Timely and accurate detection of forest cover loss is necessary to support enforcement actions in reducing unsustainable and illegal forest activities. Current near real-time (NRT) forest monitoring methods are mainly pixel based, neglecting what happens in their surroundings. The integration of neighboring information and spatial data has the potential to increase the detection accuracy of NRT forest cover loss.
The recently launched Sentinel-1 satellites, provide dense and regular C-band SAR data, at the global level and free of charge. In addition, frequently updated spatial data related to activities in forest are increasingly available, including road networks from OpenStreeMap and real-time active fire hotspot data provided by various satellite missions (e.g. MODIS, VIIRS). Additionally, very high resolution (VHR) satellite imagery, e.g. those provided by Planet Labs, now allow to track logging road developments in detail. But despite the abundance of raw data at hand, approaches to combine those dense satellite observations and frequently updated spatial data to improve near real-time forest monitoring are lacking.
Here, we present a multi-source approach that combines dense Sentinel-1 SAR observations with frequently updated spatial data, i.e. logging and major roads, and active fire alerts, to detect forest cover loss in near real-time. We provide results for the province of Madre de Dios, Peru. Forest cover loss in Madre de Dios is mainly driven by agriculture expansion, mining and timber extraction.
First, logging roads were visually detected using Planet VHR images, and combined with Peru’s major road network available through the Peruvian Ministry of Transport and Communication, to develop a temporal dynamic road network dataset. Second, Euclidean distance to road networks and fire alerts were calculated for all pixels within the study area. Third, Sentinel-1 time series observations, distance to road networks and fire alerts were transformed into non-forest probability through a random forest model. Finally, distance to roads and Sentinel-1 observations were chronologically updated, and the probability of forest cover loss was calculated based on “Bayts”, the probabilistic machine learning approach developed by Reiche et al. (2018). This approach calculates the probability of forest cover loss once a new satellite image is available, and uses future observations to confirm or reject forest cover loss detection events.
Preliminary results show an improved timing of forest cover loss detection of approximately 24 days, i.e. two Sentinel-1 images, compared to the single-source approach using Sentinel-1 only. Results show a clear relationship between the road’s temporal development and its importance to detect forest cover loss in their neighborhood. Distance to recently developed roads was found to be a strong predictor of new forest cover loss.. The spatial and temporal distance to relevant road developments and active fire events can help to reduce commission error, especially in forest areas located far away from human activities.
Future research will be conducted in assessing the forest cover loss detection events for Madre de Dios province, and the inclusion of other spatial data sets, such as water bodies.

This project was funded by the National Council for Science and Technology of Mexico (CONACYT).

References:
Reiche, Johannes et al. 2018. “Improving near-Real Time Deforestation Monitoring in Tropical Dry Forests by Combining Dense Sentinel-1 Time Series with Landsat and ALOS-2 PALSAR-2.” Remote Sensing of Environment 204(April 2017): 147–61. https://doi.org/10.1016/j.rse.2017.10.034.

Image caption: (A) Single-source (Sentinel-1 only) forest cover loss detection. (B) Multi-source (Sentinel-1 + roads) forest cover loss detection. Start of the monitoring period (black line), time flagged (red dotted line), time confirmed (red line)

Satellite imagery is the primary tool for providing information on newly deforested areas in vast and largely inaccessible forests. Most monitoring approaches have relied so far predominantly on optical remote sensing, despite drastic limitations in tropical areas due to cloud cover and haze. The use of SAR remote sensing for such applications was until recently almost inexistent, for various reasons including the lack of freely available and consistent SAR time series and the lack of SAR expertise in the research groups involved in the production of forest loss products [1]. In this regard, the recent launch of the Copernicus Sentinel-1 constellation, which offers freely available C-band SAR imagery acquired globally following a systematic acquisition scenario, is a game changer that opens a new era for the development of near-real time SAR-based forest loss monitoring applications.
In fact, SAR is one of the most promising remote sensing tools for the near-real time mapping of forest disturbances in the tropics, thanks to its all-weather ability. Nevertheless, the use of C-band SAR data in forest loss mapping has not been well developed so far, and thus its operational application has not yet been realized and still requires specific methodological developments. The first studies that aimed at exploiting the potential of the Sentinel-1 time series for forest loss detection generally rely on the assumption that forest loss is characterised by a decrease in C-band backscatter [2], [3]. However, this assumption is not always verified, because of the high sensitivity of C-band SAR to the remaining vegetation in deforested patches and to environmental parameters such as soil moisture, as demonstrated in [4]. Another approach has been proposed to tackle this issue, based on the detection of the shadows that appear in SAR images at the edges of deforested patches [4]. This approach however also has its own limitations (better results are obtained when both orbit orientations, descending and ascending, are available over the same area, when trees are relatively high, and when the canopy closure in the forest is important) and cannot be applied under all conditions.
In this presentation, we will discuss the challenges and opportunities linked to the use of Sentinel-1 for near-real time forest loss monitoring, based on analyses carried out in Amazonia (Peru, Brazil, French Guiana), in Southeast Asia (Vietnam, Laos, Cambodia), and in Africa (Gabon, Cameroon, Malawi). The algorithmic choices to be adopted are found to depend on the type of forest (tree height, canopy cover), on the type of forest loss (size of affected patches, post-deforestation management, etc) and to a lesser extent on the local availability of Sentinel-1 imagery (descending and ascending orbit, or only one orbit orientation). The potential of a synergistic use of Sentinel-1 with imagery from other sensors like L-band SAR systems (in the present, such as ALOS-2/PALSAR-2, or in the future, such as NISAR and ALOS-4/PALSAR-3) and the optical Sentinel-2 constellation will also be discussed.
References:
[1] J. Reiche et al., ‘Combining satellite data for better tropical forest monitoring’, Nature Climate Change, 27-Jan-2016. [Online]. Available: https://www.nature.com/articles/nclimate2919. [Accessed: 07-Mar-2018].
[2] J. Reiche, E. Hamunyela, J. Verbesselt, D. Hoekman, and M. Herold, ‘Improving near-real time deforestation monitoring in tropical dry forests by combining dense Sentinel-1 time series with Landsat and ALOS-2 PALSAR-2’, Remote Sens. Environ., vol. 204, pp. 147–161, Jan. 2018.
[3] J. Reiche, R. Verhoeven, J. Verbesselt, E. Hamunyela, N. Wielaard, and M. Herold, ‘Characterizing Tropical Forest Cover Loss Using Dense Sentinel-1 Data and Active Fire Alerts’, Remote Sens., vol. 10, no. 5, p. 777, May 2018.
[4] A. Bouvet et al., ‘Use of the SAR Shadowing Effect for Deforestation Detection with Sentinel-1 Time Series’, Remote Sens., vol. 10, no. 8, p. 1250, Aug. 2018.

Remote sensing and geographic information system (GIS) data have become essential for land managers. It is especially useful for monitoring forest ecosystems that can extend over vast areas, which can be difficult to access. French Guiana, a French overseas department bordering Brazil and Suriname, is one of the territories with the highest rate of forest cover in the world. The tropical rain forest dominates about 8 million hectares, which represents 95% of the territory. As most of the forests are located on state land, their management has been entrusted to the National Forest Office (ONF), whose primary mission is to ensure the sustainable and multifunctional management of forests. Given the large size of the territory and the limited resources of ONF in the department, this mission that includes the control and monitoring of forest cover integrity is largely accomplished through the use of remote sensing data.

The development of operational methods for monitoring deforestation have been largely dominated by the use of optical data, which were more democratic and more accessible than radar data. However, optical data are sensitive to atmospheric conditions and cloud cover can limit their use, especially in tropical areas such as French Guiana. The weakness of optical sensors has recently been offset by free access to Sentinel-1 radar data from the European Space Agency (ESA). These high spatial and temporal resolution data offer the opportunity to develop operational systems for monitoring deforestation that are capable to "see through" clouds and provide results all year long.

This paper details the exhaustive, semi-automatic and near-real-time deforestation monitoring method that was developed for ONF needs in French Guiana from the Sentinel-1 time series analysis. The deforestation maps generated every 15 days by the method are used daily by ONF to carry out its missions of control and surveillance of the territory. After one year of operation, the results are very promising (see illustration below). Based on experience feedback since the start of operation, this paper discusses the strengths and weaknesses of the method with respect to the characteristics of the sensor, in comparison with optical data and from quantitative assessments of accuracy.





The detection of deforestation is based on the VH polarization of the Sentinel-1 time series, which is more sensitive to vegetation density than the VV polarization (Mermoz and Le Toan, 2016). The alert system, which aim to be robust, cost-effective and near-real-time, consist in three main steps:

Data preprocessing (calibration, orthorectification, temporal filtering)

Optimization of contrast between forest and deforested areas, notably by attenuation of the seasonal effect on radiometry

Triggering an alert in forest areas in response to a phenomenon of change (i.e. deforestation) causing a large variation of radiometry over time (> 1 dB).

The second step is to extract for each date the average value of the radiometry of a forest area stable over the entire series. For each pixel, this value is used to adjust the radiometry level in order to mitigate sensor effect or seasonal effect over forested area and consequently reduce possible false alarm. In addition in order to smooth the time series we apply what is called the exponential mean. that is a cumulative mean over the time. This approach allows better detection of deforestation and reduce false alarms. As illustrated in the top figure of the illustration.

The third step is to detect radiometric decrease between one date regarding the previous ones.
For that we used the so-called modified Z-score indicator that consist first to estimate over N dates before a current date i the median and MAD (Median Absolute Deviation). Then we compute the modified Z-score :
Mod Z-scorei = (ti - Median(ti-1 to ti-N)) / MAD(ti-1 to ti-N)
where ti is the VH polarisation in dB from the date i, Median(ti to ti-N) is the median of the VH polarization calculated from date i-N to i-1 and MAD(ti-1 to ti-N) is the mad from date i-N to i-1.
This indicator is related to the probability that a value is statistically near the median which mean that more the absolute value of the modified z-score is high more the probability to have a change in the time series is important.
To detect a deforestation event we apply a fixed threshold over the Zscore.
Results are illustratedin the bottom of the illustration.

Time series of satellite data are a valuable source to detect forest disturbance pattern in time and space. Several approaches exist to explore satellite time series. Many of them make use of composites at regular observation intervals such as monthly, seasonal or annual, and sometimes even multi-annual. The generation of artifact-free composites of optical data requires high-quality and consistent pre-processing, in particular when multi-sensor data is to be fused. Cloud coverage is another challenge in many regions of the world. Therefore, SAR data is often the better choice. However, when long-term spatio-temporal forest change pattern and processes are to be explored, the optical Landsat system is often the only consistent and comprehensive data source with reasonable spatial resolution. Besides that, the spectral and spatial properties of Landsat are well-suited for forest monitoring. Common techniques to explore optical time series comprise temporal segmentation and breakpoint detection. Currently, only very few methods exist that generate time series of all artifact-free observations rather than reducing them to a series of optimal dates, e.g, by compositing. Recent approaches also make use of multitemporal metrics to characterize the spectral variability of specific land cover classes within a certain period. Here, we use time series of all available Landsat and Sentinel-2 observations to explore earth observation data for forest disturbance monitoring over southern Vancouver Island, British Columbia, Canada. We calculated the disturbance index (DI) for each image and applied a break detection method to identify clear-cut events in time and space. For each pixel, we applied support vector machine (SVM) regression to estimate the DI value of specific dates in near future. However, the DI of any given date can be estimated from the SVM regression output if desired. By subtracting the DI value of real observations (e.g. cloud-free images) from the predicted values, we are able to detect recent changes, i.e. new clear-cuts. Other types of disturbance can be detected with this approach but rarely occur in the study site. Our results demonstrate applicability of the DI to reveal forest dynamics at very high temporal resolution. Instead of years, disturbance dates are estimated at virtually daily resolution. The method results in maps of clear-cut timing at daily resolution and regrowth trends expressed as time span of DI recovery. DI recovery does not necessarily correspond to forests’ ecological development. It rather corresponds to crown closure and regreening of forests. Validation of historical changes showed high conformity with existing maps. The results of the detection of very recent changes are reasonable. The new approach allows using the full capacity of dense time series data allowing for better understanding of forest processes and phenology and near-real time change detection. The growing archive of Landsat-like and Sentinel data will widen the field of its application in future since the presented approach is not restricted to a specific data source.

Every year about 2.1 PgC is being released from biomass burning. Given its importance for climate simulations accurate global burned area (BA) products are needed. Remote sensing is an essential approach to obtain such products. Most BA products have been based on optical data, with large areas not being observed in regions with permanent cloud covers such as the equatorial and tropical belt. Recently, in the framework of the ESA’s CCI Fire project, this shortcoming was mitigated through the development of BA mapping algorithms based on Sentinel-1 C-band SAR data. One of these algorithms was tested over several regions globally with results being similar or more accurate when compared to current global BA products (e.g. MCD64) based on MODIS data.

In this contribution we present the Sentinel-1 BA detection algorithm. The algorithm uses VV and VH polarizations to calculate temporal backscatter ratios by major land cover classes. ESA’s CCI Land Cover map is used as source of ancillary information. Variations between consecutive images are assessed using the Reed-Xiaoli detector (RXD), a technique widely used when detecting changes from optical remote sensing data. Anomalous changes (AC) are detected from the RXD with likely burned and unburned pixels being extracted using the AC scores and ancillary information from hotspots (e.g. VIIRS, MODIS C6). These pixels are further used to train land cover dependent Random forests models that separate burned from unburned areas where hotspots are not available. The existence of significant temporal decorrelation between the time of fire event and the maximum variation of the backscatter coefficient was observed in most areas. This issue was solved in the algorithm through an iterative detection process that consider not only the current detection period but also the previous and past ones up to a pre-defined temporal threshold.

The change in backscatter coefficient was analysed over burned and unburned areas as well as over areas affected by commission and omission errors to better understand the source of misclassified pixels, and thus improve the detection algorithm. The analysis was essential to better understand the interaction between the effect of fire and the C-band backscatter coefficient over different ecosystems

The accuracy of the BA detections was assessed over 18 (2.2M ha) globally distributed study areas as well as for 44 (2.6M ha) study areas distributed in the Amazon basin. The detection accuracy was compared with that of current global BA products (i.e. MCD64 C6). The reference burned area used for validation was obtained from supervised classification of Landsat-7/8 and Sentinel-2 image series.

The potential of radar backscatter temporal series for the BA detection is demonstrated in this work over a wide range of vegetation types globally distributed. Similar or more accurate detections when compared to MCD64 C6 global product are achieved, except for the grasslands. Future research will be focused on improving the BA detection accuracy using active and passive temporal series from the Sentinel-1 and 2 missions.

The paper presents the automatic Autochange method for change detection using Copernicus and other satellite data. It applies hierarchical clustering to two-temporal imagery. The algorithm is meant for detecting and identifying rapid changes such as cuttings, storm or pest damage. The detected change has happened in between the acquisition times of two images. The images do not have to be from the same stage of the growing period or represent identical spectral bands. It is assumed that the majority of the image area does not include changes of which the user would be interested in. The algorithm is based on the method that was presented by the same main author earlier but its principal logic has changed.

Due to the large data volumes of the images, not all the pixels are selected for the clustering, but a sample of objects is selected (Figure 1). The sample consists of mean spectral vectors of groups of pixels representing relatively homogeneous ground targets.

A clustering is performed to the selected homogeneous pixel groups of the earlier image. We have used the k-means clustering. The clusters from the earlier image are sorted by their ‘biomass index’.
The clustering result to the primary clusters from the earlier image is used for two purposes: to define the situation in the land cover before the changes and to characterize the initial state for the second level clustering of the newer image.

The homogeneous objects of the newer image are labeled with the ID numbers of the primary clusters without making a new clustering. The spectral intensity distribution of a primary cluster will become heterogeneous when a change has occurred in between of the acquisition of the two images. A within-cluster clustering of the newer image is performed to extract the heterogeneous objects from the rest of the cluster.

The Euclidean distance between each secondary cluster and its primary cluster of the newer image is computed as the length of the spectral mean vector and it represents the magnitude of change.

The change type is computed using two spectral features. One feature is the ‘biomass index’. The other feature is the Normalized Difference Vegetation Index or NDVI.

The change types do not take into consideration the magnitude of change but any observation involved in the clustering receives one of the four possible change types.

The spectral statistics of the homogeneous objects that were selected to the analysis are used to compute the output for every pixel of the area of interest. The output of Autochange is a three-band image including the primary clusters from the earlier image, the change magnitude and change type.

The final change classification is compiled automatically by applying simple logical operations to the three-band output image. The threshold to separate the changed and unchanged pixels is defined automatically by analyzing the cumulative histogram of the change magnitude output.

The method was tested with two Sentinel-2 images from 2015 and 2016 and Landsat 8 and Sentinel-2 images from the same years. The reference data were from forest harvester measurements showing borders of clear-cut forest stands. Only five of 1866 uncut forest stands, and one of 141 clear-cut stands was incorrectly classified. Operational use of Autochange is beginning and research for slighter changes than clear cuts and Autochange application to radar images and radar/optical images is ongoing. The method will be also tested for other applications such as urban changes.