Examining patterns of spatio-temporal variation for forest type mapping 

Get Complete Project Material File(s) Now! »

Bi-directional effects and estimates of spatial variation of forest types

The following methods address the first research question, related to how bi-directional effects affect the spatial variation of spectral response of tropical forests. As elaborated in the General Introduction, the focus of the research question was to broadly understand how instrumental artefacts caused by BRDF not only affect estimates of canopy spectral response, but also the implications such artefacts have for mapping forest types in the tropics. The South American territory of French Guiana, whose approximate area is 83,534 km2, was selected as the study area due to its high estimated overall forest cover (Barret at al. 2001). In terms of the coverage of the remote sensing data used to address the research question (i.e. Landsat-5, Landsat-7 and MODIS), the territory covers some eight scenes in the Landsat World Reference System-2 (see Figure 3), and is located entirely within MODIS tile h12v08.

Selection of a method for assessing BRDF effects in reflectance data

With regard to estimating the impacts of BRDF on the reflectance data from those sources, as a number of methods of both assessing [and correcting] the bi-directional effects in Landsat data have been proposed, part of the study design involved selection of an appropriate method for such assessment. One set of methodologies originally tested over Australia, for instance, uses the reflectance estimates from overlapping satellite overpasses to develop BRDF correction factors based on sensor scan angles, among other factors (Danaher et al. 2001; Wu et al. 2001; Danaher 2002; Flood 2013; Flood et al. 2013).
On the other hand, another means of assessing and correcting bi-directional effects in Landsat data is proposed by studies piloted over the Congo Basin (Hansen et al. 2008; Potapov et al. 2012). In that method, Landsat data are compared with MODIS surface reflectance estimates to determine the bias between the two reflectance estimates. However, in that approach, the authors proposed evaluating the Landsat-MODIS bias using the MOD09 surface reflectance product which is not corrected for bi-directional effects. A modification on that approach would thus be to calculate the Landsat-MODIS bias using the MODIS surface reflectance data which are already corrected for BRDF effects (i.e. the MCD43A4 product, see (Schaaf et al. 2002), and this was the approach employed in the present study. (As a caveat, it should be mentioned that it is difficult to indicate how well the BRDF correction of the MCD43A4 product is, but this was addressed somewhat by this research.) In addition to being atmospherically corrected, both the MOD09 and MCD43A4 surface reflectance are already corrected for topographic artefacts (Strahler and Muller 1999; Vermote and Vermeulen 1999).

Generation of reference MODIS reflectance dataset

Nadir BRDF-corrected reflectance estimates from MODIS from the MCD43A4 product (MODIS collection 5) were acquired from the publicly accessible NASA Reverb portal (http://reverb.echo.nasa.gov/reverb/) for the period July 2002 – March 2014 for the h12v08 tile which covers French Guiana. The quality assessment / quality control (QA/QC) flags from the corresponding MCD43A2 product were used to filter the reflectance data by extracting only data corresponding to “best quality, full [BRDF] inversion” (Schaaf et al. 2002). Since the Landsat data to which the MODIS reflectance was to be compared corresponded to data from September (of different years), a single September composite of the filtered MCD43A4 data was generated by averaging reflectance data corresponding to the multiple, overlapping 16-day datasets which cover September. These included the data for Julian days 241 (29 August to 13 September), 249 (covering 6-21 September), and 257 (14-29 September). Averaging the data over the period indicated was used as a means of both reducing noise in the MODIS reflectance data, and maximizing the number of observations over French Guiana, necessitated by shifting cloud cover. Further, the September composite was assembled by using minimum reflectance in all of the spectral bands, with the exception of the near-infrared band, for which maximum reflectance was extracted. This was also done as a means of removing noise, as proposed in an earlier study (Hansen et al. 2003). As evidenced by earlier research, aerosols tend to increase spectral reflectance in most visible and mid-infrared parts of the electromagnetic spectrum, but they actually decrease spectral reflectance in the near-infrared part of the spectrum (Kaufman et al. 1992). Hence, compositing using the maximum near-infrared reflectance will tend to remove aerosol-contaminated pixels.

Generation of Landsat surface reflectance dataset

The assembly of a Landsat surface reflectance mosaic for French Guiana involved more data processing than that of assembling the MODIS mosaic (see Figure 3). For one, imagery from multiple dates with slightly differing illumination conditions had to be selected because of persistent cloud cover over French Guiana, which is even a problem in the dry season. The imagery used were acquired via the U.S. Geological Survey’s Glovis portal (http://glovis.usgs.gov), and had acquisition dates ranging from 27 September 2005 to 5 September 2009 (Table 1). Where reflectance is known to be impacted by solar elevation, it should be noted that the images had a maximum difference in solar elevation of only 2 degrees (Nagol et al. 2015).

Effects of seasonality on mapping the spatial variation of forest types

Complementary to the evaluation of how BRDF effects impact estimates of forest type distribution, additional analyses concentrated on understanding how estimates of forest type distribution vary temporally in data already corrected for BRDF. This was related to the second research question, which focused on how estimates of spatial distributions of tropical forest types change seasonally. Overall, the methods consisted of (i) extracting reflectance data for the forest cover classification, (ii) the mapping of forest types, (iii) evaluation of the separability of the forest types, and (iv) the estimation of the number of spectrally distinct forest types in each reflectance image. With regard to extracting the reflectance data, due to the low rate of land use change in French Guiana, estimates of mean monthly reflectance were derived, as the territory’s forests can be considered ‘stable.’ That reflectance data was compiled using the 463m resolution nadir BRDF-adjusted MCD43A4 reflectance data, which was used in place of the more commonly used MOD09 surface reflectance data. Data were acquired for the h12v08 tile in MODIS’ reference system, which includes all of French Guiana and Suriname, most of Guyana, and a part of Brazil’s Amapá state. The MCD43A4 data are available for 46 overlapping 16 day periods, which over the 15 year time period studied amounted to 688 datasets spanning Feb. 2000 – Feb. 2015. Data were extracted for 6 of MODIS’ 7 spectral bands (excluding the 1230-1250nm spectral band). Quality assessment flags from the complementary MCD43A2 data were used to extract only the highest quality reflectance data. For each 16 day period, an ‘average’ was computed as follows: for all of the spectral bands with the exception of the near-infrared band (MODIS band 2), the 15-year minima were calculated, and for the NIR band, the maxima were calculated. This mirrored the approaches used in previous studies (Hansen et al. 2003; Viennois et al. 2013) and allowed for filtering noise from the respective spectral bands. Data were then grouped according to the approximate calendar months they pertain to, and those reflectances were averaged to generate mean monthly reflectance estimates (based on the 15-year time-series). As the MODIS h12v08 tile covers much of the Guiana Shield, data were spatially subset to the extent of French Guiana. However, as shown in Figure 7 (derived from the resulting data), even using the 15 years of acquisitions, missing data due to cloud cover was still an issue, particularly during the first half of the year (the rainy season). Although only August and September had 100% coverage, July to December was generally cloud-free. Thus, for this analysis, only the data from July-December were used to evaluate forest type discrimination. A 3×3 low-pass filter was applied to the reflectance data to smooth the data and by so doing remove the ‘salt-and-pepper’ pattern observed (Lillesand et al. 2007).

Evaluating the temporal patterns of variation of tropical forests

Complementary to the evaluations related to the spatial patterns of variation of tropical forest (the first axis of this research) were analyses related to the temporal patterns of variation of tropical forest (the second axis of this research).

Assessing seasonal variation of vegetation indices at the regional level

The first part of the assessment of temporal patterns of variation addressed this thesis’ third research question, which was related to the extent to which observed seasonal variation in vegetation indices (VIs) correspond to seasonal climatic variation or to instrumental or atmospheric artefacts. This particular set of analyses also focused on VI variation at the regional level. With a wide variety of satellite-derived VIs available for such analysis, this study focused on analysis of data from 3 indices: (i) the Enhanced Vegetation Index, EVI (derived from MODIS’ MCD43B4 product), (ii) the fraction of green vegetation cover index, FCOVER (derived from VGT), and the Leaf Area Index, LAI (derived from VGT).
EVI, on the one hand, estimates “the ‘green’ vegetation signal across a global range of vegetation conditions while minimizing canopy influences associated with intimate mixing by non-vegetation related signals” (Huete and Justice 1999). FCOVER, in turn, estimates “the fraction of green vegetation covering a unit area of horizontal soil… correspond[ing] to the gap fraction in the nadir direction,” and is said to be “a very good candidate for substitution of classical vegetation indices” (Camacho et al. 2009). LAI corresponds to “half the total developed area of green elements per unit horizontal ground area” (Baret et al. 2013). EVI is derived from ratios of blue, red, and near-infrared reflectance, while FCOVER and LAI are both generated by radiative transfer model inversion and application of neural network algorithms to red, near-infrared and shortwave infrared reflectance data of VGT (Baret et al. 2013; Huete and Justice 1999). EVI and FCOVER were generated from nadir-normalized imagery, precluding potential view angle effects (Schaaf et al. 2011; Camacho et al. 2009).
Data were acquired from NASA’s Reverb system and the European Commission’s Copernicus Global Land Service (both publicly accessible). Data were acquired for the 139-month period spanning September 2002-March 2014, chosen mostly for the temporal overlap between SPOT VEGETATION and MODIS (Aqua and Terra). The study area consisted of three 10 degree x 10 degree zones located on the equator (Figure 8), and corresponding to tiles h12v08, h19v08, and h29v08 in MODIS’ reference system. The westernmost tile (1) covers most of the Guianas (i.e. the Guiana Shield), including all of French Guiana and Suriname, most of Guyana and the Brazil’s state of Amapá. The central tile (2) covers an area of central Africa including the majority of Cameroon and Equatorial Guinea, half of the Central African Republic, and parts of Nigeria, Gabon and Congo. The easternmost tile (3) covers the northern half of the island of Borneo, including all of Brunei, and parts of Malaysia and Indonesia. As the areas were extensive and covered a range of ecosystems, for standardization, only evergreen forests were considered, using a mask derived from the European Space Agency’s 2009 GlobCover map.

Assessing seasonal variation of vegetation indices at the plot level

Related to the previous evaluation, it was likewise decided to examine the issue of VI variation, but at the plot level, to complement the regional level analyses. That analysis addressed the fourth research question regarding the extent to which VI variation modelled as a function only of solar elevation change correlates with actual remotely sensed observations. As case studies for other tropical regions, three 1-ha plots within French Guiana representing distinct forest structures and topography were selected as study sites (Figure 9). As indicated in Table 4, the northernmost plot, at Paracou, possesses the forest with the greatest mean canopy height, a moderate terrain slope, a high vegetation density as indicated by a high mean Plant Area Index, and the highest degree of canopy closure of the three sites. In contrast, the plot at Nouragues was the site with the largest gap fraction, but with a mean canopy height of over 24m, a similar PAI to the plot at Paracou, and it lies on largely flat terrain. The plot at Itoupé is on an eastern-facing slope of a mountain almost 800m above sea level, and possessed the highest terrain slope of the three plots. Its forest also possessed the lowest vegetation density of the three plots, and its estimated mean canopy height was not as high as the other two plots. Overall, the three sites all possess rough canopies, albeit with a gradation from Paracou to Nouragues and Itoupé as indicated by two particular metrics, the high mean canopy height slope, and the standard deviation of the canopy surface elevation (see Table 4).

Table of contents :

General Introduction 
I. Context
II. Study Objectives
III. Geographic scope and available data
IV. PhD thesis structure
Methods
I. Overview 
II. Evaluating the spatial patterns of variation of forest type distribution 
Bi-directional effects and estimates of spatial variation of forest types
Effects of seasonality on mapping the spatial variation of forest types
III. Evaluating the temporal patterns of variation of tropical forests 
Assessing seasonal variation of vegetation indices at the regional level
Assessing seasonal variation of vegetation indices at the plot level
Assessing seasonal variation of vegetation indices by forest type
IV. Examining patterns of spatio-temporal variation for forest type mapping 
Chapter 1: Spatial variation of forests in French Guiana 
Overview of the chapter
Introduction
Results
Discussion
Chapter 2: Temporal variation of forests in French Guiana 
Overview of the chapter
Introduction
Results
Discussion
Chapter 3: Patterns of spatio-temporal variation in tropical forests 
Introduction
Results
Discussion 
General discussion
I. Summary of findings
II. Contextualizing the findings
III. Study implications
IV. Limitations
V. Perspectives on future research
VI. Overall scientific contribution
References

GET THE COMPLETE PROJECT