International Journal of Applied Earth Observation and Geoinformation
Alerts of forest disturbance from MODIS imagery
Introduction
Deforestation and forest degradation contribute 12 percent of global anthropogenic greenhouse gas emissions each year (Harris et al., 2012, Van Der Werf et al., 2009). In addition, habitat loss and fragmentation of forest landscapes threaten ecosystem resilience and biodiversity (Folke et al., 2004, Meagher, 2010). Environmental externalities do not necessarily enter the economic calculus for private decision makers (Pigou, 1920). Efforts to align private and social incentives depend crucially on measuring and evaluating the impacts of activities affecting forests. Thus, reliable, timely, and transparent information on forest disturbance is urgently needed, especially in the humid tropics, which accounts for the majority of global deforestation (FAO, 2012).
Current techniques to monitor forest disturbance must balance spatial and temporal resolution. Sampling techniques using high-resolution imagery have been shown to reliably detect small-scale disturbance. But infrequent updates, relatively high data and processing costs, and limited spatial coverage all constrain these techniques (Goetz et al., 2009, Asner et al., 2010).
Moderate- and coarse-resolution imagery from Landsat and MODIS, respectively, is acquired at higher frequency and for broader geographic areas. Such data sets sacrifice spatial detail in favor of more timely information on forest cover disturbance (DeFries et al., 2006, Hansen et al., 2008a, Shimabukuro et al., 2006, Souza et al., 2005, Souza et al., 2009, Potter et al., 2012, Anderson et al., 2005, Reymondin et al., 2012).
Despite the high frequency of coarse-resolution imagery products, monitoring is still often constrained by persistent cloud cover in the humid tropics. The DETER monitoring system (Shimabukuro et al., 2006), for example, only reports data in the legal Amazon in Brazil, where cloud cover is less problematic than in other tropical regions (Roy et al., 2006, Gunderson and Chodas, 2011). Other systems (Potter et al., 2012) use resampled or aggregated data products to reduce the impact of cloud cover.
There are two broad approaches to change detection: cross-sectional differencing and time series analysis. Cross-sectional differencing is well-suited to relatively high resolution imagery that is acquired infrequently. However, persistent cloud cover often precludes the ability to identify transient change (Coppin et al., 2004). In Indonesia, for example, the conversion from primary forest to secondary forest or oil palm plantations can be completed between compositing periods, or even within the interval required to create cloud-free image composites (Broich et al., 2011).
Time series analysis examines the spectral history of each pixel, with different techniques depending on the temporal resolution of the imagery. Algorithms from Kennedy et al. (2007), Huang et al. (2010), and Broich et al. (2011) search for specific temporal signatures of forest disturbance in Landsat pixel histories, using one observation out of a potential 22 observations each year. Imaging twice a day by MODIS sensors aboard NASA's Terra and Aqua satellites yield far more potentially usable observations, reducing the impact of a cloudy scene on the overall time series. Choosing imagery with higher temporal resolution for sub-annual forest assessment requires discerning true change from a noisier time series with strong seasonal components, even in the tropics (Verbesselt et al., 2010b).
The dense MODIS time series has inspired a varied set of techniques for identifying landuse/landcover (LULC) change more broadly. Mildrexler et al. (2009) detect disturbances in North America by generating and analyzing annual composites of maximum land surface temperature (LST) in conjunction with composited and 16-day enhanced vegetation index (EVI) values. Campos and Di Bella (2012) also identify several types of LULC around the world, but analyze full MODIS NDVI time series using wavelet transforms. In the Dry Chaco ecoregion of South America, Clark et al. (2010) used TIMESAT to analyze multispectral MODIS time series and generate annual land use change statistics. Finally, Kleynhans et al. (2010) show that time series analysis of NDVI using an extended Kalman filter and 3 × 3 pixel neighborhoods is more effective at change detection than image differencing.
This paper proposes a methodology for an alerting system for forest disturbance using imagery from the Moderate Resolution Imaging Spectroradiometer (MODIS) and statistical techniques borrowed from time series econometrics. The methodology has been implemented as an open source, automated forest monitoring system that covers the humid tropics. By leveraging the high-frequency MODIS imagery, the system produces forest disturbance alerts at 16-day, 500 m resolution. It is intended to complement high-resolution forest change data sets that are updated less frequency but capture more variation in small-scale forest disturbance.
The algorithm relies on break detection algorithms to examine the spectral time series for each pixel, searching for structural changes in temporal patterns that are historically associated with forest disturbance. The structural changes are independent of both seasonal and random variation, helping to reduce the impact of clouds within the dense time series. In addition, the analysis incorporates rainfall in order to broadly control for transient climatic conditions during the sample period. A semi-supervised learning algorithm then characterizes each pixel's time series and compares it to historical forest cover loss data for 2000–2005, as identified by Hansen et al. (2008a). The classification rule derived from this process is used to classify each pixel for each 16-day interval after December 2005, the end of the training interval. Each pixel is assigned a normalized index representing the probability of disturbance.
The paper is organized as follows: Sections 1.1 Existing methodologies: data, 1.2 Existing methodologies: algorithms present an overview of existing monitoring systems and algorithms to measure and quantify forest disturbance. Section 1.3 positions the proposed algorithm within previous research. Section 2 describes the raw imagery and other data used to develop the final data product. Section 3 provides a detailed, step-by-step explanation of the algorithm. Section 4 reports the correspondence of the proposed method with other remotely sensed forest disturbance data in Pará, Brazil. Section 5 describes the computational tradeoffs within the final feature selection, and Section 6 concludes.
Section snippets
Time series data
The proposed algorithm uses the following time series data sets: (1) Normalized Difference Vegetation Index (NDVI) from the MODIS Terra 500 m vegetation indices product (MOD13A1); (2) rainfall from the Precipitation Reconstruction over Land (PREC/L) data set; and (3) active fire detections identified by the Fire Information for Resource Management System (FIRMS).
The 500 m NDVI layer in the MOD13A1 16-day composite (NASA, 2001) is a measure of pixel-level vegetation intensity used to flag
Methods
The classification process analyzes the temporal characteristics of each pixel during the training period. A separate classification rule is generated for each terrestrial ecoregion in order to account for local vegetation, land use, or forest clearing patterns. The algorithm is then applied across the forested humid tropics.
Results
The geographic scope and high temporal frequency of the full data output is unique and there is currently no pantropical dataset with which to directly compare the output data. Furthermore, terms like forest clearing, loss, degradation and disturbance, used throughout the literature, are applied inconsistently, making direct comparisons difficult. Thus, we reiterate here that the output dataset is designed to generate alerts of forest disturbance, or forest cover change, not to estimate change
Producer accuracy
Producer accuracy of the proposed alerting system is currently low in the comparison with PRODES. The primary goal of the alerting system, however, to produce reliable alerts, i.e., high user accuracy. The results suggest that the proposed detection algorithm yields the desired features of an alerting system. A future iteration of the algorithm currently under development will move to 250 m MODIS data – supporting detection of smaller patches of disturbance – and improve the classification data
Conclusion
Timely and geographically consistent information on forest disturbance is a necessary – albeit insufficient – condition to effectively manage the use and protection of forest resources. The proposed alerting system reports timely information on large-scale forest disturbance, and while the classification rule is locally tuned the algorithm overall is globally consistent.
This has the advantage of generating consistent data that can be compared across tropical countries or ecoregions. However,
References (58)
- et al.
Time-series analysis of multi-resolution optical imagery for quantifying forest cover loss in Sumatra and Kalimantan, Indonesia
Int. J. Appl. Earth Observ. Geoinform.
(2011) - et al.
A scalable approach to mapping annual land cover at 250 m using MODIS time series data: a case study in the Dry Chaco ecoregion of South America
Remote Sens. Environ.
(2010) - et al.
Use of a green channel in remote sensing of global vegetation from EOS-MODIS
Remote Sens. Environ.
(1996) - et al.
A method for integrating MODIS and landsat data for systematic monitoring of forest cover and change in the Congo basin
Remote Sens. Environ.
(2008) - et al.
An automated approach for reconstructing recent forest disturbance history using dense Landsat time series stacks
Remote Sens. Environ.
(2010) - et al.
Trajectory-based change detection for automated characterization of forest disturbance dynamics
Remote Sens. Environ.
(2007) - et al.
Land-cover change detection using multi-temporal MODIS NDVI data
Remote Sens. Environ.
(2006) - et al.
Testing a MODIS global disturbance index across North America
Remote Sens. Environ.
(2009) - et al.
Monitoring deforestation with MODIS active fires in neotropical dry forests: an analysis of local-scale assessments in Mexico, Brazil and Bolivia
J. Arid Environ.
(2013) - et al.
Global correlation of temperature, NDVI and precipitation
Adv. Space Res.
(1993)
Detecting trend and seasonal changes in satellite image time series
Remote Sens. Environ.
Phenological change detection while accounting for abrupt and gradual trends in satellite image time series
Remote Sens. Environ.
Assessment of deforestation in near real time over the Brazilian Amazon using multitemporal fraction images derived from terra MODIS
IEEE Geosci. Remote Sens. Lett.
High-resolution forest carbon stocks and emissions in the Amazon
Proc. Natl. Acad. Sci.
Denoising and wavelet-based feature extraction of MODIS multi-temporal vegetation signatures
GISci. Remote Sens.
Multi-temporal analysis of remotely sensed information using wavelets
J. Geogr. Inform. Syst.
Ridge estimators in logistic regression
Appl. Stat.
Global land precipitation: a 50-yr monthly analysis based on gauge observations
J. Hydrometeorol.
Construction of Newton-like iteration methods for solving nonlinear equations
Numer. Math.
Locally weighted regression: an approach to regression analysis by local fitting
J. Am. Stat. Assoc.
Digital change detection methods in ecosystem monitoring: a review
Int. J. Remote Sens.
Fire information for resource management system: archiving and distributing MODIS active fire data
IEEE Trans. Geosci. Remote Sens.
Spatial patterns of NDVI variation over Indonesia and Their Relationship to ENSO warm events during the Period 1982–2006
J. Climate
Regime shifts, resilience, and biodiversity in ecosystem management
Ann. Rev. Ecol. Evol. Syst.
Mapping and monitoring carbon stocks with satellite observations: a comparison of methods
Carbon Balance Manage.
Cited by (32)
Near-real-time monitoring of land disturbance with harmonized Landsats 7–8 and Sentinel-2 data
2022, Remote Sensing of EnvironmentSub-annual tropical forest disturbance monitoring using harmonized Landsat and Sentinel-2 data
2021, International Journal of Applied Earth Observation and GeoinformationA near-real-time approach for monitoring forest disturbance using Landsat time series: stochastic continuous change detection
2021, Remote Sensing of EnvironmentCitation Excerpt :A recursive transformation for these steps can free the system from loading all historical images each time for a new observation, greatly improving computational efficiency for a near real-time scenario. Currently, most operational near real-time products are based on MODIS system, such as Terra-i (Reymondin et al., 2012) and FORMA alert system (Hammer et al., 2014). To our knowledge, Global Land Analysis and Discovery (GLAD) is the only operational system providing a near real-time monitoring of forested land at a Landsat-pixel scale (Hansen et al., 2016).
Can VIIRS continue the legacy of MODIS for near real-time monitoring of tropical forest disturbance?
2020, Remote Sensing of EnvironmentCitation Excerpt :However, comparing daily observations from MODIS can be difficult due to the variation in the size and center location of the footprint of each observation caused by day-to-day differences in view angles (Tan et al., 2006; Xin et al., 2013). Existing near real-time monitoring systems often utilize multi-day composites of MODIS data to mitigate the view angle effect (e.g. Hammer et al., 2014; Reymondin et al., 2012; Verbesselt et al., 2012). Combining radar with optical remote sensing data is another commonly adopted approach for near real-time monitoring (e.g. Hirschmugl et al., 2017; Reiche et al., 2018).
Near real-time monitoring of tropical forest disturbance: New algorithms and assessment framework
2019, Remote Sensing of EnvironmentCitation Excerpt :Currently, many operational near real-time products use multi-date composites of MODIS observations. For example, Verbesselt et al. (2012) and the operational system Terra-i (Reymondin et al., 2012) have used 16-day MODIS-derived NDVI composites for deforestation detection, and the FORMA alert system (Hammer et al., 2014) utilizes 32-day NDVI composites. The DETER system used by the Brazilian Government for deforestation surveillance (Shimabukuro et al., 2007) requires manual interpretation of images from various sensors.
A machine learning approach to map tropical selective logging
2019, Remote Sensing of EnvironmentCitation Excerpt :Large-scale monitoring of deforestation has significantly improved in recent years, and forest losses can be identified with accuracies >90% using freely available satellite data (Hansen et al., 2013). In addition, near real-time deforestation tracking and alert systems are now possible with systems like DETER (Shimabukuro et al., 2014), FORMA (Hammer et al., 2014; Hansen et al., 2013), and Global Forest Watch (Hansen et al., 2016). In contrast, methods for detecting and monitoring forest degradation are less developed.