Detecting and predicting forest degradation: A comparison of ground surveys and remote sensing in Tanzanian forests

cover. Here we present a field protocol for rapid assessments of forest condition. Using extensive field data from Tanzania, we show that a focus on remotely- sensed deforestation would not detect significant reductions in forest quality. Radar- based remote sensing of degradation had good agreement with the ground data, but the ground surveys provided more insights into the nature and drivers of degradation. We recommend the combined use of rapid field assessments and remote sensing to provide an early warning, and to allow timely and appropriately targeted conservation and policy responses.


| 269
AHRENDS Et Al. cover. Here we present a field protocol for rapid assessments of forest condition. Using extensive field data from Tanzania, we show that a focus on remotely-sensed deforestation would not detect significant reductions in forest quality. Radar-based remote sensing of degradation had good agreement with the ground data, but the ground surveys provided more insights into the nature and drivers of degradation. We recommend the combined use of rapid field assessments and remote sensing to provide an early warning, and to allow timely and appropriately targeted conservation and policy responses.

Summary
• Tropical forest degradation is widely recognised as a driver of biodiversity loss and a major source of carbon emissions. However, in contrast to deforestation, more gradual changes from degradation are challenging to detect, quantify and monitor.
Here, we present a field protocol for rapid, area-standardised quantifications of forest condition, which can also be implemented by non-specialists. Using the example of threatened high-biodiversity forests in Tanzania, we analyse and predict degradation based on this method. We also compare the field data to optical and radar remote-sensing datasets, thereby conducting a large-scale, independent test of the ability of these products to map degradation in East Africa from space.
• Our field data consist of 551 'degradation' transects collected between 1996 and 2010, covering >600 ha across 86 forests in the Eastern Arc Mountains and coastal forests.
• Degradation was widespread, with over one-third of the study forests-mostly protected areas-having more than 10% of their trees cut. Commonly used optical remote-sensing maps of complete tree cover loss only detected severe impacts (≥25% of trees cut), that is, a focus on remotely-sensed deforestation would have significantly underestimated carbon emissions and declines in forest quality.
Radar-based maps detected even low impacts (<5% of trees cut) in ~90% of cases.
The field data additionally differentiated types and drivers of harvesting, with spatial patterns suggesting that logging and charcoal production were mainly driven by demand from major cities.
• Rapid degradation surveys and radar remote sensing can provide an early warning and guide appropriate conservation and policy responses. This is particularly important in areas where forest degradation is more widespread than deforestation, such as in eastern and southern Africa.

K E Y W O R D S
biodiversity conservation, carbon emissions, community-based forest management, East Africa, global forest watch, human disturbance, synthetic aperture radar, village land forest reserves

| INTRODUC TI ON
Large areas of tropical forest are degraded through human impacts such as overexploitation, fragmentation, pollution, exotic species invasion and fire (Sloan & Sayer, 2015). While there is no globally agreed definition for forest degradation, it can be broadly defined as changes to a forest stand resulting in the long-term reduction of particular attributes and functions such as biodiversity, and the potential supply of goods and services (FAO, 2011;Ghazoul et al., 2015). Deforestation-the complete replacement of forest by another land use-is easier to define, detect and monitor, and consequently has been the focus of global policy development (Sasaki & Putz, 2009). As a result, the impacts of forest degradation on biodiversity and carbon balances are comparatively poorly understood, but they are likely to be substantial (Alroy, 2017). For instance, recent studies have shown that carbon emissions from forest degradation may have been underestimated and could account for as much as 25%-69% of the combined gross carbon losses due to deforestation and degradation in the tropics (Baccini et al., 2017;Berenguer et al., 2014;Pearson et al., 2017).
Significant progress has been made with measuring deforestation and forest degradation from space (Woodcock et al., 2020).
Changes in tree cover can now be monitored at high spatial and temporal resolution, providing policy makers and conservation planners with an unprecedented wealth of data to guide interventions (Blackman, 2013;DeVries et al., 2015;Fuller, 2006). The technology is also increasingly available to non-specialists (Asner, 2009). While there are many easily accessible datasets to assist national and global monitoring of forest cover (e.g. Hansen et al., 2013;Miettinen et al., 2011;Sexton et al., 2013), remotely-sensed forest degradation data are sparser and more challenging to obtain. At a country level, quantitative assessments of degradation are often lacking (Romijn et al., 2015). Radar data hold particular promise as they overcome the challenges presented by cloud cover and variable phenology, and they correlate with changes in biomass (McNicol et al., 2018;Mitchell et al., 2017;Ryan et al., 2012). However, using such data sources for detecting and quantifying degradation from space remains limited by the extent to which degradation is associated with a reduction in canopy cover and/or biomass (Ryan et al., 2012). Airborne radar and light detection and ranging (LiDAR; Ene et al. 2017), as well as the use of unmanned aerial vehicles (Baena et al., 2018;Ota et al., 2019) can provide higher resolution data, but these technologies require expertise, lack global coverage and historical archives, and can be prohibitively expensive. Ground-based sensing methods such as hemispherical photographs (Fournier & Hall, 2017) and terrestrial LiDAR (Decuyper et al., 2018) used to quantify stand structural attributes also hold promise, but again, their implementation requires expertise.
At the other end of the spectrum, there are detailed field assessments (Thompson et al., 2013), such as permanent sample plots for assessing changes in forest vegetation. Collecting data on species, stem diameter, height, crown cover and various biotic and abiotic parameters, they are an extremely important tool in biodiversity and environmental research (Baker et al., 2017), and are used to locally characterise biodiversity, growing stock, biomass, carbon, ecosystem function and impacts of degradation. However, permanent plots are also labour intensive and time consuming to set up, and surveying them requires expertise. Consequently, few countries conduct exhaustive plot-based inventories as part of their national forest reporting, and even fewer consistently monitor them (FAO, 2011). In addition, while permanent plots are essential to understand the impacts of degradation, they are often not the most effective method to understand the extent and patterns of degradation itself. Unless they are systematically placed to cover an entire area at high density, they rarely capture the breadth of degrading activities that occur.
On the contrary-the presence of researchers and permanent tags on trees may deter illegal activities. Plots are also often placed in a stratified random or subjective fashion, that is, purposefully located in pre-selected areas viewed as representative of a given vegetation type and/or level of disturbance. In addition, as degradation is generally not the main focus, it is often not quantified in a robustly comparable and systematic way.
Consequently, while countries increasingly monitor wall-to-wall forest cover change using remote sensing, and they also have some inventory data, they still lack representative quantitative data on forest degradation (Romijn et al., 2015). Difficulties with monitoring forest degradation and associated gaps in policy interventions create opportunities for unregulated and/or illegal logging and corruption. There can be a tendency to shift the blame for forest loss among actors, whereby existing prejudice against already marginalised groups such as farmers practising shifting cultivation or charcoal producers may be reinforced (Hosonuma et al., 2012;Ryan et al., 2014). Knowledge of which forests are degraded, where degradation is likely to spread to next, and what the main drivers are is vital for formulating appropriately targeted policy interventions and management.
Here we present a framework protocol for rapid areastandardised assessments of forest condition. The protocol sits in the middle of the spectrum between detailed ground surveys and remote sensing, and its implementation does not require professional training. The protocol assesses human use and disturbance, which depending on their levels and the forest type may lead to a deterioration of stocks and services, and thus degradation.
Using the example of threatened and highly biodiverse forests in Tanzania, we investigate: 1. how ground data collected using this protocol compare to remotely-sensed datasets; specifically, radar-based maps of biomass change (McNicol et al., 2018) and commonly used maps of complete tree cover loss (which underpin 'Global Forest Watch'; Hansen et al., 2013); 2. the value of ground data for understanding and predicting degradation in combination with spatially explicit models (e.g. whether data collected using this approach in 1996-2010 could have predicted human impacts in 2020).
The overall aim is to assess whether these rapid assessments are a useful addition to remote sensing and detailed vegetation assessments in (permanent) plots in informing conservation policy and practice.

| Protocol overview
The method presented here rapidly quantifies standing woody resources and resource extraction in forests with a view to gauging forest condition (Frontier Tanzania, 2001). While the protocol is flexible and can be adjusted to the target vegetation and area, methodological details naturally need to be standardised to facilitate comparisons. The assessment is done along transects, which typically have a width of 10 m. Their length is variable and can be adjusted to the target vegetation type and forest size. The transects are located in either a random, stratified random or systematic fashion, and should cover the forest edge as well as the interior. Within each transect, all trees, as well as stumps and other signs of human use (such as charcoal production or clearance for agriculture) are recorded. The minimum assessment threshold is typically 5 cm diameter at breast height (dbh; measured 1.3 m above ground), but this can be adjusted to the type of vegetation being surveyed. In its simplest form, the method focusses on assessing the number of cut trees versus those that are (left) standing or died naturally. Size categories can be added to distinguish cutting for different end uses. Depending on the aims of the sampling, recording can consist of simple counts within categories, or include more detailed information such as diameter (over bark), height, species identification and voucher collection. Identifying at least the commonly used timber species will indicate resource preference and hint at the likely nature of the market behind that-for example, whether trees are cut for local use or international export (Furukawa et al., 2011) (noting that timber trade names often refer to collectives of species and/or an entire genus, i.e. overharvesting of individual species can be masked when using trade names only). However, the time spent collecting, measuring and identifying trades off against the primary aim of the method-to rapidly cover many areas, often with the help of non-specialists, in order to obtain reasonably reliable estimates of degradation and to support the identification of areas in need of conservation interventions. A detailed protocol and a recommended set of core measurements are provided as part of the Table S4.

| Study area
The study area (see also Methods S1) spans the Eastern Arc Mountains and part of the coastal forests, both of which are of global importance for biodiversity conservation due to high levels of localised endemism (Mittermeier et al., 2011;Olson & Dinerstein, 2002;Stattersfield, 1998). These forest systems also provide critical ecosystem services to local communities and the nation as a whole (Ashagre et al., 2018;Fisher et al., 2011;Schaafsma et al., 2014;Swetnam et al., 2011). In southern Africa (here defined as roughly −1° to −34° latitude), the livelihoods of an estimated 150 million people are thought to be dependent on the goods and services provided by woodlands and forests (Ryan et al., 2016). Rapid urbanisation and population growth mean that demand for wood products is substantial and increasing, with fuel wood being the main source of energy for over 90% of the population (Bailis et al., 2005). The Tanzanian forestry sector-both formal and informal-is an important source of income, GDP and employment (Doggart et al., 2020;United Republic of Tanzania, 2001). While the trade in wood products is often smallscale and livelihood driven (Cavanagh et al., 2015), wood is also exported to generate foreign revenue (Lukumbuzya & Sianga, 2017).
Exact figures are difficult to obtain (Lukumbuzya & Sianga, 2017), but although Tanzania has a comprehensive legal framework for the conservation and management of forest resources, and although the forests studied here mostly occur in protected areas, overharvesting is likely to be widespread (Milledge et al., 2007). An ability to monitor and to identify drivers and patterns of forest loss and degradation is vital to the conservation of these forest systems, and to ensure the long-term provision of forest resources for sustainable livelihoods.

| Field data
The data used for this example application were collected between where there was a mismatch in recorded length and/or locality, and the length or locality inferred from the transects' GPS coordinates.

| Comparison with remotely-sensed datasets
We compared the ground data against two remotely-sensed datasets: 1. widely used maps for tree cover loss produced by the initiative 'Global Forest Watch' (Hansen et al., 2013), hereafter GFW, which are based on Landsat data and assess complete canopy loss at an approximate resolution of 28 m on the ground;

| Modelling and predicting degradation
We used a spatially explicit modelling approach to investigate which factors were most influential in explaining the spatial patterns of degradation, and whether the spread was predictable. Models were developed using Boosted Regression Trees-an ensemble method, which combines regression trees and boosting, and fits multiple simple regression trees in a forward iterative fashion. The algorithm is able to fit complex non-linear patterns and interactions, and handles different type of predictor variables (Elith et al., 2008). We focussed on three dependent variables: (a) density of charcoal kilns, (b) percentage of poles (stems 5-15 cm dbh) cut and (c) percentage of trees (>15 cm dbh) cut. A transect constituted an individual data point. For modelling the percentage of trees cut we discarded transects with an overall tree density <50 ha −1 and no reported logging (n = 25), assuming that in these areas there were hardly any trees to be cut in the first place. We considered 15 candidate predictors representative of physical accessibility, likely demand, availability of resources, forest management type and tenure (Tables S1 and   S2). For each dependent variable we tested eight models with different (pre-selected) combinations of predictors (Table S3), including a correction for spatial autocorrelation. The final models were selected based on model performance when validated against test data (cross-validation correlations on up to 25% of randomly set aside test data) and maximum parsimony in terms of the number of predictors used (Table S4). Further details on model settings, F I G U R E 1 Transects and field data. (a) The location of the disturbance transects and percent tree cover according to Hansen et al. (2013). Note that this includes tree crops, for example cashew nut, explaining the large areas of tree cover outside reserves (denoted by thin black lines). (b-d) Kernel density maps of different forest condition measures: percentage of poles cut (5-15 cm dbh) (b), larger trees cut (>15 cm dbh) (c), and the density of charcoal kilns (d). The bold black line indicates the area to which models were extrapolated (see overview map; in a-d only partly visible) parameterisation and performance are summarised in Tables S3-S5, and software notes are provided in Methods S2. In order to test the predictive ability of the models we extrapolated them at 1 km resolution for all ~12,000 km 2 of forest reserves in the study area, using predictor values for 2020 (from scenarios developed in 2010; Swetnam et al. 2011). These scenarios (broadly correctly) predicted population to increase at a rate of 3% annually, but they are conservative in that they did not make predictions around infrastructure expansion. The predictions were then compared to actual tree cover losses recorded by GFW and local reports on degradation.

| Observed rates of tree cutting
Tree cutting (here ≥5 cm dbh; see Notes S1 for trees >15 cm dbh) occurred in all but one forest between 1996 and 2010. Over one third of forests surveyed during this time had at least 10% of trees ≥5 cm dbh cut (mean among transects). Cutting levels were highly variable across forests, ranging from 0% to 81% with a mean of 10% (±15% SD) and a median of 5% (±6% MAD [median absolute deviation]). The availability of standing trees was greatly reduced in some forests, being as low as <100 stems ≥5 cm dbh per ha in some of the most degraded forests (as opposed to >1,000 in some of the least degraded forests, and a mean stem density of 849 ± 89 SE). Losses were particularly severe in the lowland coastal forests (mean across forests 20% ±28% SD; median 8% ±8% MAD), which are in direct vicinity of Dar es Salaam, a major centre of demand. Cutting levels for larger trees were similar to those of trees ≥5 cm dbh (Notes S1).
While the above statistics represent tree cutting over several years (the lifetime of a stump), the density of recent stumps can be seen as indicative of offtake rates at a given time (with a recent stump generally being 6 months or maximally 1 year old). On average (among forests) there were 3 (±0.74 SE) recent stumps >15 cm dbh per ha between 1996 and 2010. If logging rates were thus three to six trees per ha and year, then some 2.2-4.3 million trees >15 cm dbh would have been felled annually across the forest reserves in the study area (here restricted to ~7,200 km 2 with tree cover ≥50% according to GFW).
Using a very simple above-ground tree biomass function (Chave et al., 2001;FAO, 2011) (which does not assume any knowledge of species or stand-level wood densities) this would be equivalent to a gross carbon loss of 0.41-0.82 TgC per year if the cut trees were 20 cm dbh. However, establishing above-ground carbon is extremely challenging without detailed dbh measurements and at least approximate wood density estimates. In addition, recent tree cutting was highly spatially and temporally clustered. While our data thus did not allow for a robust quantification of annual carbon losses between 1996 and 2010, they did however indicate that losses were substantial. In addition, there was evidence for an increase in cutting rates over the 14 years covered by the data-from less than one tree per ha and year (approxi- averages for wood-specific gravity will considerably enhance the accuracy of these estimates.

| Comparison with remotely-sensed datasets
There was broad agreement between the spatial patterns of tree (cover) losses recorded in the field and by GFW. However, as one would expect, more subtle degradation was not picked up by this dataset focusing on complete tree cover loss in ~28 × 28 m cells.
GFW reported tree cover losses for only 20% of the transects that recorded new tree cutting between 2000 and 2005. The larger the proportion of cut trees the more often GFW detected loss (Table 1).
A very similar picture emerged when looking at a lower dbh threshold of ≥5 cm dbh (Table S6).
To illustrate this with specific examples, Figure 2 shows a comparison of ground data and remotely-sensed data for three coastal reserves visited in 2004. While GFW detected some canopy losses between 2000 and 2005 (affecting 2% of the area with ≥50% canopy cover in 2000), degradation on the ground was already severe (with a mean of 11 ± 7 SD recently cut trees ≥5 cm dbh, and 10 ± 7 SD charcoal pits per ha). GFW record large losses from these areas in the following years (26% of the area with ≥50% canopy cover in 2000), confirming the early warning signals provided by the ground data. Indeed, a field survey in 2016 estimated that, since 2004, the density of trees in these areas had halved, with timber tree densities having dropped threefold, and above-ground carbon being reduced by 40% (Ahrends et al., 2020). In one of the reserves (Vikindu) large areas of forest had entirely disappeared by 2016 (Figure 2l).

(h) [as in a] (i) [as in b] (j) [as in b] (k) [as in c]
Transects recording recent cuts

(l) Vikindu in 2016 (m)
No recent cuts >0 -1% >1 -5% >5 -10% >10% ( ) third of these transects (Table 2 and Table S7). As above, the larger the percentage of cut trees, the more often losses were detected from space. The field data did not allow for a robust quantification of specificity (

| Modelled predictions of resource harvesting
Forest resource extraction increased steeply with accessibility and proximity to centres of demand ( Figures S2-S4). Particularly in the case of charcoal production, and to some extent in the case of tree cutting, models that only considered local factors such as population density and management type performed less well than models that included predictors representative of city distance and wider population pressure (with a correlation [R] between predictions and test data under 10-fold cross validation of 0.57 as opposed to 0.75 in the case of charcoal burning, and 0.62 versus 0.68 in the case of tree cutting; Table S4). Protected area management explained some variation (Tables S4 and S5), with harvesting being highest in unreserved areas. However, it is important to note that the reserve categories conflate a range of factors, for example, all productive reserves analysed here were situated at Tanzania's easily accessible coast.
In addition, sample sizes were unequal (e.g. there were over 400 transects for 54 government forest reserves, and only 27 transects for 13 reserves on village land). Management on its own explained comparatively little variation (with cross-validation correlations of 0.39-0.56), which will in part be due to the data inadequacy mentioned above, and in part due to the overriding influence of demand and accessibility. For more details see Figure S5.
The relative importance of predictors differed for the different types of disturbances. Spatial patterns in tree cutting were almost entirely explained by urban population pressure (a distance decay function of population density; Table S1), with additional variation accounted for by distances to Dar es Salaam, roads, major cities, and steepness of terrain. Patterns in charcoal production were also mainly related to distance to Dar es Salaam and population pressure. Pole cutting, on the other hand, was best explained by a multitude of factors, including management, distances to Dar es Salaam, roads and cities, and local population density (Table S5). In interpreting the relative importance of predictors, it is important to note covariation and a degree of inter-exchangeability between them (Table S2). For instance, dropping population pressure from the full model only had a moderate effect on model performance as long as population size and city distance where still present. However, overall there was a notable difference between tree cutting and charcoal production on the one hand (almost entirely explained by variables related to accessibility from urban centres), and pole cutting on the other hand where local population density and management played a greater role in explaining the variation.
All final models performed reasonably well, achieving 10-fold cross validation correlations between 0.68 and 0.78 (Table S4).
When setting aside 20% of the reserves as test data, it was generally possible to predict the top three most degraded forests from the rest of the data.
In order to broadly investigate whether the model for tree harvesting (>15 cm dbh) was able to indicate areas under future threat, we extrapolated the model to ~2020 and compared the predictions to tree cover losses recorded by GFW between 2000 and 2018 ( Figure 3) and local reports (see below). There was general agreement between the areas predicted to face high levels of cutting by ~2020 and tree cover loss detected by GFW (Figure 3). Obvious differences arose in areas  Hill southwest of Iringa has lost a lot of tree cover due to plantation rotation, but according to local reports the non-plantation natural forest is not impacted by degradation (BirdLife International, 2013).
In several other areas the model predicted high levels of tree cutting and GFW did not report major losses; here the modelled predictions were generally confirmed by local reports suggesting that degradation has occurred, but may not (

| D ISCUSS I ON
Here, we presented a tested protocol for rapid quantitative assessments of degradation in the field, and we compared data collected with this method in Tanzanian forests with optical and radar-based remotely-sensed datasets. Covering over 600 ha our field data allowed for one of the first large-scale independent tests of these spatial datasets in southern Africa. Radar-based maps (McNicol et al., 2018) appeared to perform well, with even low levels of tree cutting generally coinciding with the detection of biomass loss.
However, our study also suggests that there still is an important role for field data, which provided valuable additional information on the types of degradation and likely drivers. For instance, patterns in the field data implied that a major driver of forest degradation is demand for woody resources emanating from larger cities-a pattern that has also been confirmed in radar-based assessments (McNicol et al., 2018). The field data additionally allowed for a finer differentiation of the underlying processes, suggesting, for example, that it is specifically the urban demand for timber and charcoal which drives a lot of harvesting (whereby charcoal production was particularly high close to Dar es Salaam, whereas timber cutting was more widespread), with important consequences for where and how to target conservation interventions.
Degradation was pervasive in the study area, meaning that a focus on deforestation would significantly underestimate losses of carbon and declines in forest quality. Indeed, the 'Global Forest Watch' data (GFW), which are commonly used in national forest inventories and conservation assessments, and which measure complete canopy loss at a 28-m spatial resolution, did not routinely detect even high levels of cutting associated with severe impacts on the ground in terms of loss of natural vegetation and carbon. This echoes findings from other studies which show that small-scale deforestation tends to be underestimated by GFW, particularly in areas with low and/or seasonally dry woody cover (Bos et al., 2019;McNicol et al., 2018) where time-series analyses (Verbesselt et al. 2010(Verbesselt et al. , 2012) may perform better (Bos et al., 2019); but also in moist forest in Tanzania (Hamunyela et al., 2020) and elsewhere (Bos et al., 2019;Milodowski et al., 2017).

F I G U R E 3
Comparison of tree cover losses according to Hansen et al. (2013;GFW) and modelled prediction of tree cutting by 2020. Note that the legends are not directly comparable.  Tables S4 and  S5 and Figure S2. The general patterns between modelled and actual tree (cover) losses appear similar. Circled areas in (a) contain reserves managed as plantations, where tree cover losses are larger than the model would suggest. Circled areas in (b) experienced less detectable tree cover losses than the model suggests but are highly degraded according to local reports This is not a critique of the data generated by GFW, but it serves as a reminder that in areas where smaller scale deforestation and degradation are a significant cause of carbon emission and biodiversity loss, such as southern and eastern Africa (Baccini et al., 2017;McNicol et al., 2018;Pearson et al., 2017;Sedano et al., 2020), it is necessary to go beyond easily accessible deforestation data and to use a combination of approaches to detect these changes.
While radar data correlated well with disturbance on the ground, they cannot detect activities that have little impact on vegetative biomass-such as low levels of harvesting, collection of non-timber products, hunting, or the introduction of invasive alien species (McNicol et al., 2018;Ryan et al., 2012). Using remotely-sensed data, it is also very challenging to distinguish types of disturbances, plantations versus natural forests, and primary vegetation versus the rapid secondary growth following logging (Asner et al., 2004).
Here we counted degradation as 'detected' even if only one pixel in or around a transect, that is, an area of up to ~20 ha, was classed as degraded or deforested. It is entirely possible that the removed tree(s) was not detected, and that the reported biomass loss was due to an unrelated co-incidental process or noise. Finally, given that almost all transects used in this study contained tree stumps, it was not possible to robustly establish the specificity (=false positive rate) of the radar dataset with our data. In summary, while radar data give increasingly accurate wall-to-wall quantifications of degradation, there is still an important role for field data in aiding their interpretation, and providing an 'even earlier' warning signal in terms of subtle changes that can be detected before there is any notable impact on canopy or biomass. Similarly, early warning signals can also be provided by ground-based sensing, for example, hemispherical photography and terrestrial LiDAR (Decuyper et al., 2018;Fournier & Hall, 2017).
Capturing the spatial patterns and types of degrading activities, particularly when they are illegal, requires surveying relatively large areas. Field-based inventories and monitoring are however frequently restricted to a small sub-sample of areas of interest (O'Connell, 2018). The framework presented here can be used for quick assessments of large areas without professional training, thereby also allowing for community participation (Danielsen et al., 2011;DeVries et al., 2016). Details can be adapted to the target system and question (but should of course be standardised to ensure comparability; for a recommended set of core measurements, please see the Supporting information: Field Protocol file). In particular, we would recommend using a higher size class resolution than used here and/or detailed dbh measurements. Our models for tree cutting performed less well than those for pole cutting and charcoal burning, which is likely due to tree harvesting >15 cm dbh serving a multitude of purposes, ranging from high-grade export timber to local construction and partly also charcoal production. Differentiating three to five size classes can still be done rapidly by eye, and even detailed dbh measurements are not too time-consuming. Particularly, if combined with the identification of main timber species, this would provide more information on likely markets and scale of operation. Such higher resolution data would also enable estimation of likely levels of sustainability of the resource extraction, whereby a decline in highvalue species and/or larger trees are often indicators of unsustainability (Ahrends et al., 2010). In addition, more details, particularly on stem sizes, would also improve estimates of above-ground carbon (loss), which could only be crudely estimated using the simple counts. Another useful potential addition is collaborative work with social scientists in order to capture local knowledge, and to understand whether the resource extraction leads to win-lose or lose-lose scenarios locally (Smith et al., 2019). The transects can be done as a stand-alone activity or in addition to more detailed assessments in long-term vegetation plots (The SEOSAW Partnership, 2020), opportunistic botanical sampling or other types of surveys. Rapid transects cannot replace the depth of assessment possible in permanent plots, and large plots are also necessary for the calibration of radar (McNicol et al., 2018) as using narrow transects to relate radar to biomass is very challenging (Réjou-Méchain et al., 2014;Smith, 2018).
A key benefit of field data is that they can provide information on the type of biomass loss (e.g. charcoal, poles, planks or agricultural clearing) and sometimes on the type and sophistication of equipment that was used, allowing insights into the likely drivers and tailoring interventions appropriately (Doggart et al., 2020). Here we showed that while pole cutting may partly be driven by local demand, activities such as tree cutting and charcoal production correlated almost entirely with distances to major cities such as Dar es Salaam.
Degradation thus appeared to be mainly driven by energy and timber demand emanating from larger cities and international markets, as opposed to mainly local demand (Ahrends et al., 2010)-a pattern that has been observed throughout southern Africa (McNicol et al., 2018;Sedano et al., 2020). Deforestation on the other hand is thought to be mainly driven by agriculture, highlighting the need for coordinated policy responses (Doggart et al., 2020;Hamunyela et al., 2020). It should also be noted that while the clear spatial patterns meant that degradation was to some extent predictable, dynamics in markets, human behaviour and policies can lead to rapid changes on the ground-such as the introduction of sesame farming in Tanzania (Brockington, 2019;Gross-Camp et al., 2019;Müller et al., 2014). Thus, although models can to some extent be used to extrapolate patterns in space and time, there is a clear need for regularly updated data (Sloan & Pelletier, 2012).
Protection on the ground has had some success in halting degradation but the type of management was less important in explaining patterns of forest condition than demand and accessibility. This suggests that any form of protection can be better than none, and putting land under the tenure and management of local communities might be a more mutually beneficial way to reserve some of the 170,000 km 2 of forest on general land in Tanzania (Mbwambo et al., 2012) than excluding rural populations from the resources their livelihoods rely upon. Tree cutting in village-owned reserves only slightly exceeded levels in protective forests and nature reserves, and this was to be expected as village land forest reserves often allow sustainable extraction. The effectiveness of village participation in forest management on government-owned land (co-management) could not be robustly assessed because much of the data were collected when joint forest management agreements were in very early stages (Mbwambo et al., 2012).
The early warning provided by both radar and field data compared to GFW is a key advantage, because severe degradation and deforestation often follow the early stages of degradation (FAO, 2011)-a sequence we also observed here. However, in terms of (temporal) data availability, a significant advantage of GFW is that the readily processed data are freely available on an annual basis with global coverage, explaining their widespread use. This is not yet true for radar-based maps; although raw data are now freely avail- mapped area size depends on levels of heterogeneity and desired accuracy). In practice, a reasonable compromise may be to produce at least annual radar-based maps of biomass change, combined with rapid field surveys at 3-5-year intervals to facilitate a better understanding of the nature and drivers of biomass loss.
Strictly speaking, the method presented here only quantifies woody resource extraction and not necessarily degradation. The latter is challenging to establish-particularly in systems where little is known about regeneration and growth rates. However, while systems adapted to frequent natural disturbance may be resilient to some resource extraction, the selective extraction of larger trees in old-growth forest can negatively impact ecosystem function and biodiversity (Jew et al., 2015;Tripathi et al., 2019;Yguel et al., 2019). In addition, while there is controversy over the role of wood products in carbon storage, the damage to the surrounding vegetation in denser forests, as well as the associated transportation and processing of the timber, tend to lead to substantial emissions (Ingerson, 2009;Pearson et al., 2014). Resource extraction in old-growth forests thus requires careful regulation. The vast majority of extraction recorded here took place in protective (as opposed to productive) reserves, and was consequently mostly unregulated and illegal with no concomitant legal revenue benefits for Tanzania as a state (Milledge et al., 2007).
In conclusion, the consideration of degradation in global forest reporting is important-particularly in southern Africa where the area affected by degradation is likely to be double the size of the area that is deforested, and overall carbon emissions from forest degradation are likely to exceed those from deforestation (McNicol et al., 2018). We recommend to routinely use radar-based monitoring combined with, wherever possible, rapid field assessments to better understand the quality of forests and the reasons for their decline, to provide an early warning, and to allow for informed and timely policy interventions.

ACK N OWLED G EM ENTS
We are very grateful for the contributions of extensive data from