The study was carried out in five sites (experiencing either FSC industrial or artisanal logging) located in three countries across the Congo Basin: the Republic of the Congo, the Democratic Republic of the Congo and Cameroon (see Figure 1). In all sites, two dry seasons (June - August and December - February) alternate with two wet seasons (March-May and September-November) during the year (Creese et al., 2019).

The two studies areas experiencing industrial logging are named Kabo (Kab, 2°34′N,17°08′E) and Mbeti (Mbe, 1°98′N,16°02′E) and are located in the northern part of the Republic of the Congo (see Figure 1) in a semi-deciduous tropical forest dominated by medium wood density tree species (Réjou-Méchain et al., 2021). They cover 2,953 km² and 5,650 km², respectively, and are exploited by a company named “Congolaise Industrielle des Bois” (CIB-Olam Concession), certified by the Forest Stewardship Council (FSC). The mean annual temperature at both sites is 25 °C and the mean annual precipitation is around 1,600 mm (Loubota Panzou et al., 2018). According to the overall inventory of forest stands in forest management units encompassing the two study areas, the average tree density was about 350 trees per ha, while basal area and above ground biomass were of 28 m²/ha and 433 Mg/ha, respectively (Forni et al., 2019; Gourlet-Fleury et al., 2023).

The three study areas subjected to artisanal logging are named Yanonge (Yan, 0°49′N,24°7′E), in the Democratic Republic of the Congo, Mindourou (Min, 3°50N, 13°44E) and Dzeng (Dzg, 3°7′N, 11°82′E) in Cameroon and cover 155, 106 and 357 km², respectively (see Figure 1). The Yan and Min sites are located in a semi-deciduous evergreen transition forest (Réjou-Méchain et al., 2021). The Dzg site is located in a degraded semi-deciduous forest (Réjou-Méchain et al., 2021). The mean annual temperature is around 25, 23 and 23 °C in Yan, Min and Dzg respectively and the mean annual precipitation is around 1,800, 1,500 and 1700 mm respectively (climate-data.org).

Over the site of Mbe (Republic of the Congo), we had access to a detailed 3D representation of canopy disturbances thanks to repeated UAV surveys. Two main data sources were used and combined: stereophotogrammetry and LiDAR. RGB imagery data were acquired on 2018-06-18 and 2019-04-05 using a Ebee Sensefly fixed wing UAV. Stereophotogrammetric processing was performed using Pix4D software. Digital Surface Models were produced at a 0.1 m resolution. LiDAR acquisition was performed on 2020-02-13 using a Surveyor Ultra (Yellowscan) LiDAR sensor. The system comprises a VLP32 scanner (Velodyne) and an Applanix 15 IMU. The scanner was carried onboard a DJI Matrice 600 UAV with scanning direction perpendicular to flight direction, at a flight altitude of ∼ 70  m above ground level (following SRTM DTM) and a flight speed of 15 m s − 1 . An interline distance of 80 m was chosen to achieve ∼ 50  % overlap at less than 45° incidence angle at actual ground level (which is below the SRTM altitude in a dense canopy). Differential correction of flight trajectories was achieved using an Emlid Reach RS2 GNSS base station on a point of location determined using a Precise Point Positioning approach using a full day of acquisition data. Post Processing Kinematic adjustment and optimisation of the trajectories was performed using the Applanix PosPac UAV software (v. 8.5). Las point clouds were exported using CloudStation software (Yellowscan) on the basis of the optimized trajectories and sensor calibrated geometry. A digital surface model (DSM) was derived at 0.6 m horizontal resolution using the Lastools software function lasgrid, with the -max option. The description of the LiDAR acquisitions are detailed in Kaçamak et al. (2022).

The total overlap area between the three surveyed dates was ∼ 374 ha . To ensure perfect spatial adjustment, a global and local adjustment was performed using the Arosics software (Scheffler et al., 2017) using the LiDAR coverage as reference. Two binary maps of small-sized forest disturbance/non disturbance were derived from the three resulting DSM. Disturbed pixels were defined as a decrease in tree canopy equal or higher than 10 m between 2019-2018 and 2019-2020. To compute the 2019-2020 disturbances, the DSM of 2019 was resampled to the spatial resolution of the 2020 DSM (0.6 m). Some pixels were classified as disturbed because of very slight spatial shifts between the DSMs. Therefore, a morphological opening operation was applied using the Binary Morphological Operation algorithm from OTB (Grizonnet et al., 2017). An erosion operation was used to remove disturbance patches smaller than 2 m, followed by a dilation of 2 m to restore the size of the remaining disturbances. Despite the differences in sensitivity to forest cover between LiDAR and stereophotogrammetry data (Goodbody et al., 2019), the use of the DSM instead of points clouds and the removal of very small detections avoid biased in UAV-derived binary forest disturbance maps. The UAV-derived binary forest disturbance maps were finally resampled to 1 m resolution to limit time-computing.

FLCD and CuSum were trained on the UAV-covered footprint and validated on the remaining Mbe area and on the five other study sites using independent GPS-based datasets. The UAV-derived binary disturbance map is available only for the Mbe site, which limits its representativeness across the full range of forest types considered in this study. However, Mbe and Kab share comparable forest types and climatic conditions, so this limitation is expected to have a limited impact on method performance in FSC-certified sites, while potentially affecting performance in the three artisanal logging sites (see Section 2.1).

Several sources of uncertainty are inherent to UAV-based canopy disturbance mapping. Time intervals between acquisitions (2018–2019 and 2019–2020) do not allow precise dating of individual tree-felling events. DSMs derived from RGB stereophotogrammetry can also be affected by illumination conditions, cloud shadows and variations in light diffusion within the canopy, introducing local height estimation errors. These effects were mitigated by focusing on large canopy height losses ( ≥ 10 m) and the UAV acquisitions were all performed under clear weather conditions.

This dataset includes coordinates (EPSG: 4,326) of felled trees and associated felling date recorded in FSC (Mbe and Kab) and artisanal logging areas (Yan, Dzg, and Min).

Contrary to the previous dataset, spatial extents of disturbances were not available. Only the location of the felled trees stumps were provided and analyzed.

Over the five study areas, S-1 C-band (5.405 GHz) time series were acquired between 2018 and 2022 (Mbe: 2018-2019 and 2021, Kab: 2021-2022, Yan: 2021, Dzg: 2021 and Min: 2021). The images were downloaded in Interferometric WideSwath (IW) mode, corresponding to a resolution of 20 × 22 m (rg × az), a pixel spacing of 10 × 10 m (rg × az) and a swath of 250 km. The revisit frequency of the S-1 constellation is 12 days in Central Africa, increasing to 24 days from December 2021 onward due to the loss of data transmission from the S-1B satellite. This may reduce sensitivity to very short-term disturbances and affect the temporal accuracy of detection. However, logging gaps persist for several months to years before canopy closure, as shown by previous studies in Amazonia (Asner et al., 2004), which limits the impact of this longer revisit interval on disturbance detection. The Level-1 Ground Range Detected (GRD) products were used, which are multilook intensity (5 and 1 looks according to slant range and azimuthal direction, respectively) projected to a ground range based on an Earth ellipsoid model. Since the S-1 satellite can only map our study areas in a single pass, images were acquired in descending mode for the study areas located in the Republic of the Congo and the Democratic Republic of the Congo (Mbe, Kab, Yan) and in ascending mode for the study areas located in Cameroon (Dzg, Min). All available images were downloaded from Google Earth Engine (GEE) “COPERNICUS/S1_GRD” collection using the open python code (https://github.com/chiaquino/S1_CuSum) provided by Aquino et al. (2022). For each observed year, the time series started 12 months before the year and ended 6 months after the observed year, e.g., for the year 2019, the time series extended from 2018-01-01 to 2020-06-01. Disturbance detection requires a post-disturbance series of 6 months for optimal results (Supplementary Appendix S5.1). Pre-processing already done on GEE “COPERNICUS/S1_GRD” collection includes thermal noise removal, radiometric calibration and terrain correction. Speckle filtering was not applied to either FLCD or CuSum. Aquino et al. (2022) showed that speckle filtering degraded the detection accuracy of CuSum. Similarly, internal analyses conducted for FLCD using Boxcar, Refined Lee, Lee, and Lee Sigma filters led to a slight increase in recall, but substantially reduced precision and increased false detections. Spatial filtering may alter the pixel backscatter signal, which is critical for detection of very small canopy gaps, while the lasso regression inherently accounts for noise. The python code from Aquino et al. (2022) further performs sigma nought to gamma nought ( γ 0 ) conversion and data mosaicking for each unique date using the GEE mosaic function. The γ 0 corresponds to sigma nought corrected for near-far range incidence angle variations. It was selected since Hoekman et al. (2020) demonstrated that γ 0 was more relevant than sigma nought to monitor structural changes in tropical forests. Given the relatively flat terrain of the Congo basin, angle-based correction is sufficient to account for backscattering variations without needing a more complex area-based correction (Small, 2011). Our study focused on the VV polarization since Ygorra et al. (2021) found that the VV polarization was more accurate than VH to monitor tropical forest changes using S-1 data. A number of 72–84 γ 0 VV S-1 images was used, depending on the site. Images were projected in WGS84 coordinate system (EPSG: 4,326).

The FLCD method aims at producing a binary map featuring small-sized forest disturbances (associated with magnitude values), while putting emphasis of disturbance date assessment. Figure 2 illustrates the workflow implemented in this study. The Table 1 indicates the use of each dataset in the workflow.

To produce maps of small-sized forest disturbances from the S-1 SAR time series, we developed a method we call Fused-Lasso Change Detection (FLCD). Figure 3 describes the main steps (detailed below) of the FLCD method. First, fused-lasso regressions are computed on pixelwise S-1 time series. Secondly, iterative local differences between successive lasso-smoothed values and corresponding sliding sums are applied to highlight signal decreases. Finally, a post-classification step (spatio-temporal filtering) based on the disturbance dates is applied to reduce false detections.

The FLCD parameters (sliding period size for the sliding sum, detection threshold and exclusion patch size) were calibrated using the UAV-derived binary forest disturbance map (over 374 ha) for the 2018-2019 period in Mbe, the Republic of the Congo (see Supplementary Appendix S5.2). A 5-fold cross-validation approach was used to select the two FLCD parameters, namely, the detection threshold and the sliding sum window. Five training and testing datasets were defined within the UAV study area. FLCD were applied to each training and corresponding test dataset. The detection threshold was tested at values of 0.005, 0.01, and 0.015; the sliding window length at 30, 60, 90, and 120 days. For each parameter configuration, the mean performance and confidence intervals were computed across the five train–test splits (Supplementary Appendix S5.2).

The reference polygons were rasterized on the basis of the detected disturbance maps grid (10 m resolution). Then, we assessed the FLCD classifications computing precision (Equation 1) and recall (Equation 2), which are suitable when focusing on small proportions of positive class, using the following equations. precision = TP TP + FP (1) recall = TP TP + FN (2) where TP = true positive, FP = false positive and FN = false negative.

The parameters yielding the best compromise between a high precision and a high recall were selected.

The trained FLCD method was validated on extensive FSC and artisanal areas using the felled-trees locations dataset. GPS points have two main limitations: they only indicate felled trees without providing data on non-disturbed pixels, and the spatial discrepancies between GPS points taken near stumps and actual canopy disturbances can cause mismatches. Thus, we calculated indicators based on overlapping between reference polygons and disturbed patches instead of a conventional pixel-wise approach, as done by Aquino et al. (2022).

FLCD was applied on Mbe 2018-2019, Mbe 2021, Kab 2021, Kab 2020, Yan 2021, Dzg 2021 and Min 2021, resulting in seven binary disturbance maps. Detected disturbed pixels around roads, identified through aerial photographs, were reclassified as undisturbed using a buffer of 30 m. All GPS points that overlapped with the train dataset (i.e., UAV-derived binary forest disturbance map in Mbe 2018) were removed. For each forest disturbance map produced, the dates of detected disturbances were filtered to match with the year of GPS data. Reference polygons were computed by applying a buffer area of 20 m around the felled trees points. The resulting polygons were used to estimate the Logging gap Detection Rate (LDR) and Unconfirmed Detected Logging gap Rate (UDLR) defined as: LDR = 100 n rp n r and UDLR = 100 n p − n rp n p (3)

Where n r is the total number of reference polygons, n p the number of detected patches and n r p the number of detected patches overlapping reference polygons. In this patch based framework, n r p is equivalent to true positives (TP) and n p - n r p to false positives (FP). The LDR metric is therefore related to TP, whereas the UDLR metric reflects FP. False negatives (FN) and true negatives (TN) have no equivalent metric, as the field measurements do not include information on undisturbed areas. The LDR index answers the question: “What is the rate of referenced logging gaps that are correctly detected?” whereas the UDLR index answers: “What is the rate of detected patches that do not correspond to logging gap events?“. This process was implemented using the “vectorize”, “patches”, “extract” and “intersect” functions from the “terra” R package (Hijmans et al., 2022).

Finally, the discrepancy between detected and reference disturbance dates were analyzed. First, GPS reference polygons were rasterized on the basis of the detected disturbance maps grid (10 m resolution). The difference between reference and detected dates were computed on the overlapping disturbance pixels. The histogram, median, mean, first and third quartiles were calculated from the differences between dates for each study area.

In this study, we developed a new method called FLCD. We compared its performances to that of the CuSum method, which is accessible and of proven effectiveness in detecting small-sized forest disturbances under 0.1 ha (i.e., 1,000 m² or 20 × 50 m). Since the RADD method was previously compared with CuSum (Aquino et al., 2022), and its minimum mapping unit is 0.1 ha, we do not compare it quantitatively with FLCD. Dupuis et al. (2023) also introduced an effective method for detecting small forest disturbances monthly, but do not aim to accurately predict felling dates. Finally, the study conducted by Carstairs et al. (2022) exhibits high potential for measuring the extent of forest disturbances at 1 ha scale. However, they state that “the presence of geolocation errors and false alarms makes this method unsuitable for confirming individual disturbances”. Although they found a linear relationship between the detected S-1 shadow and LiDAR-based canopy loss at a scale of 1 ha, they do not claim that their method could accurately identify individual disturbances or determine accurate logging dates.

Forest disturbance maps derived from the CuSum approach were computed using the open python code (https://github.com/chiaquino/S1_CuSum) provided by Aquino et al. (2022). To allow comparisons, both methods were applied to the same S-1 time series following data preprocessing described in Subsection 2.4. The CuSum method was also trained using the 2018-2019 UAV-derived binary forest disturbance map in the Republic of the Congo. A k-fold cross validation was used with the same five training and test datasets as FLCD varying the percentile at 95, 96, 97, 98, and 99. On this basis, the detection threshold was set to the 99th percentile of the CuSum maximum distribution (see the Supplementary Appendix S5.3 for details). The trained CuSum method was also validated on the extensive FSC and artisanal areas using the GPS point dataset and the 2019-2020 UAV map. Differences in LDR between FLCD and CuSum were assessed using a paired Wilcoxon signed-rank test from the LDR values computed on the height study sites. UDLR was not tested because it could only be computed for FSC-certified sites, resulting in an insufficient sample size for robust testing.

Both FLCD and CuSum share common features, as they rely on S-1 VV-polarized gamma nought SAR time series to detect small-scale canopy disturbances. Methodologically, both approaches identify significant decreases in the radar signal using percentile based detection thresholds and provide estimates of disturbance date and magnitude. FLCD explicitly models the temporal evolution of the SAR signal using fused-lasso regression, whereas CuSum is based on cumulative sums of residuals derived from the SAR signal. Disturbance dates in FLCD are derived from a sliding sum of negative local differences in the fused lasso regression. In contrast, CuSum relies on the maximum cumulative sum relative to the mean temporal signal. Finally, compared to CuSum, FLCD includes an additional post-classification spatiotemporal filtering step to remove isolated false detections.

Based on the 5-fold cross-validation approach, the sliding period size for the sliding sum was set to 3 months. The detection threshold was set as the 0.01th percentile of the distribution of all negative sliding sum values across the entire image. In other words, disturbances were identified when the summed negative values of the time series for a given pixel were more extreme than 99.99% of the corresponding values across all pixels, indicating significant forest structural change. The exclusion patch size during the post-classification process was set to 0.01 ha (i.e., one 10 m pixel corresponding to the pixel spacing of S-1). Therefore, the current trained FLCD method cannot reliably detect disturbances smaller than two adjacent 10 m pixels (i.e., 0.02 ha or 200 m²). These disturbances may be detectable if, despite their small area, they significantly affect the S-1 signal in both pixels. This is because while the individual pixel spacing of S-1 is 10 m, its actual resolution is coarser ( 20 × 22 m).

Figure 4 shows the maps produced by the FLCD method with the calibrated parameters over the 2018-2019 UAV-derived binary forest disturbance map. Throughout the monitoring period (June 2018 - April 2019), the trained method reached a precision of 55% and a recall of 23.78% The majority of disturbance detections occurred in December 2018 and January 2019 (see Figure 4 Left panel) and corresponded to a 4–10 dB decrease in the VV signal (i.e., disturbance magnitude, see Figure 4 Right panel).

For illustration purposes, Figure 5 shows the temporal behaviors of γ 0 VV, fused-lasso regressions and sliding sum values on three classes of pixels with different profiles. The first class corresponds to actually disturbed pixels detected by the FLCD method (true positive rate), the second to undisturbed pixels misclassified as disturbed (false positive rate) and the last one corresponds to undetected real disturbances (false negative rate). When the FLCD method succeeds in detecting the disturbances (true positive rate), the SAR signal is relatively steady before and after the disturbance, which itself materializes as a decrease in the signal of varying time length and magnitude. This corresponds to horizontal fused-lasso regression lines interrupted by neatly decreasing segments. The sliding sum highlights the successive negative decreases. In the case of false positive rates, the FLCD method misclassified pixels as disturbed due to either a weak decrease in the SAR signal nevertheless exceeding the percentile threshold. Finally, the FLCD method failed to detect some disturbances (false negatives) which did not alter the signal stationarity leading to fused-lasso regressions remaining constant all over the series.

The method was validated on the independent datasets made of felled trees locations in the FSC study areas. This led to a LDR of 31.08% (see Table 2). The method’s effectiveness at detecting logging gaps varied across individual sites, with the LDRs ranging between 24.32% and 37.78% on FSC study areas (Table 2) and from 28.57% to 59.09% on the artisanal study areas (see Supplementary Appendix S5.6). The UDLR varies from 31.23% to 66.89% accross the FSC study areas (Table 2).

The FLCD method was able to estimate accurately and precisely the logging dates on the FSC study areas with the median of differences between real and detected dates ranging from −6 to −2 days depending on the study area (Figure 6). In artisanal study areas, the median was −3, −6 and −20 days on Yan, Dzg and Min, respectively.

From FSC-certified areas, we found that the FLCD method systematically yielded more accurate and precise logging dates than the CuSum approach (see Figure 6). Indeed, distributions of temporal deviations between detected and reference dates across sites were larger with the CuSum than the FLCD method (Figure 6, see the Supplementary Appendix S5.6 for descriptive statistics values per study area). Concerning the smaller set of artisanal logging dates (see the Supplementary Appendix S5.6), results were contrasted across sites and fairly homogenous between methods with low bias and high precision on the Yan site (median ± sd: 0 ± 42.66 for CuSum and −3 ± 59.02 for FLCD), a moderate bias and low precision on the Dzg site (median ± sd: -29 ± 90.80 for CuSum and −6 ± 102.84 for FLCD) and a high bias on the Min site (median ± sd: -203 ± 89.65 for CuSum and −178 ± 75.63 for FLCD). Overall, LDR values of CuSum were slightly higher than those of FLCD across the five FSC-certified logging areas (i.e., three in Mbe and two in Kab, see Table 2). A similar pattern was observed on artisanal study areas (Dzg, Min, Yan) with CuSum systematically detecting 4 more logging gaps than FLCD (see the Supplementary Appendix S5.4). The paired Wilcoxon test indicated a statistically significant difference in LDR between FLCD and CuSum (V = 2, p = 0.023). On average, FLCD yielded lower LDR values than CuSum, with a median paired difference of −8.45%, indicating that CuSum detected a higher proportion of reference logging gaps. However, the confidence interval (−14.29% to −1.18%) also highlights the variability in relative performance across study sites. In FSC study areas, the number of unconfirmed logging gaps was overall high and higher for the Cusum than the FLCD method in three sites above five (see Table 2) ranging from 419 to 3,998 unconfirmed gaps for CuSum against 579 to 2,301 over-detected patches for FLCD. During the first UAV-covered period (2018–2019), which coincided with logging activities (n = 341 reference patches), UDLR were 10.61% and 9.52% for CuSum and FLCD, respectively. In the second period (2019–2020), when no logging occurred, the UAV map showed three disturbance patches likely caused by natural tree falls. During this period, CuSum and FLCD produced 17 and 61 false positives, respectively, with no overlap with the three reference patches.

From the UAV-covered reference area it was possible to analyze the detection rates in relation to the gap sizes (bearing in mind that this area was used for calibrating the two methods). Both methods agrees in displaying a sharp increase of true detections with gap size assessed from diachronic LiDAR (see Table 3). The detection rate was c. 49% for gap sizes of 0.02 ha and went above this value for larger gaps. It reached 71% and 90% for patches equal or higher than 0.05 and 0.20 ha, respectively The overlaps between detected and assessed patches also increased, albeit to a lesser extent, with FLCD yielding slightly higher overlap values. The results also highlighted that the gap size frequency distribution is overall skewed towards very small gaps with only 35% of LiDAR-assessed gaps equal or over 0.02 ha, and 9% equal or above 0.05 ha.

The FLCD method appeared relevant to monitor small-sized forest disturbances such as canopy gaps made by tree felling in both FSC industrial and artisanal logging contexts. The dates of the detected forest disturbances were estimated with a very high accuracy on FSC sites (i.e., −8.37 to +4.60 days of difference on average between the estimated and the actual disturbance date, see the Supplementary Appendix S5.6. Part of the uncertainty in the detected date is due to the fact that the actual tree felling date can occur at any time between the date of the image in which the change is first visible and the date of the previous image in the time series. This is a problem already encountered and reported in previous studies (Reiche et al., 2018a). Moreover, some pixels can still be disturbed after the felling event by delayed mortality of remaining vegetation (Dalagnol et al., 2019). Considering those two sources of uncertainty, the assessment of disturbance dates appeared of high accuracy.

Although our method targets canopy gaps as small as 0.02 ha (i.e., just two nominal S-1 pixels), our method detected an average of 31.1% of logging gaps in FSC sites (see Table 2) and 42.6% in artisanal logging areas (see the Supplementary Appendix S5.4). Moreover, detection results in the UAV-covered reference area clearly showed an increasing relationship of the detection rate with gap size, and they suggested that detection rates could be above 71% for canopy gaps in the range 0.05–0.1 ha (Table 3). These encouraging results remain however to be considered with caution and call for future verification on new UAV-based datasets of larger extent. Indeed, we assessed the parameters of the tested methods over our unique UAV-covered area of 374 ha, while our very large datasets from FSC concessions provide no information on the gap sizes associated to the felled tree locations. It is worth noting, however, that the rate of true positive we found in the UAV-covered area is very close to the rate observed in the FSC validation sites (32.9% vs. 31.1%), which suggests no blatant overfitting, as supported by the cross validation strategy implemented during calibration.

Considering false detections, the FSC datasets do not provide information on canopy disturbances apart from felled trees, whether natural tree fall gaps (i.e., chablis) or anthropogenic disturbances such as logging roads, log landing sites and even small-holder crops. As a result, all associated pixels are likely to be detected but incorrectly labeled as false positives, thereby inflating the UDLR figures. Indeed, UDLR values we observed in the FSC sites are high, i.e., ranging between 31.2% and 66.9% which is substantially greater than the 23% observed in the UAV-covered reference area, where we were able to accurately exclude canopy disturbances unrelated to tree felling. This supports the idea that a large fraction of the false positives registered in the FSC sites may well be actual gaps, though unrelated to logging or at least to tree felling operations. Additional UAV-borne data over large areas would be needed to accurately assess the fraction of actual gaps unrelated to logging, considering that this fraction seems to display notable inter-site and inter-annual variability. To get an idea, we may refer to studies on natural tree fall and deriving canopy gap creation. For instance, directly interpreting (Van Der Meer and Bongers, 1996) results suggests that the yearly proportion of standing tree that fall and actually create a canopy gap is about 1%, which would correspond to ca. 5 gaps/ha*year, compared to 0.5 to 4/ha tree felled in FSC concessions of Central Africa.

The observed UDLR values should thus be regarded as upper-limit estimates of false positives, which could be reduced using ancillary data to distinguish and sort out non-logging canopy gaps. Further research, though beyond the scope of the present paper could classify detected canopy disturbances by driver types in order to yield more accurate reference data to assess false positive values truly relating to small-area logging-induced disturbances (Dalagnol et al., 2023; Slagter et al., 2023).

FLCD was more accurate than CuSum in estimating the dates of tree fellings. The discrepancy between the actual felling date and the detected disturbance date ranged from −8.34 to +4.60 days on average using FLCD and from −38.14 to +0.32 days on average using CuSum in the FSC sites (Mbe, Kab). Those last figures suggest an intrinsic bias of CuSum towards negative error values (i.e., estimated date posterior to the actual one), which seems also corroborated by site-wise figures. Overall, the detection rate obtained with CuSum was slightly higher than with FLCD: in some FSC-certified sites (Mbe), the LDR values showed modest differences but still significant based on the Wilcoxon test, while in others, somewhat more noticeable variations could be observed (Kab). In the UAV-covered area, both methods displayed very similar trends of increasing detection rate and precision with canopy gap sizes. Both yielded rates close to 71% for gaps of 0.05 ha. This suggests that observed detection rates might fundamentally reflect an intrinsic limit of the actual spatial resolution of the S-1 data that hinders detection of gaps of very small size (0.02–0.05 ha, i.e., just 2 to 5 pixels) whatever the method applied. Although S-1 data are provided at a ground spacing of 10 m, the effective spatial resolution after multilooking is coarser (approximately 20 × 22 m), which explains the reduced detection rates observed for disturbances close to the pixel size. This casts doubt on the feasibility of using S-1 data to track all individual felling events, since a large share of logging gaps are indeed very small and poorly detectable. Moreover, some gaps are retaining understorey vegetation cover or specific species such as palms in humid situations which could reduce the drop in S-1 signal. Having a large fraction of small logging gaps is in fact a positive result of FSC certification through improved felling techniques aimed at minimizing disturbances. In spite of these difficulties, both CuSum and FLCD can contribute effectively to monitoring logging activities and logging-induced degradation fronts at landscape scales. The UDLR rate quantifying detections that do not match referenced felling locations only showed weak differences between CuSum and FLCD bearing in mind the notable variability across FSC sites of results from both methods and considering the limited number of sites mentioning UDLR, to carry out a Wilcoxon test. The same explanations evoked for FLCD seem relevant to account for the large UDLR values also found for CuSum. The CuSum method is a threshold-based change detection method, while the FLCD is a trajectory segmentation method (Hirschmugl et al., 2017), which flexibility allows more improvement opportunities, considering the increasing availability of UAV-based data that will progressively help enhance training and parameter calibration. Indeed, regarding the flexibility of FLCD, longer time series could help improve the fused-lasso fitting (Tibshirani et al., 2005). Furthermore, in time series change detection methods, seasonal effects are often preliminarily suppressed from the remote sensing time series (Hirschmugl et al., 2017). Removing the seasonal effect from the S-1 signal can be done for both CuSum and FLCD methods as (Reiche et al., 2018b) did for RADD. However, the fused-lasso allows going further with the incorporation of external factors (Arnold et al., 2022) such as topography, season, month and dry-wet periods thereby helping remove a part of the signal noise as to better focus on forest changes. This opportunity to add external factors is further developed in the next Section 4.3 Prospects for improvement. In addition, our application included a post-classification step that was substantially reducing the number of commission errors (see the Supplementary Appendix S5.2) and could prove useful for other methods.

Further studies should focus on improving the calibration of the FLCD method and on understanding how S-1-SAR characteristics contribute to the different types of detection errors. Firstly, FLCD (and also CuSum) is based on a percentile threshold (0.01 for FLCD and 99% for CuSum) calculated from all the pixels of the study area. Further research is needed to understand how integrating the properties of the land cover mosaic and the size of the study area to appropriately set this threshold. For instance, future work could investigate the sensitivity of the percentile-based detection threshold across contrasting landscape contexts, with varying forest-to-non-forest ratios. By optimizing threshold calculation such as choosing a minimum number of pixels in the study area and masking non-forests land uses, we may expect enhancing the accuracy of the felling gap detection. Secondly, based on a trajectory fitting approach (Hirschmugl et al., 2017), reference data (notably, multi-temporal UAV-borne) could be used to classify temporal S-1 signals that are characteristic of forest disturbances and sort out inappropriate signal trajectories that does not lend themselves to detecting punctual events (e.g., steadily decreasing ones as illustrated in Figure 5). These typical signal trajectories could be classified from fused-lasso regressions prior to going further into the threshold detection process. Indeed, the SAR signal variations induced by the speckle noise and local moisture fluctuations (Bouvet et al., 2018; Reiche et al., 2018b; Hirschmugl et al., 2020) increase the number of false positive detections. This is often handled through spatial smoothing. But Aquino et al. (2022) demonstrated that smoothing the time series using boxcar filters with a window of one to seven pixels systematically led to a decrease in the detection accuracy of the CuSum method. A major challenge in the tropics is the strong negative correlation between rain intensity and negative anomalies in the C-band SAR signal (Doblas et al., 2020), leading to false detections of forest disturbances. Finally, the incorporation of VH polarization has been suggested by Aquino et al. (2022) to improve the accuracy of the CuSum method as could also be done for FLCD. Although VH polarization is generally less sensitive than VV to abrupt canopy openings in dense tropical forests (Ygorra et al., 2021), it captures complementary scattering mechanisms related to volume scattering within the forest canopy. Previous studies have combined VV and VH information, or used derived indices such as VV/VH ratios, to help improve detection performance (Mermoz et al., 2021; Ygorra et al., 2021; Carstairs et al., 2022; Dupuis, 2023). Other SAR features, such as the radar vegetation index, polarimetric (Dupuis et al., 2020) and textural features (Balling et al., 2023; Hethcoat et al., 2021) have shown promising results in monitoring forest disturbances. The fused-lasso framework could be extended to incorporate SAR and external variables such as elevation, soil type, or location through a multi-type regularization approach. Devriendt et al. (2021) have presented a regression setting where different predictor types are assigned to penalty structures tailored to their specific properties. By modeling external variables to change gradually over time while allowing sudden changes in the forest signal itself, the FLCD method could better separate natural environmental effects from real logging events. For example, a drop in S-1 signal at low elevation may be likely due to human activity, while slow decrease at high elevation may reflect climatic stress. False positives could be reduced by incorporating contextual information into the fused-lasso framework by combining environmental covariates and spatial context. For example, the model could account for proximity to roads, clustering of detected disturbances, or elevation, helping to distinguish anthropogenic logging gaps from natural canopy openings. In addition, the launch of S-1C reduces the revisit interval compared to the 24-day period imposed by the loss of S-1B. This shorter revisit frequency can further improve the detection of small and short-term forest disturbances, because the growth of giant herbs and vines can be swift around the felled stumps. Moreover, even though the radar sensors are the most relevant in the humid tropics thereby reducing the gap size, complementary information can be expected from optical sensors, which for instance could be used to assess radar-detected gaps, as to reduce false positives. Indeed, while cloudiness hinders analyzing data series from optical sensors by limiting temporal regularity and frequency of usable images, a thorough analysis of such images, even if sparse, can help identify the nature and cause of the false positive. Those can be natural tree fall (rather small and isolated gaps), anthropogenic openings unrelated to tree felling (rather large and close to infrastructures), or genuine errors due to S-1 signal defects.

Tree felling, especially when unauthorized or unregistered tends to be at the forefront of the forest degradation process, which itself often pave the way to deforestation (Vancutsem et al., 2020). It is thus of importance to detect and delineate areas where new logging activities are initiated. Very often, such areas are not recognized and mapped as degraded by the current global monitoring systems (i.e., GLAD, TMF, RADD). We further note that for detecting felling gaps, both FLCD and CuSum have a clear edge against existing global monitoring methods such as GLAD (Hansen et al., 2016)/TMF (Vancutsem et al., 2020) and RADD alerts (Reiche et al., 2021), which do not pretend to address disturbances smaller than 0.09 and 0.1 ha, respectively. According to our field sites in Cameroon and the Democratic Republic of the Congo, 86% of the documented artisanal logging gaps were located in areas classified as “intact forest” in the TMF classification, while the median size of those documented gaps was ranging from 0.04 to 0.06 ha in the three sites. In a certified FSC concession in Aquino et al. (2022) found that only 2% of the logging-induced disturbances were detected by RADD against 59% by CuSum. In our FSC certified concessions, and as expected, RADD was able to detect only the larger forest disturbances in the UAV reference area (i.e., log landing sites and roads, see the Supplementary Appendix S5.7). In industrial logging concessions most of which are not certified, it is also needed to detect felling operations that may not respect the boundaries of annual coupes or the seasonal schedules that aim at soil protection. The method we propose here for detecting individual felling gaps is likely to highlight a surge of detections in any area experiencing new felling activities, even though a large share of felling events corresponds to individual gaps that are too small to be detectable from S-1 data. Such surges of spatially aggregated detections are clearly apparent in the dated detection maps presented in Supplementary Appendix S5.5. Industrial logging operation are generally concentrated in time, and on a monthly or seasonally basis, the resulting spatial patterns of detected logging disturbances are unequivocally different from what could be expected from natural tree falls. Getting sufficiently accurate dates of canopy disturbance is here quintessential to monitor the displacement of logging operations by delineating areas in which logging activities are in progress and to verify whether they are in compliance with existing guidelines and regulations. Concerning the monitoring of informal or artisanal logging, we provide here the first hitherto results on the detectability of such activities through S-1 data. Although our dataset is modest compared to the massive sets provided by certified corporations, it enabled us to assess the proportion of detected trees among a set of recent felling sites that we thoroughly ascertained in the field in three distinct regions of Central Africa. Our study which is a first foray into the important question of measuring impacts of informal logging suggests that the detectability of canopy gaps made by artisanal logger may be at very similar levels than for logging corporations. Moreover, results are promising on two artisanal sites (Yan and Dzg) with histograms of date differences centered around zero as in the FSC sites (see the Supplementary Appendix S5.6). More comprehensive datasets, for instance based on UAV-borne diachronic LiDAR acquisitions are obviously needed to provide results on false detections. Regardless of the different logging types, the foreseeable increasing availability of such drone-borne datasets will provide a robust basis for consolidating the evaluation of the S-1-based gap detection techniques we attempted in the present study. Of course collecting UAV-borne data remains challenging in tropical areas of poor accessibility. However, difficulties are eased in industrial logging concessions where logging roads and camps allow accessing electricity sources for devices. Equipment costs for integrated UAV-Lidar systems are dwindling. Currently, once on the spot, a trained operator is able to cover about 1,000 ha in about 4 days of full work. For sampling a large logging concession, several sample patches of 1,000 ha each could be a good size. In addition to better evaluating false negatives and false positives, several avenues of investigation are possible as to draw more information from S-1 images. First, the spatiotemporal distribution of natural tree fall gaps should be compared with patterns of gaps generated by the different types of logging activities. Second, the spatial distribution of detected felling gaps is to be analyzed in relation to villages, road network and deforested areas. Thirdly, the return time of disturbances is a key parameter for evaluating forest degradation (Vásquez-Grandón et al., 2018) and disturbance detection could be repeated throughout the time series to detect recurring disturbances on the same pixel (or in neighboring pixels) and assess the closure time of the canopy by detecting when the signal returns close to its initial pre-disturbance value, as has been done in previous studies investigating forest recovery after wildfires and clear-cuts using optical satellite imagery (White et al., 2017; Chirici et al., 2020).

Regarding this third and last point, fused lasso regression constitutes a relevant framework for modeling and analyzing the temporal trajectories of the radar signal at the pixel scale. It will make it possible to take advantage of both the increasing availability of UAV-borne data and the lengthening of the S-1 image series to distinguish in the signal environmental and instrumental effects from those reflecting actual canopy changes.