A multivariable approach for mapping sub-pixel land cover distributions using MISR and MODIS: Application in the Brazilian Amazon region

https://doi.org/10.1016/j.rse.2003.06.002Get rights and content

Abstract

Accurate mapping of land cover at continental to global scales is currently limited by our ability to exploit the spatial, temporal, and radiometric characteristics of the available satellite data. Many ecologically and biogeochemically important landscape features are spatially extensive, but occur at scales much smaller than the ∼1-km footprint of wide-swath, polar orbiting radiometers. This is especially true for land cover changes associated with human activities. Satellite instruments that offer the appropriate spatial detail have much smaller swaths and longer repeat times, resulting in compositing intervals that are too large to resolve the time scales of these changes. In addition, the cost and effort associated with acquisition and processing of high-resolution data for large areas is often prohibitive. Methods for taking advantage of information contained in multiple-scale observations by combining data from high-resolution and moderate resolution sensors are thus of great current interest.

In this paper, we retrieve land cover distributions in two different parts of the Brazilian Amazon region by estimating relationships between land cover fractions derived from 30-m resolution ETM+ and reflectance data from ∼1-km resolution MODIS and MISR. The scaling relationships are derived using a Bayesian-regularized artificial neural network (ANN) and compared to results using linear unmixing (LU). We explore the simultaneous use of two significant independent variables in terrestrial optical remote sensing, wavelength, and sun-sensor geometry, by combining nadir-adjusted MODIS reflectances in seven bands (VIS-SWIR) with multiangular (−71° to +71°) bidirectional reflectance data from MISR. This research was motivated by evidence from modeling and field studies demonstrating that: (a) the angular dependence of reflectance (e.g., from MISR) contains information about the structural composition of canopies that is complementary to the wavelength dependence; and (b) the SWIR portion of the spectrum (e.g., from MODIS) is sensitive to canopy moisture and shading conditions and, therefore, to the successional status of the ecosystem. This case study, using the Bayesian artificial neural network with combined MODIS-MISR data to estimate sub-pixel land cover fractions, yielded a quantitative improvement over spectral linear unmixing of single-angle, multispectral data. Our results suggest potential for broad-scale applicability despite a number of challenges related to tropical atmospheric conditions.

Introduction

The Brazilian Amazon region consists of over 4,000,000 km2 of tropical forest, representing one of the largest and most diverse contiguous ecosystems in the world. Over the past two decades, dramatic land cover changes in Amazonia have resulted in a wide variety of ecological and biogeochemical impacts, ranging in scale from local to global, and including changes in forest productivity and composition, nutrient dynamics, species diversity, stream chemistry, and atmospheric carbon dioxide Houghton et al., 2000, Potter et al., 2001. Furthermore, alterations in forest structure are thought to impact regional climate via biophysical feedbacks between the atmosphere and biosphere Costa & Foley, 2000, Walker et al., 1995. Therefore, quantification of the magnitude, timing, and spatial extent of human modification of the landscape in this region is currently an important Earth science research topic.

While deforestation-related land cover changes are widespread, affecting nearly 600,000 km2 in the Brazilian Amazon region alone, they occur at small spatial scales and exhibit highly dynamic interannual variability (INPE, 2000). The affected areas are thus a constantly evolving mosaic of cleared land and secondary vegetation fragments of varying size and age, woven into a background of relatively undisturbed forest. Because the underlying causes and resulting patterns of land use activity in Amazonian forests depend upon a variety of economic, social, and ecological factors that are extremely difficult to document, quantifying disturbance and recovery of Amazonian forests is typically addressed with satellite remote sensing techniques.

The objective of this paper is to present an approach for mapping land cover distributions using data from multiple-scale satellite observations, with specific application to estimating patterns of deforestation and recovery in Brazil. We present and evaluate a method for estimating sub-pixel land cover fractions that is unique in the following ways: (1) the scaling between Landsat ETM+ and Terra (MISR and MODIS) data is accomplished using an artificial neural network (ANN) method that is designed to prevent overfitting; and (2) we combine multiangle and multispectral data from the two Terra sensors and compare the utility of various band-angle combinations. In Section 2, we will review some of the strategies that have been previously used, and briefly discuss the surrounding analytical and remote sensing issues that form the foundation of our approach.

Section snippets

Resolving land cover change patterns

The availability of wide-swath, coarse spatial resolution data from polar-orbiting instruments (e.g., AVHRR, MODIS, and MISR) allows classification of large regions at kilometer scales and greater. A number of these products exist (e.g., DeFries et al., 1998, Loveland et al., 2000), and they are often intercalibrated with one another or linked to field data and censuses (e.g. Cardille et al., 2002, Frolking et al., 2002). These data sets have been useful in a wide variety of applications,

Methods

In this study, we produced a set of sub-pixel land cover fraction estimates at a moderate spatial scale (∼1 km) in two heavily impacted regions of the Brazilian Amazon region. The estimates were predicted using reflectance data from MODIS and MISR, based on known land cover estimates derived from unsupervised classification of ETM+ with manual image editing. Thus, the coarse resolution observations, derived from the available bands and view angles, are treated as independent variables in a

Results and discussion

The results of the 11 cases reveal a large range in ability to estimate sub-pixel land cover fractions (Table 1; note that the following section numbers correspond to the analysis section numbers in Table 1). The sub-pixel fraction estimates for our base case (Case A; using the BANN, and trained and tested in Ruropolis) show good reproduction of spatial land cover distribution patterns. However, some overestimates of secondary vegetation, and corresponding underestimates of forest, are evident

Conclusions

The results presented in previous sections suggest that nonlinear regression offers significant improvement relative to linear unmixing for estimation of sub-pixel land cover fractions in the heterogeneous disturbed areas of Brazilian Amazonia. This improvement is likely due to the fact that linear unmixing assumes the existence of pure sub-pixel classes with fixed reflectance signatures (end-members). In this application, two of the land cover classes (cleared land and secondary vegetation)

Acknowledgements

We are grateful for the constructive comments of two anonymous reviewers. We also thank Michael Palace for helpful suggestions about the manuscript. Thanks to Dr. David Diner for providing advice and support for our use of the MISR land data products. The MISR and MODIS data were obtained through the NASA EOS Data Gateway system. The Landsat ETM+ data was purchased as part of a NASA Large Scale Biosphere Atmosphere Experiment in Amazonia (LBA) project. This study was funded by grants from the

References (45)

  • C. Bishop

    Neural networks for pattern recognition

    (1995)
  • B. Braswell et al.

    Extracting ecological and biophysical information from AVHRR optical data: An integrated algorithm based on inverse modeling

    Journal of Geophysical Research

    (1996)
  • J. Cardille et al.

    Characterizing patterns of agricultural land use in Amazonia by merging satellite classifications and census data

    Global Biogeochemical Cycles

    (2002)
  • S. Chaterjee et al.
  • M.H. Costa et al.

    Combined effects of deforestation and doubled atmospheric CO2 concentrations on the climate of Amazonia

    Journal of Climate

    (2000)
  • A. Cross et al.

    Subpixel measurement of tropical forest cover using AVHRR data

    International Journal of Remote Sensing

    (1991)
  • D. Deering

    Field measurements of bidirectional reflectance

  • R. DeFries et al.

    Global land cover classifications at 8 km spatial resolution: The use of training data derived from Landsat Imagery in decision tree classifiers

    International Journal of Remote Sensing

    (1998)
  • P. Fisher

    The pixel: A snare and a delusion

    International Journal of Remote Sensing

    (1997)
  • G. Foody et al.

    Non-linear mixture modeling without end-members using an artificial neural network

    International Journal of Remote Sensing

    (1997)
  • S. Frolking et al.

    Combining remote sensing and ground census data to develop new maps of the distribution of rice agriculture in China

    Global Biogeochemical Cycles

    (2002)
  • I. Grant

    Investigation of the variability of the directional reflectance of Australian land cover types

    Remote Sensing Reviews

    (2000)
  • Cited by (88)

    • Simultaneous retrieval of sugarcane variables from Sentinel-2 data using Bayesian regularized neural network

      2023, International Journal of Applied Earth Observation and Geoinformation
    • Spatial homogeneity from temporal stability: Exploiting the combined hyper-frequent revisit of Terra and Aqua to guide Earth System Science

      2021, Remote Sensing of Environment
      Citation Excerpt :

      A prominent example is agricultural monitoring (Duveiller and Defourny, 2010), where the successful use of crop-specific time series requires isolating only those pixels which adequately fall in the appropriate fields using ancillary information at finer spatial resolution (Duveiller et al., 2011). Other applications that are confronted to this problem of identifying homogeneous pixels that are representative of the target object include detecting deforestation (Morton et al., 2005), land cover change (Braswell et al., 2003), fires (van der Werf et al., 2017) or water bodies (Huang et al., 2018). Progress in satellite Earth Observation technology should eventually provide higher spatial resolution with sufficiently frequent revisit times, such as with construction of virtual constellations like the combination of Sentinel2 and Landsat-8 (Claverie et al., 2018), but it will not be possible to extend it to the past.

    • A comparison of two downscaling procedures to increase the spatial resolution of mapping actual evapotranspiration

      2017, ISPRS Journal of Photogrammetry and Remote Sensing
      Citation Excerpt :

      All these models require data from the thermal bands of satellite sensors. Daily ET maps estimated from available and accessible RS data are too coarse to be used in precision farm management, as the pixel sizes are commonly too large to well present spatial variability within individual fields, thus causing substantial errors (Braswell et al., 2003). Such errors are therefore caused by pixels that have diverse land cover and various vegetation types, roughness conditions and soil moisture contents (Kustas, 2004).

    View all citing articles on Scopus
    View full text