Review Article
Burned area estimations derived from Landsat ETM+ and OLI data: Comparing Genetic Programming with Maximum Likelihood and Classification and Regression Trees

https://doi.org/10.1016/j.isprsjprs.2018.05.007Get rights and content

Highlights

  • Genetic Programming, a methodology to mapping burned areas using satellite images.

  • Supervised classification methods for burnt detection in tropical regions.

  • Burnt detection in savanna and woodlands located in tropical regions.

  • Comparison of Maximum likelihood and Classification and Regression Tree algorithms.

  • Landsat satellite images, a valuable source of information to map burned areas.

Abstract

Every year, large areas of savannas and woodlands burn due to natural conditions and land management practices. Given the relevant level of greenhouse gas emissions produced by biomass burning in tropical regions, it is becoming even more important to clearly define historic fire regimes so that prospective emission reduction management strategies can be well informed, and their results Measured, Reported, and Verified (MRV). Thus, developing tools for accurately, and periodically mapping burned areas, based on cost advantageous, expedite, and repeatable rigorous approaches, is important. The main objective of this study is to investigate the potential of novel Genetic Programming (GP) methodologies for classifying burned areas in satellite imagery over savannas and tropical woodlands and to assess if they can improve over the popular and currently applied methods of Maximum Likelihood classification and Classification and Regression Tree analysis. The tests are performed using three Landsat images from Brazil (South America), Guinea-Bissau (West Africa) and the Democratic Republic of Congo (Central Africa). Burned areas were digitized on-screen to produce mapped information serving as surrogate ground-truth. Validation results show that all methods provide an overestimation of burned area, but GP achieves higher accuracy values in two of the three cases. GP is the most versatile machine learning method available today, but still largely underused in remote sensing. This study shows that standard GP can produce better results than two classical methods, and illustrates its versatility and potential in becoming a mainstream method for more difficult tasks involving the large amounts of newly available data.

Introduction

In tropical regions, large areas of savanna and woodlands burn every year. Occurring mainly during the dry season when herbaceous vegetation has dried out, fires are one of the main drivers of ecosystem transformation or maintenance (Bucini and Lambin, 2002), also releasing gases and particles into the atmosphere (Smith et al., 2007). In fact, estimates show that burning of savannas and woodlands in Sub-Saharan Africa accounts for more than 50% of the total global emissions from biomass burning during any typical year (Williams et al., 2012).

Land management practices induced by human activities are at the base of the majority of fire occurrences in the tropics. Shifting cultivation, agricultural expansion, deforestation and harvesting are some of the practices involving fires that may contribute to partial or complete destruction of vegetation cover, depending on fire intensity and combustion efficiency (Bucini and Lambin, 2002, Daldegan et al., 2014). Significant intensification of fire frequency or avoidance of fire occurrence can negatively affect existing ecosystems and have impacts on vegetation composition, landscape patterns, habitat types, and soil erosion processes, which in turn affect hydrological processes and the carbon cycle. Therefore, accurate and multi-temporal burned area maps are important tools that can help fire and land managers understand and assess the impacts of specific interventions, while informing landscape management strategies.

Multi-temporal data records of fire distribution, extent, and timing, correspond to historical activity data, which together with vegetation emission factors, support the quantification of emissions. The establishment of emission baselines against which the results of subsequent vegetation and fire management actions can be compared is essential for the Measuring, Reporting and Verification (MRV) activities necessary in carbon accounting procedures. Thus, accurately and frequently mapping burned areas over large extents, using cost advantageous, periodic, and expedite approaches is very desirable.

In the last decade, several methodologies based on remote sensing techniques have been developed and applied to recurrently map burned areas in tropical ecosystems. Some were based on coarse spatial resolution satellite data such as Moderate Resolution Imaging Spectroradiometer (MODIS), Along Track Scanning Radiometer (ATSR)/Advanced ATSR (AATSR), Satellite pour l’Observation de la Terre (SPOT) Vegetation (VGT) and National Oceanic and Atmospheric Administration (NOAA)/Advanced Very High Resolution Radiometer (AVHRR) (Brivio and Maggi, 2003, Giglio et al., 2009, Silva et al., 2005, Grégoire et al., 2003, Zhang et al., 2015). These are adequate for global and regional scale studies but are insufficient for local applications where higher detail is needed and medium to high resolution sensors are preferable for accurately mapping burned areas (Stroppiana et al., 2012).

The higher resolution sensors Landsat TM (Thematic Mapper), ETM+ (Enhanced Thematic Mapper) and OLI (Operational Land Imager) are a valuable source of information and have been widely used in the development of automated methods to detect burned areas (Bastarrika et al., 2011, Chen et al., 2016, Daldegan et al., 2014, Hudak et al., 2004, Júnior et al., 2014, Korontzi et al., 2003, Hawbaker et al., 2017, Meddens et al., 2016, Laris, 2005, Liu et al., 2018, Melchiori et al., 2014, Matricardi et al., 2013, Oumar, 2015, Smith et al., 2007, Stroppiana et al., 2012, Trisakti et al., 2016). Even though classification of burned areas using Landsat images provide satisfactory results with classical approaches (Morton et al., 2011), the spectral similarities between burnt surfaces and other land cover categories, such as, water bodies, shadows, and mixed water-vegetation, still introduce spectral confusion and overlap with other classes. Therefore, it is important to explore new methods capable of increasing the discrimination between burns and other landscape features, minimizing the uncertainties (Giglio et al., 2010, Jain, 2007).

This study aims at investigating if there are comparative advantages in using Genetic Programming (GP) – one of the most powerful and underused flavors of machine learning – for identifying and mapping burned areas in Landsat ETM+/OLI imagery when compared to Maximum Likelihood classification (MLK) and Classification and Regression Tree analysis (CART) – two classical classification methods. Our research is conducted over three study areas located in Brazil, Guinea-Bissau, and the Democratic Republic of Congo. The respective tropical territories are subject to frequent and extensive fires, mainly due to human activity.

The merit of each approach is assessed by calculating the overall accuracy, Dice and kappa coefficients, and omission and commission errors over a representative sample grid of points extracted from the images. According to Padilla et al. (2014), measures such as the Dice coefficient that are focused on a single category (i.e. burned), are the most appropriate in the validation of Burned area products. Additionally, the agreement of the classifications with surrogate ground-truth burned area maps is calculated based on precision and recall measures (Powers, 2007). Surrogate ground-truth is obtained from visual interpretation and on-screen digitizing of burned area perimeters over the entire images. In order to assess the similitude of the overall landscape structure obtained from the on-screen digitizing with that obtained from the classifications, a set of landscape metrics are also derived and compared.

Several authors applied MLK and CART to map burned areas (Chen et al., 2016, Henry, 2008, Meddens et al., 2016, Sá et al., 2003, Sertel and Alganci, 2016, Silva et al., 2003, Thariqa et al., 2016, Verlinden and Laamanen, 2006), however, very few studies exist for GP (Silva et al., 2010). Djerriri and Mimoun (2015) successfully applied a new approach combining unsupervised classification and GP to automatically extract burned areas from Landsat 8 imagery. Also, Brumby et al. (2001) obtained encouraging results when applying GP to extract wildfire scars from Landsat 7 imagery, but found some confusion with dark cloud shadows and bare ground/rock outcrops. More recently, a different type of GP, called Geometric Semantic Genetic Programming (GSGP) (Vanneschi, 2017), was used by Castelli et al. (2015) for identification of burned areas. Although GSGP is a very powerful method, it does not provide readable models. Even though very few applications of GP for classification/data extraction of remote sensing images can be found in the literature, GP has been successfully used in several other areas, e.g., modeling and regression, image and signal processing, time series prediction, control, medicine, biology and bioinformatics, and even arts and entertainment (Poli et al., 2008). GP often yields results that are not merely academically interesting, but competitive with the work developed by humans (Koza, 2010). It is the master algorithm of evolutionary computation, and the only one with the potential to emulate all the other machine learning approaches (GP can evolve decision trees, neural networks, Bayesian networks, and almost anything else one can think of) (Domingos, 2015).

New sensors, such as those on board of the European Union (EU) Sentinel satellites1 provide free full global coverage and high frequency optical and radar imagery. The EU Copernicus program, which also aims at providing environmental monitoring services for South America and Africa,2 can become a driver for the systematic and high periodicity production of high resolution burned area maps over tropical regions. Therefore, methodological developments that contribute to improve operational processes while improving output accuracy may increase the usefulness of products and facilitate their respective diffusion. Recent studies have shown the feasibility of using distributed GP in long running systems dealing with big data (Hodjat et al., 2014).

Section snippets

Study areas

One study area located in Brazil and two in Africa were chosen to test the performance of the burned area mapping methods: the southeastern Amazonian region of Brazil, the Coastal western region of Guinea-Bissau, and the central eastern region of Congo; each corresponding to one complete Landsat image as shown in Fig. 1.

The first area, located in eastern Amazonia, in southeastern Pará, Brazil (BRZ site) lies to the south of the Amazon River which is drier than the central and western parts of

Reference dataset

Visual inspection of the combination of Landsat ETM+ bands 7, 4 and 3 and Landsat OLI bands 7, 5 and 4 allows a clear visual depicting of burned areas (Pereira et al., 1999). Using the 7-4-3 or 7-5-4 combination for display, according to the sensor, all visible polygons of burned areas perimeters were manually delimited on-screen to constitute a burned area ground-truth map for each of the study areas. These burned areas, and their complementary non-burned polygons, constitute the binary base

Results and discussion

Visual comparison between the classified maps and the satellite images shows distinct performance patterns according to the site location and classification method adopted. A detail of fire perimeter agreement between the ground-truth maps and the classified burned areas is shown for Area 1 and Area 2 of each of the three locations in Fig. 4.

The Precision and Recall values obtained for each burned area map are reported in Table 4. These values are also depicted as points in a plot to facilitate

Conclusions

The present study has compared three different methods (GP, CART and MLK) for detecting burned areas in three different sites (BRZ, GB and DRC). The results have shown that, depending on the study area and sensor type, the three methods achieved different accuracies. Nevertheless, the accuracies of the burned area maps produced by the GP methods were always higher than those produced by CART, and only at the BRZ site GP performed worse than MLK, being affected by a small percentage of

Acknowledgments

The projects leading to this work have received funding from the European Union’s Horizon 2020 Research and innovation programme under the Marie Skłodowska - Curie grant agreement No. 691053, and from the FCT/MCTES/PIDDAC (Portuguese Foundation for Science and Technology) which funded strategic projects UID/AGR/002389/2013 (CEF) and UID/MULTI/04046/2013 (BioISI).

References (70)

  • J. Liu et al.

    Burned area detection based on Landsat time series in savannas of southern Burkina Faso

    Int. J. Appl. Earth Observ. Geoinform.

    (2018)
  • A.J.H. Meddens et al.

    Detecting unburned areas within wildfire perimeters using Landsat and ancillary data across the northwestern United States

    Remote Sens. Environ.

    (2016)
  • D.C. Morton et al.

    Mapping canopy damage from understory fires in Amazon forests using annual time series of Landsat and Modis data

    Remote Sens. Environ.

    (2011)
  • M. Padilla et al.

    Validation of the 2008 MODIS-MCD45 global burned area product using stratified random sampling

    Remote Sens. Environ.

    (2014)
  • J.M.N. Silva et al.

    Comparison of burned area estimates derived from SPOT-Vegetation and Landsat ETM+ data in Africa: influence of spatial pattern and vegetation type

    Remote Sens. Environ.

    (2005)
  • D. Stroppiana et al.

    A method for extracting burned areas from Landsat TM/ETM+ images by soft aggregation of multiple spectral indices and a region growing algorithm

    ISPRS J. Photogr. Remote Sensing

    (2012)
  • L. Breiman et al.

    Classification and Regression Trees

    (1984)
  • P.A. Brivio et al.

    Mapping burned surfaces in Sub-Saharan Africa based on multi-temporal neural classification

    Int. J. Remote Sens.

    (2003)
  • Brumby, S.P., Theiler, J., Perkins, S., Harvey, N.R., Szymanski, J.J., 2001. Genetic programming approach to extracting...
  • A.I.R. Cabral et al.

    A land cover map of Southern hemisphere Africa based on SPOT-4 Vegetation data

    Int. J. Remote Sens.

    (2006)
  • M. Castelli et al.

    Predicting burned areas of forest fires: an artificial intelligence approach

    Fire Ecol.

    (2015)
  • L.M.F. Catarino

    Fitogeografia da Guiné-Bissau. Provas de doutoramento em Engenharia Agronómica

    (2004)
  • L. Catarino et al.

    Cashew cultivation in Guinea-Bissau – risks and challenges of the sources of a cash crop

    Sci. Agricola

    (2015)
  • W. Chen et al.

    Mapping a burned area from Landsat TM data by multiple methods

    Geomatics, Nat. Hazards Risk

    (2016)
  • G.A. Daldegan et al.

    Spatial patterns of fire recurrence using remote sensing and GIS in the Brazilian Savanna: Serra do Tombador Nature Reserve, Brazil

    Remote Sens.

    (2014)
  • K. Djerriri et al.

    Genetic programming and one-class classification for discovering useful spectral transformations

    IGARSS

    (2015)
  • Domingos, P., 2015. The Master Algorithm: How the Quest for the Ultimate Learning Machine Will Remake Our World. Basic...
  • L. Giglio et al.

    Assessing variability and long-term trends in burned area by merging multiple satellite fire products

    Biogeosciences

    (2010)
  • J.-M. Grégoire et al.

    The GBA2000 initiative: developing a Global Burned Area Database from SPOT-VEGETATION imagery

    Int. J. Remote Sens.

    (2003)
  • T. Hayes et al.

    Using Classification and Regression Trees (CART) and Random Forests to analyze attrition: results from two simulations

    Psychol. Aging

    (2015)
  • M.C. Henry

    Comparison of single- and multi-date Landsat data for mapping wildfire scars in Ocala National forest, Florida

    Photogr. Eng. Remote Sens.

    (2008)
  • B. Hodjat et al.

    Maintenance of a long running distributed genetic programming system for solving problems requiring big data

  • Ickowitz, A., Slayback, D., Asanzi, P., Nasi, R., 2015. Agriculture and deforestation in the Democratic Republic of the...
  • A.C.P. Júnior et al.

    Modeling fire frequency in a cerrado savanna Protected area

    PLOS ONE

    (2014)
  • Koza, J.R., 1992. Genetic Programming – On the Programming of Computers by Means of Natural Selection. MIT Press, 813...
  • Cited by (51)

    • Ecological landscape pattern changes and security from 1990 to 2021 in Ebinur Lake Wetland Reserve, China

      2022, Ecological Indicators
      Citation Excerpt :

      Scale has always been the focus and difficulty of the LULC and ecosystem services research. Numerous scholars have explored the evolution of LULC changes and ecological landscape patterns from different angles, many of which have supported the applicability of remote sensing technology in monitoring LULC changes and landscape ecology at local, regional (Cabral et al., 2018), watershed (Sisay et al., 2021), and global scales (Skariah and Suriyakala, 2022; Sutton and Costanza, 2002; Trinh et al., 2018). The LULC research on a global scale focuses on the change monitoring of artificial surfaces, forests, farmland, and wetland, as well as their relationship with soil processes, climate change, and biodiversity.

    • Sentinel-2 sampling design and reference fire perimeters to assess accuracy of Burned Area products over Sub-Saharan Africa for the year 2019

      2022, ISPRS Journal of Photogrammetry and Remote Sensing
      Citation Excerpt :

      The Dice Coefficient (DC) (Dice, 1945) is a synthetic metric combining omission and commission varying in the range [0, 1], the closest to 1 the best. It is a similarity index, it provides a single measure that combines the information of the separate class-specific metrics Ce and Oe (Padilla et al., 2015) and it has been exploited before as a measure of the accuracy of the burned category (Cabral et al., 2018). The Relative Bias (relB) is the difference between total burned area in the classified (S2) and reference (Planet) BA maps normalized by the total reference BA (Planet): positive (negative) relB values highlight overestimation (underestimation) errors.

    • Pixel- and Object-Based ensemble learning for forest burn severity using USGS FIREMON and Mediterranean condition dNBRs in Aegean ecosystem (Turkey)

      2022, Advances in Space Research
      Citation Excerpt :

      In this regard, supervised classification of remotely sensed data for burn severity mapping offers possible capabilities compared to the use of a single index (Gibson et al., 2020). Many authors have used classification algorithms involving the pixel- and object-based approaches for burned area mapping (Mitri and Gitas, 2014; Sertel and Alganci, 2016; Kavzoglu et al., 2016; Cabral et al., 2018; Gibson et al., 2020; Hamilton et al., 2021). Although classification of burnt areas with traditional approaches using satellite images offers satisfactory results, spectral similarities between burnt areas and other land cover classes, such as mixed soil-vegetation, water bodies and shadows, still poses a problem in spectral confusion within classes (Cabral et al., 2018).

    • Detecting high-temperature anomalies from Sentinel-2 MSI images

      2021, ISPRS Journal of Photogrammetry and Remote Sensing
      Citation Excerpt :

      However, the LR detector elements hinder small HTA detection due to their coarse resolution and impede the ability to retrieve HTAs' spatial information (location, size, shape), as they will typically be subpixel in size (Murphy et al., 2016). Moderate-resolution (MR; tens of meters) optical sensors, such as Landsat-8 Operational Land Imager (OLI) and Sentinel-2A/2B Multispectral Instrument (MSI), provide more spatially-specific information at a fine-scale and have played an essential role in HTA-related studies, especially in detailed mapping of the affected extent of fires (Cabral et al., 2018; Hawbaker et al., 2017; Roteta et al., 2019). However, relatively little attention has been paid to HTA detection from MR imaging, which may be attributed to the following reasons: (i) MR sensors usually lack optimum ‘fire’ bands (Giglio et al., 2008), making short-wave infrared (SWIR; ~1.1–2.5 μm) bands the only available option; (ii) in daytime SWIR bands, the radiance emitted from HTAs is mixed with the reflected solar radiation (Schroeder et al., 2008), whereas the former varies with HTA's size and temperature relative to non-burning components (Kaufman et al. 1998; Giglio and Justice 2003); (iii) in contrast to burned areas that exhibit a persistent surface signal in RS images, only HTAs that are actively burning/operating at the time of satellite observation can be sensed (Kumar and Roy, 2018).

    • Urban structure and its implication of heat stress by using remote sensing and simulation tool

      2021, Sustainable Cities and Society
      Citation Excerpt :

      This classification system divided the urban structures into more detailed categories based on local characteristics (Wang et al., 2019). The MLC method used in this research is a reliable and fully developed method, according to previous research (Cabral et al., 2018). This research combined the classification system with a method to identify urban structures, and a confusion matrix was selected to validate the accuracy of the results.

    View all citing articles on Scopus
    View full text