Measurement and monitoring needs, capabilities and potential for addressing reduced emissions from deforestation and forest degradation under REDD+

Remotely sensed data enable mapping and monitoring of forest cover and change over large areas at regular intervals, providing information on where and how changes are taking place at annual or even finer time scales. Satellite remote sensing provides the capability to measure and monitor changes taking place at relatively fine spatial resolution (<1 ha) over very large areas (regions to continents). Moreover, because satellite-based sensors such as Landsat (owned and operated by the US Geological Survey and NASA) collect data on a daily basis as they circle the Earth, and because the data are typically distributed at minimal to no cost to users, monitoring of forest clearing can be implemented consistently and accurately across large regions as compared to a much smaller portion of the landscape that can be monitored as part of traditional forest inventory efforts. For example, 30 years ago (in the mid-1980s), it would cost a researcher $5000 to obtain a single Landsat image (covering an area 170 km by 183 km). A country the size of Brazil requires at least 20 such images to entirely cover its territory, even for a single acquisition timeframe, and many times that number may be needed to obtain cloud-free coverage. Now, anyone with a fast internet connection can obtain these data for free, and processing time has been reduced dramatically as a result of advances in computing and data serving.

The Landsat program provides a robust data set for measuring and monitoring deforestation. Three developments have facilitated dramatic advances in measuring and monitoring deforestation from Landsat data in recent years. These are: (i) opening of the Landsat archive in 2008 making the data freely available to and downloadable by users, (ii) improved computing capabilities enabling rapid processing of large data volumes, and (iii) development of mapping methodologies designed to take advantage of these large volumes. Enabling the use of the entire Landsat database has allowed users to largely overcome issues of cloud cover and other factors that previously resulted in missing coverage. Landsat is in many ways ideal for measuring deforestation because unlike other Earth observation systems it: (i) has a global acquisition strategy, meaning every part of the land surface is imaged systematically over time, (ii) makes data freely available through the USGS, (iii) provides data access through multiple delivery portals, (iv) ensures high fidelity image characteristics, meaning data are of consistently high quality and (v) provides pre-processed data to enhance usability, including geometric and radiometric corrections applied to all archived images. No other system currently has these features, which have been implemented with Landsat since 1999. Data from 1972 to 1999 lack the systematic global acquisition strategy but have all other features.

A number of satellite-based national programs for monitoring forest cover loss are already operational. The standard to date for national monitoring is Brazil's Amazon Deforestation Monitoring Project (PRODES) and associated data produced and distributed by the Brazilian National Institute for Space Research (INPE). INPE provides annual updates of deforestation for the states of the legal Amazon, including spatially detailed digital maps. While other countries also have satellite-based deforestation mapping programs (e.g. Australia and India), they do not provide map products to the public as INPE does. We expect more countries will emulate Brazil in delivering timely and publicly available information on historic and current deforestation patterns and rates.

The continuity of the satellite observational systems into the future will enable the quantification of longer-term trends in deforestation, a requirement for carbon measurement, monitoring and reporting relative to baselines⁶ (GFOI 2013, GOFC-GOLD 2014, Hewson et al 2014). Having a consistent set of measurements is critical to establishing trends in forest cover over time. Because of the advances noted above, an unprecedented global map of forest losses and gains, based on changes in tree cover density, was developed from Landsat data for the period 2000–2012 (Hansen et al 2013) (figure 1). This data set can be visualized by all interested parties via a Google Earth Engine web platform⁷ . The data products are also included in a web platform called Global Forest Watch (globalforestwatch.org), which makes the data available to anyone in a user-friendly format, thereby advancing transparency across the entire user base and ensuring users of the data can aid development of improved data products through time. The continued availability of forest monitoring capabilities like those conveyed by Landsat are critical for a wide range of Earth science and resource management applications, including REDD+. However, it is important to note that the maintenance and replacement of existing sensors is not guaranteed so continuous efforts are needed to ensure operational funding. Broad-based awareness of the value of Landsat and other satellite missions, and associated political support, will help to ensure that the line of continuous and consistent earth observation remains unbroken. Moreover, the implementation of earth observation sensor systems with similar and expanded capabilities and data policies is needed. For example, Brazil's open and free data policy will be made more valuable to the global community as it extends the coverage of its China-Brazil Earth Resources Satellite data acquisitions⁸ , and the European Space Agency's Sentinel series of satellites has a planned data access policy similar to that of the Landsat system⁹ . Moving such land imaging systems into operational modes will ensure the accurate, consistent and long-term monitoring of global forest resources.

The second key component required for tracking emissions associated with the conversion of forest land to other land uses is the measurement of emission factors, or the carbon stocks associated with areas that have undergone change. Regardless of the vegetation type, about half of its aboveground woody biomass consists of carbon, and is thus commonly referred to as the vegetation aboveground carbon density or carbon stock. Information on carbon stocks is not only important for REDD+ but it is also commonly used in the modeling of global carbon budgets (see www.globalcarbonproject.org), providing valuable information on the potential for carbon sequestration from the atmosphere via plant growth (i.e. through net photosynthesis).

All tropical forests are not equal with respect to their potential to mitigate emissions from deforestation and degradation. In the context of REDD+, data on the distribution of carbon stock across landscapes is critically important because carbon stock density can be quite uneven and patchy. If deforestation takes place in an area of high carbon stock, more carbon will be emitted to the atmosphere, all else being equal. A comparably sized area of deforestation in a low carbon stock area will, conversely, emit less carbon to the atmosphere. This is important when considering how extensively to distribute field sites in order to adequately measure changes in emissions from deforestation or forest degradation. Many field plots have historically been located in sites that were not altered or disturbed by human activity, in so-called primary forests, and so were not necessarily representative of many landscapes, particularly in the tropics where disturbance is commonplace. Moreover, because field measurements are time consuming and thus relatively expensive, it is difficult to achieve sampling densities that can adequately capture local spatial variability across diverse landscapes and large tropical countries such as Brazil, Indonesia or the Democratic Republic of the Congo. Nonetheless, field data are essential for providing a basis to extend local measurements over larger areas, and remote sensing provides the means to do so. In addition to allowing for the calibration of aircraft or satellite observations, subsets of field measurements held in reserve are necessary for validating (i.e. assessing the accuracy) of carbon stock maps derived from the combination of field and remote sensing measurements.

2.1.2.1. Aboveground carbon stocks/emission factors

2.1.2.2. Synthesis of field plots and national forest inventories with cover type classification

2.1.2.3. Synthesis of field plots and aircraft remote sensing measurements

2.1.2.4. Synthesis of field plots and satellite measurements for monitoring

Finally, maps of carbon stock can be estimated over very large areas (e.g national to global scales) by calibrating satellite measurements with estimates of carbon stock density from field plots. Estimating the carbon content of forests and its dynamics using aircraft has driven a desire for a space-based LiDAR mission appropriate for carbon stock mapping and monitoring. New carbon stock maps have been developed at sub-kilometer resolution using a combination of LiDAR data from an instrument designed for ice sheet monitoring (the Geoscience Laser Altimetry System, GLAS, onboard a satellite called ICESAT-1) and imaging satellite observations from MODerate spatial resolution Imaging Spectroradiometer (the MODIS). These maps have been produced for large areas of Canada (Boudreau et al 2008), Siberia (Neigh et al 2013) and even the entire tropical region (Saatchi et al 2011, Baccini and Asner 2013) (see figure 2). Reported errors on the plot data used to develop the models for these maps were on the order of 5% whereas individual location errors in the derived maps are typically higher. When aggregated nationally or to a continental scale, errors are again reduced via the effect of spatial averaging to around 5%. Once carbon stock maps exist they can be readily combined with activity data using either sampling (e.g. Achard et al 2014) or wall to wall approaches (e.g. Baccini et al 2012, Harris et al 2012) to estimate and spatially map emissions. Several group efforts are now underway to extend the synthetic approach to mapping carbon stock at higher spatial resolution, through time, across the globe. Radiowave Detection and Ranging (RaDAR) imagery can assist this effort as well because, unlike optical imagery or LiDAR, it has the unique ability to penetrate cloud cover and can be acquired either during the day or night.

Zoom In Zoom Out Reset image size

A significant advance in monitoring of forest carbon stocks between 50° North andSouth latitude will be possible in the next few years with the installation of a LiDAR instrument on the International Space Station (ISS). Development of the mission, called Global Ecosystem Dynamics Investigation (the GEDI) was initiated in July 2014. The ISS orbits around the equator and GEDI will acquire extensive data sampling (some 15 billion locations within a year of operation) that will allow derivation of carbon stock maps across large areas (e.g. regions, continents, the pantropical domain) at higher resolution and lower error than is currently possible. This capability will also vastly improve the potential for assessing changes in carbon stocks across tropical forest areas. Other upcoming missions advancing remote sensing technological capability are summarized in section 5.

2.1.2.5. Belowground carbon stocks/emission factors

Calculating emissions from degradation is more difficult than calculating emissions from deforestation for two reasons: estimates of biomass density at multiple points in time are required, and carbon sequestration due to regrowth must be considered. As with estimating emissions from deforestation, described in section 2.1, both forest area (extent) and carbon stock density are needed to capture the components of forest degradation emissions to the atmosphere. Unlike the case of deforestation, however, emissions from forest degradation are more uncertain for two reasons. First, to estimate the emissions from degradation, multiple estimates of biomass density (before and after degradation) must be obtained, from which the difference is an estimate of the committed emissions. In contrast, emissions from deforestation require only a single estimate of biomass density. Degradation may also be more difficult to measure in the field than deforestation because, unlike deforestation, which is usually a short-term conversion from forest to non-forest, degradation can take place gradually over years.

The second issue that adds to the complexity of estimating the emissions from degradation is the possibility that degrading processes (logging, burning, grazing) may be followed by recovery. When wood is selectively harvested from a forest, for example, the carbon density of the forest is reduced as a result. But if the land is not converted to another use, the harvested forest will accumulate carbon again as the forest grows back. Does the accounting for carbon in forest degradation include only the gross emissions, or does it include the net effect of loss and recovery? If one is interested in only the gross emissions of carbon from degradation, then the recovery is irrelevant. But if one is interested in the net emissions, and if a harvested forest returns to its initial carbon stock density, the net effect over the longer term is zero. Fully accounting for stock changes requires repeated observations over a time scale of decades. We consider the issue of regrowth in the section 2.3, but it is important to acknowledge its relevance to the measurement of degradation as well.

Imagery with higher resolution than Landsat can be used to map local losses of tree cover that result from various types of degradation. For example, logging roads can be readily mapped using Landsat, but degradation associated with canopy gaps of felled trees and skid trails where trees are dragged to roads typically requires imagery of higher resolution (on the order of a few meters). The area affected by these canopy disturbances can accumulate to be many times larger than disturbances associated with tree felling alone (Laporte et al 2007) (figure 3). This is important in the context of REDD+ in order to assess the fuller impacts of degradation on canopy cover and biomass. Use of high resolution (e.g. 5–10 m) optical satellite imagery (such as IKONOS or Quickbird) is helpful for delineating the affected areas, and is a low-technology solution since such imagery can be visually inspected and hand-digitized.

Zoom In Zoom Out Reset image size

A related variation of this approach is informing finer scale assessments based on moderate resolution imagery (such as Landsat). In the case of selective logging, which targets specific tree species of high value and large size (e.g. African mahogany), loss of tree cover (and diversity) may be small relative to the biomass removal, but canopy cover removal can be substantial and therefore detectable with Landsat (Souza et al 2005). Other types of degradation, such as fuel wood removal, also require higher resolution imagery to detect. As with the case of deforestation, attributing the changes from such land uses and management requires field observations in combination with remote sensing. Currently, the detection of lower levels of carbon stock degradation (<30% biomass loss) using remote sensing is difficult unless based on a combination of field and LiDAR data sets, which means they are currently restricted to relatively small areas where aircraft LiDAR are typically flown (hundreds of km²).

Consistently and accurately mapping most types of forest degradation at regional scales, particularly over short time intervals, benefits from the use of nested multi-scale imagery. This is because the ability to accurately map the area and associated change in carbon stock associated with degradation is, as described earlier, a function of the time period over which the various types of degradation take place. A multi-scale approach, i.e. making use of several data sources at variable resolution, will allow one to increase the ability to detect degradation-related changes in forest cover, while also increasing the likelihood of detecting degradation over shorter time intervals. Like any other approach, this is most effective and accurate when accompanied by appropriate validation efforts and relatively frequent (e.g. annual) repeat measurements. A good example of this type of nested multi-scale approach, making use of a wide range of image data sources, is that being conducted in Guyana as part of a bilateral agreement with Norway¹¹ , which has robust MRV procedures in place.

A quicker and easier but less accurate method to estimate degradation over larger areas is to extrapolate local-scale assessments of degradation using approaches that incorporate more indirect indicators of degradation, such as buffers around or distances from villages and roads (including logging roads) or proximity to previous forest clearing (Potapov et al 2008, Herold et al 2011, GOFC-GOLD 2014). For example, a reasonable estimate of fuelwood removal from a forest area surrounding a group of villages could be obtained by assuming that degradation declines at a constant rate as forests are increasingly remote from villages. Assessing errors in such methods is more difficult than assessing errors based on more direct approaches, but indirect approaches can help approximate degradation and plan more targeted observations (such as high resolution imagery or LiDAR acquisition).

Recent work indicates that a combination of indirect proxies and field observations of the area and carbon stock of degraded lands can provide assessments of emissions associated with forest degradation from both selective logging and shifting agriculture. For example, Pearson et al (2014) used field measurements of changes in carbon stocks in selectively logged areas to estimate the carbon emissions from three different components of wood harvest: the extracted log, incidental damage to surrounding trees, and infrastructure (e.g. logging roads). In contrast, Margono et al (2012) and Zhuravleva et al (2013) linked degradation to remote sensing in the Democratic Republic of the Congo by using Landsat imagery and LiDAR data to quantify structural differences between primary intact, primary degraded and other forests. While these two approaches to estimating forest biomass loss (Pearson et al 2014) and forest cover degradation (Margono et al 2012, Zhuravleva et al 2013) advance what can be done using indirect proxies, extending field measurements of degradation to a national scale benefits from additional, more direct links with remote sensing of biomass losses (and gains) such as those described in the previous section 2.2.2.

Summary: degradation can be assessed using approaches that capture change in canopy cover coupled with measurements of carbon stocks. In this sense measurement of degradation is similar to that of deforestation, but degradation is complicated by biomass gains that can partially or wholly offset biomass losses in the same forest. To simplify this complexity for purposes of REDD+ accounting, one can measure net degradation (including both losses and accumulations) of carbon stock at repeated time intervals. The accuracy of doing so depends on the combined accuracy of field measurements of aboveground stocks, and the sensitivity of remote sensing to the density of those aboveground stocks. Detecting the net effect of gains and losses within areas that remain forest thus benefits from use of high resolution imagery, and can be effectively augmented with field measurements of carbon stock changes as well as the biomass sensitivity conferred by LiDAR measurements. Related approaches make use of proxies of degradation, such as the area of forests that have been disturbed relative to areas that have no record of disturbance in the satellite record.

Given rapid growth in the commercial oil palm industry across the tropics, particularly in Indonesia, the vast majority of remote sensing research to address the monitoring of plantation forests has focused on the conversion of natural forest to oil palm plantations. Oil palm thrives in regions otherwise occupied by dense humid tropical forest, which means regions that are also frequently obscured by heavy cloud cover. Hence, sensors such as Landsat (30 m) and MODIS (250 + m), which provide for relatively frequent image acquisitions (e.g. nominally every 16 days for Landsat and daily for MODIS) and so have a greater likelihood of capturing cloud-free imagery, are useful for one-time mapping as well as longer-term change monitoring. For example, Carlson et al (2013) used visual interpretation of Landsat time-series data to assess carbon emissions from oil palm plantations in Kalimantan, Indonesia, over a 20 year period (1990–2010). Where image coverage overlapped, oil palm detected with Landsat was confirmed using high-resolution optical data available in Google Earth and/or RaDAR imagery, in this case from the Japanese Advanced Land Observing Satellite (ALOS).

In another recent example, Gutiérrez-Vélez and DeFries (2013) assessed the tradeoffs between spatial and temporal resolutions among different satellite sensors to map oil palm expansion from 2001 to 2010 for the Peruvian Amazon. This analysis automatically classified MODIS time-series data to map annual forest conversion to oil palm for events larger than 50 ha with an accuracy of 73% at 250 m resolution. Total forest conversion to oil palm (939 204 km²) was subsequently calculated in annual time steps based on visual delineation of Landsat images. Analyses were also conducted using a combination of Landsat and ALOS RaDAR data (both at 30 m resolution) within a subset of the area (2158 km²) for conversions larger than 5 ha with an accuracy of 94%. The data and approaches used for these studies of oil palm are applicable to other plantation forests, such as eucalyptus.

As more high-resolution imagery is acquired and archived, monitoring the conversion of natural forests to plantations with such imagery can become more routine, as can larger area assessments of the kind now only practical with moderate-resolution imagery. A number of countries, such as Mexico, are systematically acquiring high-resolution imagery (e.g. 5 m resolution imagery) for the purpose of producing detailed classifications of land cover including the discrimination of natural versus managed lands (see figure 5). Because high resolution imagery can be expensive and quickly adds up to large volumes of data, using it routinely to map large areas (1000s of km²) requires the availability of significant computing and processing capabilities. With increasingly more powerful computers the former issue is generally being overcome; however, consistent data processing of large data volumes requires systems that can properly address cloud cover, calibration and many other issues that arise among adjacent scenes through time.

Zoom In Zoom Out Reset image size

As described above, these issues are more easily overcome with moderate resolution imagery (i.e. Landsat and MODIS), but are increasingly available for application to higher resolution data sets like those being used in Mexico and elsewhere. With the advent of constellations of 'micro-satellites', including the recent acquisition of Skybox by Google (www.skyboximaging.com) and the launch of Planet Labs (www.planet.com), these challenges are increasingly likely to be overcome in the near term (next few years) rather than decades from now. Moreover, taking advantage of platforms like GoogleEarth allows users in many parts of the world with limited computational infrastructure to conduct their own interpretations using high resolution imagery, as long as they have reliable internet access.

Summary: using multi-sensor data sets that combine the advantages of each data source allows accurate discrimination of oil palm plantations from natural forests, and thus the ability to monitor the extent of plantations and the associated conversion of natural forests to plantations. Because many oil palm plantations are large (hundreds of hectares), moderate resolution imagery like that of Landsat and even coarser resolution MODIS is often adequate for mapping and monitoring plantation forests relative to natural forests, as well as the conversion of the latter to the former. Higher resolution imagery is becoming increasingly more common, with some countries acquiring 'wall to wall' mapping of their forested lands, and these data provide additional accuracy in mapping plantations and the conversion of natural forests. While high resolution image analysis of large areas requires greater data storage, computing and processing capacity to ensure continuity of mapping efforts, these capabilities are rapidly advancing. Also, such imagery will continue to become more readily accessible via web platforms such as GoogleEarth, and can be relatively easily interpreted visually by trained technicians, so use of this technology is accessible to countries with varied levels of technical capacity—particularly for more local mapping efforts.

Characterizing the tree species diversity of forests, understood here as the number and distribution of individual species or assemblages of species, was considered beyond the capability of remote sensing technologies until relatively recently. Tropical rainforests present a particularly unique challenge given the large number of species per unit area together with high structural complexity (e.g. number of different shapes and sizes of trees) and spatial heterogeneity (i.e. number of species within tens of square meters or less). Recent advances in remote sensing of tree reflectance spectra combined with canopy structure from lidar has changed our view of what is possible in terms of mapping aspects of biodiversity (e.g. Leutner et al 2012). As with approaches to measuring and monitoring forest cover and carbon stocks, these approaches to mapping patterns of forest traits and canopy chemistry can be divided into so-called direct approaches capable of detecting the individual tree diversity or communities themselves, and indirect approaches that depend on proxies such as environmental characteristics (i.e. aspect, elevation, soil type, etc) or physical structural features (e.g. tree height, crown size, etc).

Technologies to measure and monitor forest biodiversity in tropical ecosystems have advanced rapidly as a result of the evolution of new sensor classes including LiDAR and hyperspectral remote sensing, with the latter also known as imaging spectroscopy. Whereas LiDAR is capable of characterizing forest biodiversity based on detailed measurements of canopy three-dimensional structure (Bergen et al 2009), imaging spectroscopy sensors are capable of sensing the canopy's chemical and physiological properties or 'fingerprints' of plant spectra (Asner et al 2015). Imaging spectroscopy is currently operational onboard both airborne and satellite remote sensing platforms. However, because satellite sensors provide only moderate resolution data (Hyperion: 30 m; CHRIS: 17 m) and both air- and satellite acquisitions are generally only available to researchers on a geographically limited basis, the overall utility of these data for detailed species-level mapping across large areas remains limited. As a result, and because a number of studies have successfully demonstrated the potential of imaging spectroscopy in a variety of environments (Schimel et al 2015), there are ongoing efforts to advance our capability to distinguish and map forest canopy chemistry and trait diversity from a spaceborne platform at high resolution.

Being able to more directly map the distribution patterns of animal species diversity would provide additional REDD+ benefits of ensuring biodiversity conservation (e.g. of threatened and endangered species). Animal diversity distributions have been estimated using field surveys, often rapid assessments, linked with vegetation type maps, climate and other environmental variables to extrapolate over larger areas. Valuable global biodiversity data sets distributed by the IUCN, UNEP-WCMC and Birdlife International, among others, are based on this approach. In addition, one of the best-known proxies for forest species diversity is canopy habitat and three-dimensional canopy structural diversity (i.e. the horizontal and vertical distribution of canopy elements), particularly for arboreal species like birds and primates (Davies and Asner 2014). Canopy structure data sets have not been widely available before but with the advent of LiDAR remote sensing, these data are now much more widely and consistently available than was possible from limited field studies. A number of studies have shown the utility of airborne LiDAR data for predicting not only bird richness (Vierling et al 2008) but also individual species preferences and competition for specific habitats and nesting locations (e.g. Goetz et al 2010). Thus characterizing multi-dimensional habitat heterogeneity with remote sensing, particularly LiDAR coupled with field observations, is extremely useful for identifying and mapping biodiversity patterns.

Summary: the state-of-the-art in species-level mapping and monitoring of subtropical and tropical forest ecosystems combines the spectral sensitivity of imaging spectroscopy with the canopy structural sensitivity of LiDAR measurements. This combination provides for unprecedented discriminating power in three dimensions, improving upon the information that either sensor type is capable of providing alone, and thereby also providing capability to map tree composition, animal species richness, habitat diversity and use, and changes in these ecosystem attributes through time. These data will not be available everywhere but they provide valuable information on the biodiversity of forests that is useful for REDD+ safeguards by augmenting what is currently available more generally around the globe.

Because REDD+ is a pay-for-performance mechanism, it is important to distinguish between anthropogenic (i.e. of human origin) and non-anthropogenic (i.e. natural) causes of deforestation. This allows agencies in forest countries to penalize (or reward) the emission (or sequestration) of carbon resulting from management activities but not those resulting from natural processes. In the tropics, most clearing is the result of human action on the landscape, but there are also natural changes resulting from wildfires, wind storms or other causes. A recent example attributing deforestation from natural versus anthropogenic causes in Peru (Potapov et al 2014) nicely demonstrates the potential of distinguishing these different causes using remote sensing, as well as the value of close partnerships between the policy and research communities (see also Pelletier and Goetz 2015).

Natural and human causes of fire also interact, for example where degraded forests may have more fuel conducive to carrying fires initiated from a natural cause (like lightning). Proximate causes of human-induced forest loss include mechanical removal of vegetation and the intentional setting of fire, both of which are often employed in the clearing of tropical forests. Natural processes like drought also exacerbate the likelihood of fire. Thus identifying the direct cause of forest disturbance would improve carbon emissions estimations and their attribution to human or natural origins.

Remote sensing of fire disturbance is well advanced and routinely used in a number of temperate countries for fire management efforts. Aside from 'hotspot' detection of fires and burned area delineation (using MODIS and Landsat) (e.g. Justice et al 2002, Boschetti et al 2010), remote sensing of fire disturbance in the tropics has focused largely on direct fire event emissions into the atmosphere (Giglio et al 2013) or the proximity of fire events to populated areas and roads (Kumar et al 2014). Using frequent repeat remote sensing can often assist in identifying the sources of different disturbance types, given timber harvest, forest cutting for shifting agriculture, and other forest conversion and degradation processes have unique patterns on the landscape. However it is usually necessary to inform remote sensing of these patterns with additional ancillary information, such as the indirect methods discussed in sections 2.2.3 and 3.1.

To the extent it is possible to distinguish the land use to which previously forested land is converted, attribution to particular land uses, owners or industries can be made. As noted in the previous section, it is often possible to match patterns of deforestation with obvious indication of human activity, such as extension of roads into the affected areas. When available, data on land ownership and land use rights can provide additional information on the probability of accurate attribution to human drivers as agents of change. A nice example of this approach is Killeen et al (2008) documenting land transformations by different social and economic groups in Bolivia over multiple decades.

Land use in recently deforested or disturbed areas can also be discerned by statistically relating (regressing) the patterns and the texture of the land surface to areas where deforestation and degradation has already been attributed to human land use (e.g. pastoralists and shifting agriculture). Data from observational platforms such as Landsat are a prerequisite to these efforts.

Where multiple causal factors contribute to deforestation, statistical techniques can be used to assess their relative contributions and suggest promising interventions. These techniques, termed spatially explicit econometrics, analyze the relationship between spatial patterns of deforestation and maps of potential driver variables (Ferretti-Gallon and Busch 2014).

Sensors that can detect hotspots of forest-cover change activity on a frequent basis can be used to target law-enforcement actions. For example, INPE's DETER program detects, in near-real time, clearings within the Brazilian Amazon. DETER data have been used to support enforcement of land use regulations by prosecuting illegal deforestation. The system was launched in May 2004 and has been instrumental in allowing rapid reaction to signs of deforestation and thereby reducing emissions associated with such activities (figure 6) (Arima et al 2014). Monthly alerts are generated by INPE and sent to the EAs at the federal and state level, including IBAMA (the Brazilian Institute of Environment and Renewable Natural Resources). Based on this information the EAs can define priority areas for action in the field as well as refine the process of distinguishing legal and illegal deforestation.

Zoom In Zoom Out Reset image size

Because the DETER data and maps are made publically available via the internet, the program has allowed other institutions to develop their own monitoring systems, such as System of Alerts on Deforestation (SAD) from IMAZON (Amazon Institute of People and the Environment). This transparency also decreases the potential for corruption since available data allow investigation of illegal activity by all interested parties. The DETER and SAD systems have the potential to be extended to the entire humid forests of the tropics as part of the FORest Monitoring for Action (FORMA) program, launched in 2014 but developed and tested for years earlier (at the Center for Global Development and the World Resources Institute). FORMA is now part of the Global Forest Watch effort (see section 2.1).

Brazil's DETER system, and the extension of it to other tropical forest areas via FORMA, are based on data from NASA's MODIS sensors, which acquire relatively coarse spatial resolution imagery but with high frequency (i.e. 250–1000 m data acquired daily). This high-temporal acquisition frequency provides the most rapid update of the land surface currently available while increasing the probability of cloud-free acquisitions. MODIS thereby provides the ability to detect deforestation associated with many human activities, such as the expansion of frontiers of deforestation, soy agriculture, and other activities.

The best use of MODIS for forest monitoring is as an alarm or indicator of larger 'hotspots' of forest loss rather than to calculate specific changes in forest area. This is because the moderate spatial resolution of MODIS is somewhat limited for observations of forest cover loss because forest disturbances often occur over areas that are smaller than MODIS pixels (i.e. sub-pixel scale). Landsat, RaDAR and other higher resolution data sources can then be used to establish specific measurements of forest loss in the affected areas. Using these systems together, along with other data sources, is a powerful means to enforce policies. Moreover, because MODIS acquires data more frequently than Landsat, it can be used to determine the timing of deforestation, which can be important in some cases for attributing change to human versus natural causes. Recently, an updated version of DETER (called DETER-B) was developed using higher resolution imagery from Landsat and the Brazilian AWIFs system, rather than the 250 m MODIS imagery (Diniz et al 2015). The updated system allows detection of areas smaller than 25 ha and has improved accuracy resulting in more efficient enforcement action.

Another set of technologies that can aid detection of illegal activity and enhance enforcement activities is in situ monitoring devices. These include motion-detection cameras and audio devices that can detect the sound of chain saws (from logging) and gun shots (from poaching). Motion activated cameras are now widely used for wildlife monitoring, but also have utility for detecting illegal logging, particularly when installed to image larger areas (e.g from high in a forest canopy). Audio detection and alert transmission systems have been in development for a number of years and some make use of proven technology that is widely available (e.g. wildlandsecurity.org). Related recent efforts make use of discarded cellular telephones for detecting audio signals and then transmitting alerts to enforcement authorities or conservation groups¹⁸ . These techniques are likely to become more readily available and so available for broader adoption.

Summary: the distinction between forest change attributable to human activity versus natural processes is advanced by systematic remote sensing of forest conversion though time, and by incorporating additional data sources to make attribution more reliable. It is often possible to infer attribution using remote sensing based on patterns across the landscape unique to timber harvest or plantations, as well as proximity to the expansion of roads and towns, coupled with local knowledge and field observations. Detection systems that make use of frequent satellite observations to identify hot-spots of change are also valuable for attribution and are used to alert EAs to deforestation and to deter illegal operations, with Brazil having the most advanced capabilities for this type of near real-time monitoring. Related capabilities that help with enforcement include in situ motion cameras and audio sensors that are triggered by potentially illegal activities. These detection and alert systems have had a measurable impact on reducing deforestation related emissions and associated illegal land conversion in Brazil. Similar systems can be implemented in other countries by learning from and building upon these advances.