the temporal dynamics of carbon losses and compare the temporal profiles across our study region. Our approach was not to provide stratified area estimations36 associated with forest disturbances but we used this concept in the sense that we had a stratified sample of higher quality reference data18 to estimate the omission and commission errors and consider those in our uncertainty estimation on the pixel level. The analysis showed that omission and commission errors are small and rather balanced, and thus do not result in a major area bias for the forest disturbances. The uncertainties of the aboveground biomass product32 were adjusted for known regional biases using regional forest biomass plot data sources. With this approach, the original aboveground biomass map bias was partly corrected using a model-based approach deemed to be an alternative to a sample-based approach whenever country data are unavailable37. Our uncertainty analysis and error reduction showed that we expect only minor bias in the forest disturbance and the biomass data and the remaining uncertainties are propagated in our pixel-based uncertainty layer. More
