Elsevier

Remote Sensing of Environment

Volume 204, January 2018, Pages 147-161
Remote Sensing of Environment

Improving near-real time deforestation monitoring in tropical dry forests by combining dense Sentinel-1 time series with Landsat and ALOS-2 PALSAR-2

https://doi.org/10.1016/j.rse.2017.10.034Get rights and content

Highlights

  • Dense Sentinel-1 time series combined with ALOS-2 PALSAR-2 and Landsat 7 and 8

  • Deforestation events at a dry tropical site in Bolivia detected with high accuracy

  • Sentinel-1 detects deforestation more timely than ALOS-2 PALSAR-2 or Landsat.

  • Combination of time series increased observation density and temporal accuracy.

  • Impact of uncertainties assessed and compared for four key user scenarios.

Abstract

Combining observations from multiple optical and synthetic aperture radar (SAR) satellites can provide temporally dense and regular information at medium resolution scale, independently of weather, season, and location. This has the potential to improve near real-time deforestation monitoring in dry tropical regions, where traditional optical only monitoring systems typically suffer from limited data availability due to persistent cloud cover. In this context, the recently launched Sentinel-1 satellites promise unprecedented potential, because for the first time dense and regular SAR observations are free and openly available. We demonstrate multi-sensor near real-time deforestation detection in tropical dry forests, through the combination of Sentinel-1 C-band SAR time series with ALSO-2 PALSAR-2 L-band SAR, and Landsat-7/ETM+ and 8/OLI. We used spatial normalisation to reduce the dry forest seasonality in the optical and SAR time series, and combined them within a probabilistic approach to detect deforestation in near real-time. Our results for a dry tropical forest site in Bolivia, showed that, as a result of high observation availability of Sentinel-1, deforestation events were detected more timely with Sentinel-1 than compared to Landsat and ALOS-2 PALSAR-2. The spatial and temporal accuracies of the multi-sensor approach were higher than the single-sensor results. We improved the precision of the reference data derived from the multi-sensor satellite time series, which enabled a more robust estimation of the temporal accuracy. We quantified how the near real-time deforestation detection is associated with a trade-off between the confidence in detection and the temporal accuracy. We showed that the trade-off affects the choice on how to use the near-real time data for different applications such as fast alerting with high temporal accuracy but lower confidence versus accurate detection at lower temporal detail. When aiming for a high confidence in change area estimates for example, deforestation was detected with a user's accuracy of 88%, a producer's accuracy of 89% (low area bias), and a mean time lag of 31 days using all sensors. This is on average 7 days earlier than when using only Sentinel-1 observations, and six weeks earlier than when relying only on Landsat observations. We showed that confident near real-time deforestation alerts can be provided with a mean time lag of 22 days, but these are associated with a higher commission error. With more dense time series data expected from the Sentinel-1 and -2 sensors for the upcoming decade, spatial and temporal detection accuracy of multi-sensor deforestation monitoring in the tropics will improve further.

Introduction

Satellite-based monitoring systems are the primary tools for providing near real-time (NRT) information on newly deforested areas in vast and inaccessible tropical forests. Their potential to empower governments and communities to enact timely actions against illegal and unsustainable forest activities, and to respond to natural disasters is increasingly recognised (Assunção et al., 2013, Lynch et al., 2013, Wheeler et al., 2014, Hansen et al., 2016). In the context of satellite-based monitoring, NRT refers to the capacity to detect new changes in satellite images once they are available (Reiche et al., 2015a). At an operationalised level, the NRT detection of deforestation is currently realized by monitoring systems that mainly rely on near daily observations of the coarse resolution MODIS (Moderate Resolution Imaging Spectroradiometer) sensor. The near daily observations provide a satisfactory temporal coverage in the frequently cloud covered tropical region (Diniz et al., 2015, Shimabukuro et al., 2006, Hammer et al., 2009, Wheeler et al., 2014). Due to the coarse resolution of MODIS data of 250–500 m, however, many small scale changes are missed (Anderson et al., 2005, Hammer et al., 2014, Hansen and Loveland, 2012). Hansen and Loveland (2012) demonstrated that MODIS misses up to 50% of the forest changes when compared to medium resolution Landsat data (30 m).

The opening of the Landsat archive in combination with the ability to download fully pre-processed images (Wulder et al., 2016), stimulated a shift of efforts in recent years towards operational and large area Landsat based deforestation monitoring at annual (Hansen et al., 2013, Souza et al., 2013) and sub-annual scales (Hansen et al., 2016). Several optical time series approaches with NRT capabilities have been developed to exploit the entire temporal detail of Landsat data (Verbesselt et al., 2012, Xin et al., 2013, Zhu et al., 2012, Hansen et al., 2016). Hansen et al. (2016) demonstrated the potential for and constraints of operational Landsat based deforestation alerts for the humid tropics. The major limitation of Landsat-based tropical deforestation alerts, is the limited availability of cloud-free observations (Hansen et al., 2016, Souza et al., 2013, Sannier et al., 2014). In particular, cloud-free observations are rare during the wet season. Other regions, such as the Peruvian cloud forests for example (Hansen et al., 2016), suffer from pervasive cloud cover throughout the entire year. In extreme cases, Landsat data gaps remain for more than one year (Hansen et al., 2016, Potapov et al., 2012, Sannier et al., 2014). In summary, a monitoring system that relies only on medium resolution optical data, will not provide a sufficient number of cloud-free observations throughout all seasons and geographical locations. A reduced number of cloud-free observations results in delayed detection of new deforestation events (Reiche et al., 2015b).

Synthetic aperture radar (SAR) can penetrate through clouds, and therefore has the potential to complement optical-based forest monitoring systems (De Sy et al., 2012, Joshi et al., 2016, Vaglio Laurin et al., 2013). To monitor tropical deforestation at larger scales, mainly long wavelength L-band SAR (~ 23.5 cm) data has been utilized (Shimada et al., 2014, Whittle et al., 2012). Shorter wavelength C-band (~ 5.6 cm) SAR is generally less useful for forest change monitoring because of the lower penetration depth, and rapid saturation of the signal over forests (Woodhouse, 2005). Fragmented and inconsistent data acquisitions and/or commercial data distribution of key SAR missions in the past have hampered their operational application, and have limited opportunities to integrate optical and SAR data (Reiche et al., 2016). With the launch of the Sentinel-1A and -1B C-band SAR satellites in 2014 and 2016 (Torres et al., 2012), for the first time, dense SAR time series are free and openly available for the tropical region. Since temporal sampling frequency is key for NRT deforestation monitoring, Sentinel-1 could be a milestone when it comes to more precise tracking of forest change events and activities; a potential that is yet to be explored. A key question is the extent to which the high temporal observation density of Sentinel-1 C-band SAR can compensate for the lower sensitivity to detect deforestation, when compared to longer-wavelength L-band SAR observations. L-band SAR data are currently available from the ALOS-2 PALSAR-2 mission (launched 2014, Rosenqvist et al., 2014), but only a few images per year are available for most tropical regions and the commercial data distribution limits the operational uptake. However, freely accessible L-band data are in sight with the upcoming SAOCOM-1 (Satélite Argentino de Observación Con Microondas; planned launch 2017) and NISAR (NASA/ISRO Synthetic Aperture Radar; planned launch 2020) missions.

Space agencies and international organisations, i.e. the Global Forest Observation Initiative (GFOI, 2015), emphasize the need and potential for joint exploitation of the impeding stream of free-of-charge SAR and optical data streams to improve tropical forest monitoring (Reiche et al., 2016). Using dense Sentinel-1 time series for NRT deforestation monitoring could provide a significant step forward in its own right. Combining time series from multiple optical and SAR sensors, however, has the potential to improve the robustness of NRT deforestation monitoring at medium resolution scale by increasing the number of available observations and ensuring a minimum number of observations for all seasons and geographical locations. In recent years several studies have demonstrated the increase of spatial, and temporal accuracy of deforestation detection, when combining optical and SAR time series (Lehmann et al., 2012, Lehmann et al., 2015, Reiche et al., 2013, Reiche et al., 2015a, Reiche et al., 2015b). These studies have developed some of the methodological underpinnings for the combination of these datasets using probabilistic approaches (Lehmann et al., 2012, Lehmann et al., 2015, Reiche et al., 2015a), and focused on combining Landsat and ALOS PALSAR L-band SAR data at small test sites and humid tropical forest conditions. The expansion of these methods, however, to dry tropical forest conditions, to combine more than two sensors, and the consideration of new dense Sentinel-1 C-band SAR data, is outstanding.

Dry tropical forest accounts for ~ 40% of the total tropical forest area (Murphy, 1986) and had the highest deforestation rates in the past 15 years (Hansen et al., 2013). Detecting deforestation in satellite image time series in dry tropical forest conditions requires the removal of the seasonal forest component. Otherwise, the seasonal variations may lead to substantial false detection of deforestation (Hamunyela et al., 2016b). While the seasonal forest component in the optical time series signal is driven by changes in phenology (Tucker, 1979), changes in canopy structure and moisture cause seasonal variations in the SAR backscatter signal (Ulaby et al., 1986). The majority of time series approaches propose seasonal model fitting to account for forest seasonality (Verbesselt et al., 2010, Zhu et al., 2012). Robust model fitting requires sufficient historical observations. Due to cloud cover, historical Landsat time series in tropical regions are often too sparse for robust fitting of a seasonal model. The same applies for ALOS PALSAR-1/2 L-band SAR time series. Hamunyela et al. (2016b) proposed spatial normalisation to reduce seasonality in Landsat time series, and successfully applied it to a dry tropical forest. Each pixel is normalised with the value of dense forest in the spatial surrounding of the pixel to be normalised. The application of spatial normalisation could also be applied in SAR time series to reduce seasonal variability; a potential that has yet to be demonstrated.

Assessing the temporal accuracy alongside the spatial accuracy gets increasingly important when evaluating NRT monitoring systems. While consolidated guidelines and methods exist for estimating the spatial accuracy of detected change (Foody, 2002, Olofsson et al., 2013, Olofsson et al., 2014, Stehman, 2009), assessing the temporal accuracy is challenging due to the lack of temporally dense reference information. As a result, reference data are commonly derived from the satellite time series itself using visual image interpretation (Cohen et al., 2010, Zhu et al., 2012, DeVries et al., 2015). Commonly, the date at which the change is first visible in the image time series is used as reference data to calculate the temporal accuracy or temporal detection delay (Zhu and Woodcock, 2014a, Pratihast et al., 2015, Hansen et al., 2016, Reiche et al., 2015b, DeVries et al., 2015). This approach results in time-biased reference data. The bias is related to the fact that true date of deforestation can occur at any date between the date of the image at which the change is first visible (commonly considered as reference date), and the date of the previous image in the time series. This imprecision is variable in time and space, and becomes even larger when relying on sparse and/or irregular time series, such as Landsat in some tropical regions. Since NRT monitoring is aiming for temporal accuracies in the order of weeks or days, imprecision in the reference data becomes critical when assessing the performance. Despite the fundamental importance, the consideration of the temporal precision in the reference data and the impact for users has not been thoroughly studied for forest change alerting.

In this paper we address some key challenges for multi-sensor NRT deforestation monitoring building on newly available dense Sentinel-1 time series data in combination with ALOS-2 PALSAR-2 and Landsat. We aim to:

  • (i)

    Demonstrate how spatial normalisation reduces dry forest seasonality in Sentinel-1, ALOS-2 PALSAR-2 and Landsat time series in order to combine them for NRT deforestation detection using a probabilistic approach. We put particular emphasis on how environmental effects and dry forest seasonality affect the dense Sentinel-1 C-band SAR time series signal when compared to Landsat NDVI and PALSAR-2 L-band SAR.

  • (ii)

    Compare the spatial and temporal accuracy of single-sensor Sentinel-1, PALSAR-2 and Landsat results versus multi-sensor results. We take the precision of the reference data into account to provide a more robust estimate of how quickly a change event happening on the ground can be detected.

  • (iii)

    Evaluate the impact of uncertainties for different user scenarios, by taking a critical look at the trade-off between spatial and temporal accuracy associated with NRT deforestation monitoring.

With this study we highlight the unprecedented potential of dense Sentinel-1 time series and its combination with other optical and SAR sensors to improve the robustness of NRT deforestation monitoring in dry tropical forest environment.

Section snippets

Study area

The study was conducted at a dry tropical forest site (10,000 km2), located in the southeast of the province of Santa Cruz, Bolivia (centred at Lat. 18.39″S, Lon. 62.36″W) (Fig. 1). Being one of the wettest regions of Bolivia, this area is characterized by a humid tropical climate with distinct wet (~ October–May) and dry seasons (~ June–September). The change from wet and dry seasons is associated with a strong change in photosynthetic activity of the forest. Deforestation in the area is mainly

Data and methods

An overview of the methods used in this study is shown in Fig. 2. We acquired Sentinel-1 VV-polarised C-band SAR, ALOS-2 PALSAR-2 HV-polarised L-band SAR and Landsat (7/ETM+ and 8/OLI) NDVI time series data for the two year period between 2014/10/01 and 2016/09/30; corresponding to the first two years of available Sentinel-1 images. Data from the first year was used to derive training information (training period), and data from the second year was used to detect new deforestation events

Spatial normalisation and derived F and NF pdfs

Fig. 6 depicts the F and NF distributions overlaid with fitted pdfs separately for the original and spatially normalised time series: S1VV (A1) and S1VVn (A2), P2HV (B1) and P2HVn (B2), LNDVI (C1) and LNDVIn (C2). For the original time series we found LNDVI to have the weakest F/NF class separability (JM = 0.66) when compared to S1VV (JM = 1.14) and P2HV (JM = 1.91). The bimodal LNDVI class distribution for F corresponds to wet and dry season observations, whereby the dry season observations largely

Discussion

In this paper we presented the first study on multi-senor SAR-optical NRT deforestation detection in a tropical dry forest, and we combined newly available dense Sentinel-1 time series with ALOS-2 PALSAR-2 and Landsat 7/ETM+ and 8/OLI data. We demonstrated that spatial normalisation can be used to reduce dry forest seasonality in the SAR time series in order to combine them with optical time series for NRT deforestation detection using a probabilistic approach. Our results for a dry tropical

Conclusion

We demonstrated multi-sensor SAR-optical NRT deforestation detection in a tropical dry forest, and we combined time series observations from Sentinel-1, ALOS-2 PALSAR-2 and Landsat 7/ETM+ and 8/OLI. We successfully applied spatial normalisation to reduce the dry forest seasonality in the SAR and optical time series, and combined them using a probabilistic approach. Our results show that deforestation events were detected with higher spatial and temporal accuracies when combining observations

Acknowledgments

The work of Johannes Reiche and Martin Herold was funded by the European Commission Horizon 2020 BACI project (grant agreement 640176). The work of Eliakim Hamunyela and Jan Verbesselt was funded by the European Space Agency (ESA) ForMoSa project — Forest Degradation Monitoring with Satellite Data Project (grant agreement 5160957022). The research has been supported with funding from the European Commission Horizon 2020 LandSense project (grant agreement 689812). This work was supported by the

References (65)

  • T. Motohka et al.

    Using time series PALSAR gamma nought mosaics for automatic detection of tropical deforestation: a test study in Riau, Indonesia

    Remote Sens. Environ.

    (2014)
  • P. Olofsson et al.

    Making better use of accuracy data in land change studies: estimating accuracy and area and quantifying uncertainty using stratified estimation

    Remote Sens. Environ.

    (2013)
  • P. Olofsson et al.

    Good practices for estimating area and assessing accuracy of land change

    Remote Sens. Environ.

    (2014)
  • P.V. Potapov et al.

    Quantifying forest cover loss in Democratic Republic of the Congo, 2000–2010, with Landsat ETM + data

    Remote Sens. Environ.

    (2012)
  • J. Reiche et al.

    Fusing Landsat and SAR time series to detect deforestation in the tropics

    Remote Sens. Environ.

    (2015)
  • A. Rosenqvist et al.

    Operational performance of the ALOS global systematic acquisition strategy and observation plans for ALOS-2 PALSAR-2

    Remote Sens. Environ.

    (2014)
  • C. Sannier et al.

    Using the regression estimator with Landsat data to estimate proportion forest cover and net proportion deforestation in Gabon

    Remote Sens. Environ.

    (2014)
  • M. Shimada et al.

    New global forest/non-forest maps from ALOS PALSAR data (2007–2010)

    Remote Sens. Environ.

    (2014)
  • R.B. Thapa et al.

    The tropical forest in south east Asia: monitoring and scenario modeling using synthetic aperture radar data

    Appl. Geogr.

    (2013)
  • R. Torres et al.

    GMES Sentinel-1 mission

    Remote Sens. Environ.

    (2012)
  • C.J. Tucker

    Red and photographic infrared linear combinations for monitoring vegetation

    Remote Sens. Environ.

    (1979)
  • G. Vaglio Laurin et al.

    Optical and SAR sensor synergies for forest and land cover mapping in a tropical site in West Africa

    Int. J. Appl. Earth Obs. Geoinf.

    (2013)
  • J. Verbesselt et al.

    Detecting trend and seasonal changes in satellite image time series

    Remote Sens. Environ.

    (2010)
  • J. Verbesselt et al.

    Near real-time disturbance detection using satellite image time series

    Remote Sens. Environ.

    (2012)
  • M. Whittle et al.

    Detection of tropical deforestation using ALOS-PALSAR: a Sumatran case study

    Remote Sens. Environ.

    (2012)
  • M.A. Wulder et al.

    The global Landsat archive: status, consolidation, and direction

    Remote Sens. Environ.

    (2016)
  • Q. Xin et al.

    Toward near real-time monitoring of forest disturbance by fusion of MODIS and Landsat data

    Remote Sens. Environ.

    (2013)
  • Z. Zhu et al.

    Automated cloud, cloud shadow, and snow detection in multitemporal Landsat data: an algorithm designed specifically for monitoring land cover change

    Remote Sens. Environ.

    (2014)
  • Z. Zhu et al.

    Continuous change detection and classification of land cover using all available Landsat data

    Remote Sens. Environ.

    (2014)
  • Z. Zhu et al.

    Continuous monitoring of forest disturbance using all available Landsat imagery

    Remote Sens. Environ.

    (2012)
  • L.O. Anderson et al.

    Assessment of deforestation in near real time over the Brazilian Amazon using multitemporal fraction images derived from Terra MODIS

    IEEE Geosci. Remote Sens. Lett.

    (2005)
  • J. Assunção et al.

    DETERring Deforestation in the Brazilian Amazon: Environmental Monitoring and Law Enforcement

    (2013)
  • Cited by (180)

    View all citing articles on Scopus
    View full text