Forest disturbance alerts for the Congo Basin using Sentinel-1

Johannes Reiche; Adugna Mullissa; Bart Slagter; Yaqing Gou; Nandin-Erdene Tsendbazar; Christelle Odongo-Braun; Andreas Vollrath; Mikaela J Weisse; Fred Stolle; Amy Pickens; Gennadii Donchyts; Nicholas Clinton; Noel Gorelick; Martin Herold

doi:10.1088/1748-9326/abd0a8

1. Introduction

The Congo Basin rainforest is the second-largest in the world, covering almost 200 million ha of humid tropical forest. It plays a crucial role in the global climate cycle and provides local livelihoods and resources for more than 100 million people across six countries: Cameroon, Central African Republic, the Democratic Republic of the Congo, Equatorial Guinea, Gabon and the Republic of the Congo (Somorin et al 2012). Despite its importance, forest dynamics in the Congo Basin remain understudied compared to those in Amazon and Southeast Asian rainforests, and it is only recently that scientists have begun to systematically evaluate how and why Congo Basin forests are changing (Tyukavina et al 2018, Creese et al 2019, Kleinschroth et al 2019).

The rapidly growing population in the Congo Basin has led to an acceleration of forest disturbance rates over the past decades, with an estimated 16 million ha of forest cleared between 2000 and 2014 (Tyukavina et al 2018). Smallholder agriculture causes the large majority of forest disturbances. Other major drivers include selective logging, mining, and road expansion (Potapov et al 2012, Tyukavina et al 2018, Umunay et al 2019). Artisanal and industrial logging account for 10% of the total forest disturbance across the Congo Basin, but up to 60% in individual countries, such as Gabon (Tyukavina et al 2018). A large part of the forest disturbances in the Congo Basin is considered unsustainable and/or illegal (Lescuyer et al 2014, Kleinschroth et al 2019, Umunay et al 2019).

In the past 10 years, satellite-based alert systems (Diniz et al 2015, Hansen et al 2016, Watanabe et al 2018) have emerged as the primary tool to provide near real-time information on newly disturbed tropical forest areas. A wide range of stakeholders, including governments, NGOs, private sector actors and communities across the tropics have recognized the value of satellite-based disturbance alert products to empower sustainable land management and law enforcement actions against illegal forest activities (Lynch et al 2013, Finer et al 2018, Weisse et al 2019, Tabor and Holland 2020). Open distribution and availability of key forest disturbance alerts via nationally-hosted web portals and the World Resources Institute's Global Forest Watch platform have led to further increases in transparency of ongoing forest activities globally (Finer et al 2018, Tabor and Holland 2020).

Current operational systems rely predominantly on freely available medium scale resolution (30–100 m) optical satellite data (Souza et al 2009, Diniz et al 2015, Hansen et al 2016, Vargas et al 2019). The Brazilian Real-Time System for Detection of Deforestation uses data from the Advanced Wide Field Sensor onboard the Indian Remote Sensing satellites and provides monthly forest disturbance information at 56 m spatial scale (Diniz et al 2015). Other operational systems use medium resolution Landsat data to provide pixel-based disturbance alerts at 30 m scale, including the Peruvian Geobosques system (Vargas et al 2019) and the Global Land Analysis and Discovery (GLAD) pan-tropical forest disturbance alerts (Hansen et al 2016). The limited availability of cloud-free Landsat data in many parts of the humid tropics reduces the ability to track forest change events consistently on a near real-time basis (Souza et al 2013, Sannier et al 2014, Hansen et al 2016). In the Congo Basin, cloud-free observations are rare during the wet season in particular. In persistently cloud covered regions such as the western part of the Congo Basin, cloud-free Landsat observations can be more than 1 year apart (Sannier et al 2014, Hansen et al 2016, Tyukavina et al 2018).

In addition to preventing detection of disturbances in near real-time, long data gaps also represent a major challenge for detecting small-scale changes. After selective logging, for example, any sign of disturbance within remotely sensed data often disappears within weeks or months due to fast regeneration, canopy closure and understory revegetation (figure 1) (Asner et al 2004, Souza et al 2005, Verhegghen et al 2015). Disturbances at the scale of a single tree canopy are unlikely to be captured by a 30 m Landsat alert regardless of cloud-free observation availability and require finer scale satellite data (Verhegghen et al 2015).

**Figure 1.** Logging road expansion (January, February, March) followed by selective logging (September and October) and rapid canopy closure (November), depicted by monthly PlanetScope imagery with a spatial resolution of 3 m (Planet Team 2020). Circles indicate canopy openings after logging (October) and canopy closure within one to two months (November). Location: Sangha-Mbaéré district, Central African Republic (center coordinate: 3.95°N, 15.99°E).
Download figure:
Standard image High-resolution image

Satellite-based high-resolution radar sensors use long-wavelength energy that penetrates through clouds and smoke and is sensitive to changes in the physical structure of forests, resulting in major advantages for tropical forest disturbance monitoring (Joshi et al 2016, Reiche et al 2016). In the past, inconsistent data acquisitions and commercial data distribution of key radar missions restricted opportunities for operational large-scale forest monitoring (Reiche et al 2016). New and near-future radar satellites now provide a wealth of free, consistent radar data for global forest monitoring (Reiche et al 2016). For example, the JiCA-JAXA Forest Early Warning System in the Tropics (JJ-FAST) employs long-wavelength ALOS-2 PALSAR-2 L-band radar data with a spatial scale of 50 m, and provides event-based forest disturbance detection for the pan-tropics updated every 1.5 months (Watanabe et al 2018). A minimum event area size of 2 ha (version 3), however, limits the capacities of the JJ-FAST system to detect small-scale changes.

With the European Sentinel-1A and 1B C-band radar satellites launched in 2014 and 2016 (Torres et al 2012), global temporally dense and high resolution radar data at 10 m spatial scale are freely available for the first time. While the majority of the tropics, including the Congo Basin, is consistently covered every 12 days by one ascending or one descending orbit, tectonically and volcanically active regions such as western South America, East Africa and parts of South East Asia are covered every 6 days (Potin et al 2016). Compared to long-wavelength L-band (∼23 cm) radar, short wavelength C-band (∼5.6 cm) radar is considered less suitable for forest change monitoring because of the rapid saturation of the signal over forests, and the higher sensitivity to surface moisture fluctuations (Ulaby et al 1986). A number of recent studies, however, have shown that the dense observation frequency of Sentinel-1 in combination with high spatial resolution can overcome major shortcomings of C-band radar and have demonstrated the potential of Sentinel-1 for accurate and timely tracking of small-scale forest disturbance, including selective logging (Bouvet et al 2018, Reiche et al 2018b, Ballère et al 2021, Hethcoat et al 2020, Hirschmugl et al 2020, Hoekman et al 2020).

Here we present a Sentinel-1-based forest disturbance alert system deployed and validated for the humid tropical forest of the Congo Basin. The system is implemented in Google Earth Engine (Gorelick et al 2017), and developed in collaboration with World Resource Institute's Global Forest Watch program and Google.

2. Study area and definitions

We defined the study area as humid tropical forest within the six Congo Basin countries (figure 2A). The climate in the Congo Basin countries is warm and humid with mean annual temperatures between 24 °C and 26 °C and mean annual rainfall between 1400 and 2200 mm (Harris et al 2014).

We used the global forest change products (version 1.7) developed by Hansen et al (2013) to derive a benchmark forest map for the year 2018 which we used to limit the detection of new disturbance alerts to undisturbed humid tropical forest at the beginning of our monitoring in 2019. We used the year 2000 tree canopy cover product with a threshold of >50% tree canopy cover and excluded all historic forest loss from 2000 to 2018 (Hansen et al 2013). The evergreen forest layer of the Collection 2 Copernicus Global Land Cover dataset (Buchhorn et al 2020) was used to exclude dry tropical forests. We further refined and removed errors from the map with the aid of a radar-based global forest map (Martone et al 2018). Our forest benchmark map covered 186 115 kha (1 kha = 1000 ha) humid tropical forest area (figure 2A).

We defined forest disturbance as the complete or partial removal of tree cover within a 10 m × 10 m Sentinel-1 pixel (∼0.01 ha). Complete tree cover removal is associated with stand-replacement disturbance at the Sentinel-1 pixel scale, while partial removal mainly represents disturbances associated with boundary pixels and selective logging. This definition is similar to other operational satellite-based forest change products (Hansen et al 2013, 2016, Vargas et al 2019).

We applied a conservative minimum mapping unit of 0.2 ha (20 Sentinel-1 pixels, 2.2 Landsat pixels) as we aimed for an alert with a low false detection rate. Mapping very fine-scale disturbance events covering few Sentinel-1 pixels is naturally associated with increasing false detection rates, for example due to remaining speckle noise or local moisture fluctuations (Bouvet et al 2018, Reiche et al 2018b, Hirschmugl et al 2020).

Our alert is a generic one for detecting forest disturbances. We do not distinguish human-induced from natural forest disturbances, similar to other forest disturbance alerting products (Hansen et al 2016, Watanabe et al 2018). Natural disturbances may include windthrows, landslides, or meandering rivers.

3. Sentinel-1 satellite data

We employed dual-polarized (VV and VH) high resolution Sentinel-1 Ground Range Detected (GRD) products acquired in interferometric wide swath and available in the Google Earth Engine collection (Google Earth Engine 2020).

The GRD images have a pixel spacing of 10 m at which the full information detail is guaranteed, and a spatial resolution, i.e. the ability to separate between adjacent target objects on the ground, of approximately 20 m × 22 m (European Space Agency 2020a). In this study, GRD images acquired in ascending and descending orbits were considered (figure 2B), which corresponded to an annual total of approximately 5300 images. For the period between January 2019 and July 2020, 72% of the benchmark forest map was covered every 12 days, 21% every 6 days, 7% less than 6 days, and 0.28% was not covered.

4. Methods

4.1. Sentinel-1 data processing

Sentinel-1 GRD images available in the Google Earth Engine collection (Google Earth Engine 2020) have already undergone pre-processing using the European Space Agency's Sentinel-1 Toolbox (European Space Agency 2020b). This includes the application of the orbit files, thermal noise and GRD border noise removal, radiometric calibration to sigma naught, and range-Doppler terrain correction. We applied additional pre-processing steps to further enhance the data, including removing remaining GRD border noise and artefacts occasionally caused by heavy convective rain cells (Danklmayer et al 2009) as well as applying radiometric slope correction (Hoekman and Reiche 2015, Vollrath et al 2020) and adaptive speckle filtering (Quegan and Yu 2001). The final output was geocoded and topographically normalized gamma-naught VV- and VH-polarized backscatter images at 10 m pixel spacing.

We generated historical time-series metrics of backscatter using all Sentinel-1 GRD images available for 2017 and 2018, including the median and standard deviation derived individually for ascending and descending orbits and VV- and VH-polarization. The time-series metrics were used to describe the backscatter distribution of stable forest at the pixel level (Reiche et al 2018b). We assumed that all observations in the 2 year historical period represented stable forest and did not consider possible changes or regrowth processes.

4.2. Sentinel-1-based forest disturbance detection

In our system, a forest disturbance alert is triggered based on a single observation from the latest Sentinel-1 image. Subsequent observations are used to increase confidence and confirm or reject the alert. The date of the alert is set to the date of the image that first triggered the alert (Reiche et al 2015, 2018b).

Figure 3 depicts the flowchart of the presented Sentinel-1-based forest disturbance alerting method. The latest Sentinel-1 GRD image was accessed via Google Earth Engine. Quality control, radiometric slope correction and speckle filtering were applied (see section 4.1) before we normalized the image by matching the median backscatter distribution of forest to the expected median distribution defined by the historical time-series metrics in order to mitigate dry season effects (Reiche et al 2018a). Forest disturbances were detected using the probabilistic algorithm described in Reiche et al (2018b). First, VV and VH backscatter observations were converted into forest and non-forest probabilities using pixel-specific Gaussian Mixture Models derived from the historical time-series metrics, and the larger non-forest probability was selected. An alert was triggered for non-forest probabilities >0.75. For triggered alerts, Bayesian updating (Reiche et al 2015) was used to calculate the forest disturbance probability and iteratively update it with the non-forest probability of later observations. Unconfirmed, low confidence alerts were provided for a forest disturbance probability >0.85. Alerts were confirmed with high confidence for forest disturbance probabilities >0.975 within a maximum period of 90 days from first detection.

The resulting images have a pixel size of 10 m. We clustered alert pixels in eight connected directions and removed clusters smaller than the minimum mapping unit of 0.2 ha.

4.3. Validation

We validated the Sentinel-1-based forest disturbance alert product for the year 2019 using high-resolution optical satellite data. The 2019 product was generated in an emulated near real-time mode, similar to other studies validating forest disturbance alerts (Hansen et al 2016, Reiche et al 2018a). We used probability sampling (Stehman et al 2003) and generated three strata with a total of 1100 sample points. We allocated 500 sample points to the stratum 'forest disturbance' to have a good estimate of the rate of false detection (commission error), important for assessing near real-time systems in particular. To target omission errors, which are more likely to occur in spatial proximity to existing forest disturbances, we allocated 300 sample points to the stratum 'No disturbance within a 20 pixel buffer zone' following Olofsson et al (2020). Additionally, we allocated 300 sample points to the stratum 'No disturbance outside the buffer zone'. The sampling and population unit corresponded to our 10 m × 10 m Sentinel-1 pixel (∼0.01 ha).

We checked each of the sample locations for forest disturbance by visually examining monthly PlanetScope image mosaics (3 m spatial resolution) (Planet Team 2020), and supported the analysis with Sentinel-2 imagery (10 m spatial resolution). Cloud-free PlanetScope and Sentinel-2 data were infrequent in many parts of the Congo Basin, and often limited the verification process to quarterly or half-yearly time steps. In extreme cases where no cloud-free data was available during 2019, we depended on the Sentinel-1 time series imagery itself.

Our benchmark forest map relied on a Landsat-based annual tree cover loss product (Hansen et al 2013) to exclude forest disturbance events that occurred in 2018 and before. Some prior disturbance events were not detected by the Landsat-based algorithm, for example due to a lack of cloud-free Landsat data at the end of 2018. Disturbances detected by the Sentinel-1-based disturbance alerts in 2019 that originally occurred before 1 January 2019 (as visible in PlanetScope time series) were labelled as 'pre-2019 disturbance', but not reported as false detections (commission error).

Boundary pixels often represent partial tree cover at the edge of larger disturbance events, which makes validation ambiguous and difficult (Hansen et al 2016), in particular when dealing with higher spatial resolution data such as from Sentinel-1. Sample pixels that were ambiguous and that were on the boundary of larger disturbance events clearly visible in Planet imagery were labelled as 'boundary pixel', but not reported as false detections (commission error). This was done to not penalize our alerting system since the goal of alerting systems is the correct detection of new events and not the unbiased estimation of areas (Tang et al 2019).

In cases where a forest disturbance was visible for a sample location and not detected by our alerts (omission error), we digitized the omitted disturbance event and reported its area.

To estimate the accuracy of forest disturbance detection, we accounted for unequal inclusion probabilities between different strata as sample points were not allocated proportionally to the strata areas (Stehman et al 2003). Sample inclusion probability was calculated based on the number of sample points and strata areas. The estimation weights, i.e. the inverse of inclusion probability, were then used to construct an area weighted confusion matrix and calculate user's accuracy (1—false detection rate, 1—commission error) and producer's accuracy (detection rate, 1—omission error) (Stehman et al 2003, Stehman 2014). We excluded samples representing events <0.2 ha to estimate the producer's accuracy of our product at the applied minimum mapping unit of 0.2 ha. We used all samples to get an estimate of the producer's accuracy including disturbance events <0.2 ha which were not mapped due to the applied minimum mapping unit.

To assess the accuracy of unconfirmed alert pixels that were eventually confirmed, we calculated the total number of unconfirmed alerts throughout 2019 at four randomly selected 1° tiles. We used the resulting map from 1 July 2020 to retrieve the total number of confirmed alerts for 2019 and calculate the percentage.

For confirmed alert pixels, we assessed the time between the date when the disturbance was first detected by the algorithm and when it was confirmed (Reiche et al 2018b).

5. Results and discussion

5.1. Forest disturbance alerts

Disturbance alerts for the humid tropical forest of the Congo Basin for the period between 1 January 2019 and 1 July 2020 are depicted in figure 4. Two detail maps show the expansion of logging roads and selective logging in the Central African Republic (B), and smallholder agriculture in the Democratic Republic of the Congo (C). The white box in (B) encloses the area depicted by high-resolution PlanetScope imagery in figure 1. It shows the detection of some of the canopy openings during the short period between disturbance and canopy closure and/or understory revegetation.

While this product is designed to provide near real-time alerts and not to represent the true area of disturbances, the mapped areas can be used to evaluate relative temporal and spatial trends. We mapped 1431 kha of disturbance (of which 57 kha represented an unconfirmed alert), and a total of 4.07 million disturbance events (table 1). Forest disturbance area totals varied greatly by country. For the year 2019, for example, the mapped disturbance area totals ranged from 13 kha for Equatorial Guinea to 541 kha for the Democratic Republic of the Congo.

Table 1. Mapped forest disturbances (confirmed and unconfirmed) for the period 2019–1 July 2020 grouped by different size ranges.

Alert event size (ha)	No. Events	% Events	Area (kha)	% Area
0.2–<0.5	3238 503	79.6	469.1	32.8
0.5–<1	523 541	12.8	361.2	25.2
1–<5	295 872	7.3	502.6	35.1
⩾5	10 228	0.3	98.1	6.9
Total	4068 144	100.0	1431.0	100.0

Small forest disturbance events dominated the detected alerts, with 79.6% of all events having a mapped size between 0.2 and <0.5 ha (32.8% of the total mapped area). We assume that many of these are related to selective logging as seen in figure 4(B) for example. Disturbance events with a mapped size between 0.5 and <1 ha accounted for 12.8% (25.2% of the total mapped area), medium-scale events between 1 and <5 ha for 7.3% (35.1% of the total mapped area), and large-scale events ⩾5 ha for 0.3% (6.9% of the total mapped area).

The monthly distribution of forest disturbance areas shows a large variation between the six Congo Basin countries (figure 5). This can be linked to regional differences in the dry and wet season and related rainfall patterns. In general, forest disturbance rates were much higher during dry season months with low rainfall totals. Lower disturbance rates during the wet season are common in the tropics (Hansen et al 2016, Vargas et al 2019), as heavy rainfall makes many logging roads inaccessible and forest clearing operations are less feasible (Kleinschroth and Healey 2017).

The delivery of the Sentinel-1-based forest disturbance alerts is illustrated in figure 6. We provide information for each 10 m pixel within the boundary of the baseline forest map on (A) whether an alert was confirmed (high confidence alert) or is still unconfirmed (low confidence alert) at the date of the latest Sentinel-1 image, and (B) the date when the forest disturbance was first detected.

5.2. Validation results

The accuracy assessment yielded consistently high results (table 2). The user's and producer's accuracies of confirmed disturbance alerts were 97.6% and 95.0%, respectively, suggesting confident detection of forest disturbances larger than or equal to 0.2 ha. When including samples representing disturbance events <0.2 ha, the producer's accuracy was 83.5% indicating a high rate of fine-scale disturbance events which were not detected due to the application of a minimum mapping unit of 0.2 ha.

Table 2. Estimated user's accuracy (1—false detection rate) and producer's accuracy (detection rate) for confirmed alerts (% ± standard error).

User's accuracy
Disturbances ⩾0.2 ha	97.6 ± 4.8
Producer's accuracy
Disturbances ⩾0.2 ha	95.0 ± 25.8
All disturbances, including those smaller than the minimum mapping unit of 0.2 ha	83.5 ± 37.5

Out of the 488 correctly detected disturbance samples (out of a total of 500), 46 samples included alert detections of forest disturbance that occurred before 1 January 2019 ('pre-2019 disturbance'), but were not included in the Landsat-based tree cover loss products used to generate our benchmark forest map. We also identified 24 boundary pixels ('boundary pixel') that could not be validated unambiguously and were thus reported as correctly detected.

We examined the 12 samples labeled as commission error and identified three primary sources of false detections: dynamics in non-forested swamp and savannah areas that remained in our benchmark forest map (7 samples), swamp forest dynamics such as seasonal inundation changes that cause strong radar backscatter variations (2 samples), and unidentifiable causes of falsely detected small-scale events (3 samples). For the first source of error, false detections in non-forested land can be decreased by introducing a more accurate benchmark forest product.

The percentage of unconfirmed alerts that were eventually confirmed was found to be 62 ± 28% (% ± standard error). This means that unconfirmed alerts provide a good early indication of new disturbances, but only confirmed alerts provide high confidence information.

Figure 7 depicts the number of days between the first detection (trigger) of a disturbance alert and its confirmation, separately for boundary and core pixels of mapped disturbance events. In general, core pixels represent complete (or close to complete) tree cover removal, and therefore the algorithm only requires a few observations to confirm core pixel with high confidence. We found 74% of all core pixels being confirmed within 24 days and 95% within 48 d. Boundary pixels, in contrast to core pixels, often represent partial tree cover removal and on average require more observations to be confirmed with high confidence. Results showed 53% of boundary pixels being confirmed within 24 days and 83% within 48 d.

We also assessed the confirmation time for different Sentinel-1 revisit times. Within 24 days, 62% of the alerts (boundary and core pixels) were confirmed for areas with a 12 days revisit time and 66% for ⩽6 days revisit time. Within 48 days, 88% and 89% were confirmed for 12 days and ⩽6 days revisit times, respectively.

Depending on the Sentinel-1 revisit times, the actual date on which the forest disturbance occurred is usually not more than 6–12 days before the reported date of the Sentinel-1 image that first triggered the alert (Reiche et al 2018b). In some circumstances, the first detection of new disturbances can be delayed due to, for example, increased soil moisture after strong rain events or remaining stems after logging, both of which can cause the Sentinel-1 C-band radar backscatter of newly disturbed areas to remain at or increase back to the level of undisturbed forest (Woodhouse et al 1999, Reiche et al 2018a).

5.3. Comparison with Landsat-based products

Compared to the GLAD alerts, which are conservative by design (Hansen et al 2016), our detected forest disturbance area was found to be 3 (Democratic Republic of the Congo) to 40 (Equatorial Guinea) times higher (table 3). The ratio was largest for the western Congo Basin countries of Cameroon, Equatorial Guinea and Gabon, where cloud cover is more persistent than in other parts of the Congo Basin (Tyukavina et al 2018). Comparison with Landsat-based annual tree cover loss showed similar magnitudes of detected disturbance areas with ratios ranging between 0.7 (Democratic Republic of the Congo) up to 1.6 (Central African Republic). The key advantage of the Sentinel-1-based alert product is its availability in near real-time.

Table 3. Mapped forest disturbance area (in kha) for the year 2019 compared for Sentinel-1-based alerts (this study), Landsat-based GLAD alerts (Hansen et al 2016), and Landsat-based annual tree cover loss (Hansen et al 2013).

	Sentinel-1-based alerts (this study)	GLAD alerts (Hansen et al 2016)	Annual tree cover loss (Hansen et al 2013)
Cameroon	134.3	6.0	86.3
Central African Republic	26.0	3.6	16.5
Democratic Republic of the Congo	541.2	179.3	827.3
Equatorial Guinea	13.4	0.3	8.3
Gabon	31.1	1.4	22.8
Republic of the Congo	28.5	4.7	41.6
Total	774.5	195.3	1002.7

Visual comparison with the GLAD alerts and Landsat-based annual tree cover loss shows the ability of our Sentinel-1-based alerts for improved detection of logging roads and selective logging as well as improved spatial detail of mapping smallholder agriculture (figure 8).

**Figure 8.** Visual comparison between our Sentinel-1-based alerts for 2019 with the Landsat-based GLAD alerts (Hansen et al 2016), and Landsat-based annual tree cover loss (Hansen et al 2013). The first area (A)–(C) depicts logging roads and selective logging expansion in the province Haut-Ogooué, Gabon (center coordinate: 1.68°S, 13.17°E). The second area (D)–(F) depicts smallholder agriculture in the Équateur Province, Democratic Republic of the Congo (center coordinate: 1.37°N, 20.59°E). A Sentinel-2 composite was used as base map.
Download figure:
Standard image High-resolution image

6. Conclusions

Here we present a Sentinel-1-based forest disturbance alert product for the humid tropical forest of the Congo Basin. Our disturbance alerts provide confident and rapid detection of events larger than or equal to 0.2 ha. Almost 80% of all mapped disturbance events were smaller than 0.5 ha and likely represented selective logging activities in many cases. The availability of consistent, gap-free Sentinel-1 radar observations every 6–12 days at 10 m spatial scale enabled the timely detection of such small-scale disturbances, and the confirmation of nearly all alert pixels within a few weeks after their first detection.

The new alert product helps overcome some of the data scarcity of up-to-date logging road and selective logging information in the Congo Basin region (Umunay et al 2019, Jackson and Adam 2020) and offers a more accurate look at the spatio-temporal forest dynamics than ever before. Results revealed a strong difference in the monthly distribution of forest disturbances across the six countries, with most disturbances occurring in the dry season month.

The primary limitations of the presented alerts are twofold. First, the high sensitivity of C-band radar to moisture variations caused a number of false alerts in swamp forests. Second, fine-scale forest disturbances <0.2 ha were not mapped due to the applied minimum mapping unit of 0.2 ha. Our results suggest that disturbance rates in the Congo Basin are therefore even higher than what was reported. In future research, we aim to decrease the minimum mapping unit, while preserving a low commission error of the alerts. The verification of fine-scale and low impact disturbances such as single tree canopy damages, however, is very challenging due to the shortage of high and very high-resolution and cloud-free reference data at weekly to monthly time steps.

This study highlights the value of Sentinel-1 dense time series data for large-area and rapid tropical forest monitoring that can be expanded to more areas in the humid tropics. The guaranteed availability of Sentinel-1 data for at least 10+ years through the upcoming Sentinel-1C and -1D satellites (Torres and Davidson 2019) provides the necessary data continuity for long term operational products. The integration with data from new high spatial and temporal resolution radar sensors, for example from the upcoming L-band NISAR (NASA/ISRO Synthetic Aperture Radar, planned launch in 2021) mission (Rosen et al 2016), and the integration with optical satellite data (e.g. with Sentinel-2) or products (e.g. Landsat-based GLAD alerts, Hansen et al 2016) offers opportunities to further improve the timeliness of confident forest disturbance alerting (Reiche et al 2016).

The new alert product is made available as a public good via the Global Forest Watch platform. It contributes to the efforts of Global Forest Watch to make changes in tropical forests more transparent and actionable, and provides stakeholders in the Congo Basin with improved near real-time forest disturbance information that can support their sustainable forest management and law enforcement efforts.

Acknowledgments

This project received funding through Norway's Climate and Forest Initiative (NICFI), the US Government's SilvaCarbon program, and the European Union's Horizon 2020 project 'Capacity for Copernicus REDD+ and Forest Monitoring Services (REDDCopernicus)' project (Grant Agreement No 821880). This work was also supported by the Global Forest Observation Initiative (GFOI) research and development program. PlanetScope data was provided through the Planet ambassador program. Contains modified Copernicus Sentinel data (2015–2020). We thank the two anonymous reviewers for their valuable comments.

Data availability statement

The data that support the findings of this study are available upon reasonable request from the authors.

Forest disturbance alerts for the Congo Basin using Sentinel-1

Article metrics

Submit

Author e-mails

Author affiliations

ORCID iDs

Dates

Peer review information

Abstract

1. Introduction

2. Study area and definitions

3. Sentinel-1 satellite data