1 Introduction

Using satellite-based maps, Ceccherini et al. (2020) report abruptly increasing harvested area estimates in several EU countries beginning in 2015. They identify Finland and Sweden as countries with the largest harvest increases and the biggest potential effect on the EU’s climate policy strategy. In a response to comments (Palahí et al. 2021; Wernick et al. 2021) regarding the original study, Ceccherini et al. (2021) reduce their estimates markedly but generally maintain their conclusion that harvested area increased abruptly. Using more than 120,000 field reference observations to analyze the satellite-based map employed by Ceccherini et al. (2020), we confirm the hypothesis by Palahí et al. (2021) that it is not a harvested area but the map’s ability to detect harvested areas that abruptly increases after 2015. While the abrupt detected increase in harvest is an artifact, Ceccherini et al. (2020) interpret this difference as an indicator of increasing intensity in forest management and harvesting practice.

Ceccherini et al. (2020) use satellite-based Global Forest Change (GFC) (Hansen et al. 2013) data to estimate the yearly harvest area in each of 26 EU states over the period 2004 to 2018. They claim that an increase of harvested areas will impede the EU’s forest-related climate-change mitigation strategy, triggering additional required efforts in other sectors to reach the EU climate neutrality target by 2050.

2 Discussion

In their response to comments, Ceccherini et al. (2021) carry out a stratified estimate of harvested area for the combined area of Finland and Sweden with more than 5000 visually classified reference points based on manual interpretation, using high-resolution aerial images and Landsat data. They compare the time periods 2011–2015 and 2016–2018 to find a 35% increase in harvested area in the second period which is a considerable reduction compared to the original article, where a 54% and 36% increase was reported for Finland and Sweden, respectively. Although this approach is more robust than the “pixel counting” (Palahí et al. 2021) of the original article, as can be seen below, this is still a gross overestimation of the change in harvested area. The main issue is the use of Landsat to determine the timing of forest cover losses. Because Landsat became more sensitive in detecting forest cover loss over time, many losses that occurred in or before the first period are thus detected in the second period. This causes errors in the reference data which propagate in the reported estimate. Moreover, Landsat provides the primary data input for GFC, which results in circular reasoning when using Landsat as reference data for GFC. In other words, Landsat cannot be used to validate a Landsat-based product.

Furthermore, Ceccherini et al.’s (2021) argument that abrupt changes in harvested area were not observed in all countries and therefore cannot be caused by data artifacts is inappropriate because the algorithms used to create the GFC map are, due to the modeling procedure applied, inherently non-linear (Hansen et al. 2013). Unexpected changes can therefore happen in some regions but not in others.

Finally, Ceccherini et al. (2021) claim inconsistencies in GFC were unknown. Though inconsistencies in GFC’s time series have previously been reported (Rossi et al. 2019; Galiatsatos et al. 2020), this may indeed not have been well-known. However, it is a well-established fact that Earth observation data and related products can be unreliable and inconsistent (McRoberts 2011; Olofsson et al. 2014). Important interpretations and decisions should therefore not be based on “pixel counting” estimates.

We employ more than 120,000 field observations from repeated measurements in 44,000 sample plots from the Finnish and Swedish national forest inventories (NFIs) as reference data in statistically rigorous estimators in order to analyze the accuracy of Ceccherini et al.’s (2020) findings (see Appendix and our dataset (Breidenbach et al. 2021b)). We find that GFC’s ability to detect harvested areas and thinnings abruptly increases after 2015 (Fig. 1). When the ability to detect harvest improves, the overall harvested area in GFC will increase, even without a real change in management activity. As a result, more harvested areas and thinnings were detected by GFC after 2015, and this explains why the “harvested area” reported by Ceccherini et al. (2020) abruptly increases. In other words, the reported abrupt increase in harvest is to a large degree simply a technical artifact (bias) caused by the advancement of GFC over time. Ceccherini et al.'s (2020) conclusions, however, are the product and direct consequence of an inconsistent time series and are thus both incorrect and misleading.

Fig. 1
figure 1

Proportion and 95% confidence interval of correctly detected areas by GFC given change cause as represented by NFI data. A Finland. B Sweden. The inverse of the y-axis is the omission error

Full size image

Assuming the average proportion of correctly identified harvested areas before 2015 also applies after 2015, the GFC area after 2015 can be modeled without this increasing sensitivity. This indicates that the GFC recorded increases in “harvested area” of 54% and 36% in Finland and Sweden, reported by Ceccherini et al. (2020) represents an overestimate of 188% and 851%, respectively, compared to our reference data (Fig. 2). Because this modeled area still includes commission error, thinnings, and other harvests, additional calculations would be required to provide improved estimates of the actual harvested area change (Rossi et al. 2019). We further highlight that Ceccherini et al.’s (2021) more recent findings do not in any way alter or affect these basic, validated findings.

Fig. 2
figure 2

GFC harvested area estimate based on NFI plots with and without correction for an increase in GFC’s detection ability after 2015. The two top figures provide the uncorrected timeseries of GFC harvested area for A Finland and B Sweden along with their field-observed management outcomes (final fellings, other harvest, thinnings, no loss recorded in the field = commission error). The area with final fellings is relatively stable while the area with detected thinnings increases considerably after 2015. The two bottom figures provide the timeseries of GFC harvested area corrected for GFC’s increased detection ability after 2015 for C Finland and D Sweden. For the period 2016–2018, the area is estimated assuming the correct detection proportion would have stayed the same as before. Based on these corrected area estimates, there is no abrupt increase in the harvested area after 2015. See Tables 1 and 2 in the Appendix for standard errors of estimates

Full size image

In addition to generating harvested area estimates subject to systematic error, Ceccherini et al. (2020) do not provide any estimates of uncertainty and further assume all biomass in their mapped harvested areas was in fact removed. Given that a considerable share of the harvested areas in the period 2016–2018 are thinnings and not final harvests (Fig. 2), the latter results in even larger errors with respect to C-losses. Ceccherini et al. (2020) likewise assume the biomass map they utilize is accurate and without uncertainty, which is unrealistic (Næsset et al. 2020). We focus on the problems related to the harvested area estimate in Ceccherini et al. (2020) as this is the most fundamental issue and is adequate for illustrating the erroneous conclusions drawn by the authors.

We acknowledge the strong desire for sound and independently verifiable monitoring strategies driven by their potential for supporting the promotion of forest-related climate benefits (Griscom et al. 2017; Bastin et al. 2019; Brancalion et al. 2019). Without this, much hesitation has accompanied interest in mobilizing forest resources behind the climate change mitigation challenge. Earth observation remote sensing (RS) and related mapping efforts embody the promise of providing very important tools for monitoring land use change, tropical deforestation, and forest restoration (Hansen et al. 2013; Baccini et al. 2017; Harris et al. 2021). As such, they likewise hold the promise of supporting efforts to better integrate forest resources into the framework of climate change mitigation strategies.

RS products, however, can be used in ways that potentially result in estimates subject to severe systematic error, as we have seen in this and other studies (Næsset et al. 2020). Ceccherini et al.'s (2020) claim that a 30% increase of harvested area in France corresponds with national statistics has been invalidated by Picard et al. (2021). These authors find that, despite the already very weak correlation between national statistics and Ceccherini et al.’s (2020) results, this “correlation” itself was caused by one single year that was heavily affected by a storm event. In reliable surveys, Picard et al. (2021) were unable to find signs of increased harvested area in France.

Because RS data measure reflections of electromagnetic waves (e.g., visual light in the case of optical satellites) rather than the direct object of interest such as forest cover loss and carbon stock, algorithmic models are required for interpreting these reflections. Models, however, are frequently imprecise tools (Box 1976) and generally require reference data to correct their data output and thereby provide unbiased estimates (Næsset et al. 2020; Breidenbach et al. 2021a). The compilation of RS data results in nice, colorful maps and scientific-looking figures further distract attention. The collection of the required reference data, on the other hand, is tedious, expensive, and their enormous importance not well understood (McRoberts 2011). Combining the GFC map with adequate reference data into reliable estimators can prove very useful for estimating harvested area and related C-stock losses, as illustrated in various studies (Turubanova et al. 2018; Rossi et al. 2019; Galiatsatos et al. 2020; Næsset et al. 2020; Breidenbach et al. 2021a).

3 Conclusion

We certainly agree with the authors that one of the more important elements of the Paris Agreement is to “achieve a balance between anthropogenic emissions by sources and removals by sinks of greenhouse gases in the second half of this century” (UN 2015). Based on the data at hand, however, it would be erroneous to lay blame for the failure to achieve these goals at the feet of the forestry sector.

We nonetheless remain hopeful future debate over the role of the European forest sector will remain rooted in more scientific foundations. Certainly, the use of large-scale open data in carbon monitoring and reporting, as Ceccherini et al. (2020) also (as well as many others before them) propose, represents the next great trend and should generally be applauded. However, strong systematic errors in estimated results clearly need to be avoided. This demonstrates why work of this kind should always be accompanied by rigorous collection of field observations and appropriate statistical estimates. Future work should therefore continue in the direction of further combining the use of large-scale, field-based sampling methods with remote sensing data resources.