Going beyond the green: senesced vegetation material predicts basal area and biomass in remote sensing of tree cover conditions in an African tropical dry forest (miombo woodland) landscape

Our study area is located in central Tabora Province, western Tanzania, in the northeastern miombo woodlands (5.0 °S, 32.8 °E) (figure 2). Tree covered areas span dense forest and woodland areas (canopy cover 50%–100%), areas of sparse canopy cover (< 20%) in lower-lying seasonal drainages (locally called mbuga), and agricultural areas largely but not completely cleared of tree cover (canopy cover <10%) (figures 3(b)−(d)) (Frost 1996). Mean annual temperature is 23.9 °C and mean annual precipitation is 770 mm (2000–2009), with 90% falling between November and May. In 2012, Tabora had a population of 2.3 million people and experienced high rates of forest turnover and net deforestation from the mid-1990s–2010s typical of many miombo landscapes during this time. Forests in the study region fall under a variety of management regimes, including nationally managed reserves (Igombe Forest Reserve) and village-managed forest areas (figure 2(a)). Of standing regional forest cover in 2011, an estimated 16% consisted of regrowth since 1995 (Mayes et al 2015).

Zoom In Zoom Out Reset image size

Zoom In Zoom Out Reset image size

We measured tree stem structure and land cover features across 18 woodland sites spanning a forest regrowth chronosequence of ages 3 yr to maturity in the early to mid-dry season (July 2015) (figure 3(a), supplementary figure S1). At this time, agricultural fields are harvested, while most trees retain moderate canopy leaf cover, which maximizes the landscape-scale contrast in land cover materials across gradients of tree cover. We chose regrowth field sites using a combination of remote sensing, field visits and interviews with local people. All sites had > 3 yr since fire, prior history of maize cultivation, flat topography (slope < 2%) and sandy soils (texture > 70% sand). Of accessible, intact mature forest sites, three were under national jurisdiction of the Tanzania Forest Service in Igombe Forest Reserve and three were in village-managed areas (figure 3(a)).

Field site sizes and locations were established balancing needs to measure large enough areas to relate to 30 m resolution Landsat data (Mayes et al 2015, Elmore et al 2000), adequately sample representative distributions of forest structure across space (Woollen et al 2012), and find comparably sized areas across the landscape without severe recent disturbance. We chose a 0.25 ha field site design (50 × 50 m) used previously to quantify carbon stock accumulations across miombo forest chronosequences in Mozambique (Williams et al 2008). Sites were located as far as possible from boundaries such as roads or field edges. We measured tree stem structure, canopy and land cover in a spatially aligned manner within four circular sub-plots of 10 m radius, located using a stratified random sampling design in each quadrant of field sites (Williams et al 2008).

For tree cover structure, we inventoried stem basal diameter and diameter-at-breast-height (dbh) at 1.3 m for all trees > 5 cm dbh in the sub-plots. Allometric relationships specific to the miombo and global dry tropics were used to convert stem sizes to biomass (table 1). Basal area and biomass were summed across all plots and normalized to the sum of sub-plot areas (∼ 1/8 ha) to calculate tree basal area and biomass density (Williams et al 2008).

For canopy structure, we recorded two types of data at circular sub-plot centers, and at 5 m and 10 m positions along three radii at 120 degrees to each other for a total of seven data points per sub-plot. First, we measured plant-area index (PAI) data using an LAI-2200 photosensor (LiCOR, Lincoln, NE) at 1.3 m height. Second, we took spherical densitometer (grid-etched convex hemispherical mirror) surveys of cover conditions, following a classical United States Forest Service method (Strickler 1959, Korhonen et al 2006), accounting for four cover types (green leaf, senesced leaf, woody material and open sky (no material)) (see supplementary text 1-ST1 available at stacks.iop.org/ERL/12/085004/mmedia). For ground cover data, at the same seven points where canopy data were taken, we used a modified point-line transect method and tallied, summed and averaged proportions of the following land cover materials: leaf litter, green grass, senesced grass, woody litter, roots/stems, charred organic matter, ash, and bare substrate. The point-line transect method has been used widely in measuring land cover in semi-arid regions and has been shown to scale successfully between ground and satellite data (Pilliod and Arkle 2013, Elmore et al 2000).

We acquired six Landsat 8 OLI scenes from June–July 2015 from the NASA/USGS ESPA system covering Tabora Province (paths 170-2, rows 63–65), which were pre-processed to surface reflectance using the L8SR algorithm (Vermote et al 2016). The image mosaic was subset to a 2750 km² study area encompassing field sites (figure 2(b), figure 3(a)). Using the six visible, near and shortwave infrared optical bands with 30 m resolution, we conducted spectral mixture analyses (SMA) using field-validated endmembers (EM) to model land cover (Mayes et al 2015). SMA models pixel reflectance values as linear multiplicative sums of the reflectance of known materials, subject to constraints that EM fractions must equal 1 and there are no negative values, and acts as a physical modelling technique that estimates the proportion of known materials that comprise landscape surfaces (Adams and Gillespie 2006, Adams et al 1995). We used a fixed-endmember SMA approach, using single endmembers for four materials consistent with categories used for canopy and ground cover data: green vegetation (GV), non-photosynthetic vegetation (NPV), substrate and shade, which were chosen from georeferenced field sites with high proportions of these materials, and successfully used for forest change mapping in Tabora, Tanzania (Mayes et al 2015) (supplementary figure S2). Outputs of SMA are images with pixel values from 0–1 representing modeled proportions of endmember materials comprising land cover surfaces. We calculated NDVI images to compare how this commonly used greenness-based measure for land cover performed against SMA for representing miombo tree cover structure. To deal with image variability from scene to scene, we normalized EMs and NDVI to scene summary statistics by calculating Z-scores on a per-pixel basis.

We assessed patterns among field data on forest stem structure and land cover (canopy and ground cover) and satellite data across forest sites and site age groups, using analysis of variance (ANOVA), Tukey's honest significant difference means comparisons (p < 0.05) and ordinary least-squares linear regression. Given log-normal distributions of basal area and biomass, we developed log-linear models to predict tree structure (natural log-transformed basal area, biomass) from satellite data using three Landsat-derived indices: two of the SMA output images (Landsat SMA-NPV and SMA-GV) and NDVI. Landsat SMA-NPV, SMA-GV and NDVI values were extracted for 3 × 3 pixel regions chosen over GPS points for field sites. Of the 18 field sites, 17 were used to develop linear models—one site (S07, mature village forest) showed evidence of moderate burned area on the ground that confounded its SMA results (supplementary table S1) and relationships between stem structure and land cover. Using these models, we mapped basal area and biomass maps for the 2750 km² (275 000 ha) training region (figure 3(a)).

We evaluated the performance basal area and biomass maps based on non-green (Landsat SMA-NPV) and green land cover (Landsat SMA-GV, Landsat-NDVI) metrics in three ways. First, we back-compared satellite-based estimates of tree structure to field site data. Second, we compared map estimates of average basal area and biomass for forest and non-forest land cover across the training area, delineated by a Landsat SMA-based thresholding method developed for mapping miombo woodland land cover changes (Mayes et al 2015), and compared forest vs. non-forest structure estimates to other published miombo forest data. Third, we compared the performance of map biomass estimates across regions of dense, sparse and low tree cover conditions, shown in figures 3(b)−(d), using 38–41 image validation points from two subset regions: a nationally managed forest reserve (Igombe Reserve) and a village-managed regions (Millennium Villages Project communities), shown in figure 3(a). We conducted analysis of variance (ANOVA) for tree biomass estimates with tree cover type (dense woodland, mbuga and cleared) and remote sensing index as independent variables, with the goal of assessing the degree to which NPV (SMA-NPV) and greenness-based models (SMA-GV, NDVI) could differentiate tree structure across known landscape-scale gradients.

Forest stature, defined by various indices (basal area, biomass, stem density, height, volume), all increased with site age (p < 0.05, figure 4, table 2, all ANVOA F-values in table S1). Basal area differed significantly among all regrowth site age classes, ranging from 1.04 ± 0.37 m² ha⁻¹ (mean ± standard error) at 3–4 yr sites to 29.4 ± 6.03 m² ha⁻¹ at village-managed mature forest sites (figure 4(a), table 2). Biomass varied from 1.7 ± 0.7 Mg ha⁻¹ at 3–4 yr sites to 99.7 ± 27.7 Mg ha⁻¹ at village forest sites. (figure 4(b), table 2). Though differences were not statistically significant, in site-level data, village-managed forest field sites had strong trends of higher stem density, basal area and biomass compared to nationally managed forest sites (table 2, figure 4).

Zoom In Zoom Out Reset image size

Total plant canopy area measured by photosensor (plant-area index, PAI, ratio m² m⁻²) increased with site ages, from 0.308 ± 0.163 at 3–4 yr sites to 1.16 ± 0.382 at mature village sites (p < 0.05, ANOVA F-values in table S1) (figure 5(a)). Conversely, the proportion open-sky area of the canopy, tallied from spherical densitometer survey (units of proportional canopy area), decreased significantly with increasing site age from 0.699 ± 0.105 at 3–4 yr sites to 0.187 ± 0.050 at village-managed mature forest sites (figure 5(c)). Trends in canopy material fractions with site age were moderately positive for woody material, and were weak for green leaf material (figures 5(b) and (d)). Below these canopy structure changes, ground cover materials were dominated by leaf litter and senesced grass fractions across all sites (>0.6) (figure 4(d)).

Zoom In Zoom Out Reset image size

In continuous relationships between canopy cover and forest structure across all sites, PAI related positively to basal area (r² = 0.815, p < 0.01) and biomass (r² = 0.635, p < 0.01) (figure 6(a)). Canopy open sky fraction related inversely to basal area (r² = 0.763, p < 0.01) and biomass (r² = 0.552, p = 0.002) (figure 6(b)). No significant relationship existed between the proportion of green leaf area in the canopy to basal area (r² = 0.105, p = 0.103) or biomass (r² < 0.01, p = 0.330) (figure 6(c)).

Zoom In Zoom Out Reset image size

Landsat SMA successfully modeled land cover with low root-mean square error (< 0.01) and greater than 90% of pixels with reasonable endmember proportions (table 3). The SMA-NPV fraction comprised the highest mean proportion of land cover materials per pixel (0.40 ± 0.15) and showed the highest sensitivity to varying land cover mixtures (table 3). Mean per-pixel green vegetation fractions of land cover were lower (0.15 ± 0.07) and less variable.

Landsat SMA-NPV metrics had equal or stronger relationships than greenness metrics to field data on land cover and stem structure. Landsat SMA-NPV had a significant positive linear relationships with open-sky fraction of the canopy (r² = 0.632, p < 0.001), consistent with increasing exposure of NPV ground cover (figure 7(a)). Positive linear relationships occurred but were not as strong between canopy PAI and Landsat SMA-derived GV (r² = 0.514, p < 0.001) and NDVI (r² = 0.484, p < 0.001) (figures 7(b) and (c)). Inverse correlations of Landsat SMA-NPV to basal area and biomass were stronger than direct correlations of stem structure greenness metrics (Landsat SMA-GV and NDVI) (figure 8). Log-linear regression models showed significant relationships among scene-normalized (Z-score) SMA-NPV and green satellite metrics (Landsat SMA-GV and NDVI), and tree basal area and biomass (table 4). The predictive models of forest structure based on satellite NPV metrics were stronger than models based on greenness metrics.

Zoom In Zoom Out Reset image size

Zoom In Zoom Out Reset image size

In Landsat-derived maps, > 90% of pixels were modeled within published ranges for basal area (0–35 m² ha⁻¹) and miomass (0–200 Mg ha⁻¹) in miombo forests (supplementary table S2). In the first map performance assessment, back-comparison of map biomass estimates to field site data, SMA-NPV maps estimated biomass with mean absolute value errors of 12 Mg ha⁻¹ across all sites, 30% lower than mean absolute errors for SMA-GV (19 Mg ha⁻¹) and NDVI maps (18 Mg ha⁻¹) (table 5). All metrics tended to overestimate biomass at 3–4 and 10–24 yr regrowth site classes, with SMA-NPV metrics showing the closest estimates to field values (supplementary tables S3–S4). The magnitudes of overall average satellite map errors relative to field data were larger than the field biomass values for 3–4 and 10–24 yr sites, but smaller than regrowth and mature forest stand biomass.

In the second assessment of map performance, comparison of mean forest and non-forest area stem structure estimates at landscape scales, both SMA-NPV and greenness metric-based maps estimated mean forest cover basal area and biomass to be 300% larger than non-forest land cover areas (table 6, figure 9). For forest cover, SMA-NPV estimates of average basal area (16.0 ± 9.55 m² ha⁻¹) and biomass (46.1 ± 43.0 Mg ha⁻¹) were higher than greenness metrics (basal area < 13 m² ha⁻¹; biomass < 36 Mg ha⁻¹) (table 6). For non-forest, the SMA-NPV-based maps' estimates for basal area (4.23 ± 4.01 m² ha⁻¹ ) and biomass (7.14 ± 11.2 Mg ha⁻¹) were in the middle of the range of field-measured basal area and biomasses for 3–4 and 10–24 yr regrowth sites (table 2).

Zoom In Zoom Out Reset image size

In the third assessment, comparison of map biomass estimates across known gradients in tree structure at image validation points, biomass varied significantly by tree cover type and remote sensing index (table 7). The SMA-NPV maps significantly differentiated biomass among all woodland types in both national reserve and village-managed landscape areas (figure 10, supplementary table S5). The greenness indices differentiated biomass between dense woodlands and other types, but not between sparse mbuga and cleared areas (figure 10, table S5).

Zoom In Zoom Out Reset image size

Relationships among field-measured stem and land cover features yielded practical new insights for optical remote sensing of stem structure in an African dry tropical woodland landscape. Canopy cover-mediated exposure of NPV ground cover materials emerged as a land cover pattern that related strongly to tree structure in the early dry season. There was an inverse relationship of high NPV ground cover exposure to the sky at low basal area/biomass sites, and low NPV exposure at high basal area/biomass sites. Gradients in total canopy cover, illustrated by large ranges in PAI (< 0.4–1.2 m² m⁻²) and densiometer open-sky fraction data (>80% to 20%), were larger and related much more consistently to stem structure than the patterns of any single canopy component such as green-leaf material (varying from less than 10% to about 40% in densiometer measurements) or woody material. There was no significant relationship between tree structure and the proportion of green leaf cover in the canopy at this time of year, after the onset of senescence. This result makes sense in the context of ecological processes affecting deciduous forest green leaf cover beyond stand age or disturbance. Green leaf cover alone is affected by many factors that could confound its relationship to basal area and biomass in the early dry season, including variable phenology across tree species (Frost 1996), and landscape-driven edaphic factors such as topographic position and soil moisture (Fisher et al 2006).

We used land cover relationships to tree structure in the field to test the scalability of our hypotheses to Landsat satellite-derived metrics. Strong inverse correlations among satellite-derived SMA-NPV fractions to basal area (r = −0.85) and biomass (r = −0.86) at spatial scales of about 1 ha (3 × 3 30 m pixel groups averaged over field sites), were driven by the net land cover signal of decreasing ground NPV exposure with increasing canopy cover, which correlated directly with basal area and biomass. Inverse relations of satellite- NPV to basal area (r² = 0.72) and biomass (r² = 0.75) were stronger than direct relationships of Landsat SMA-GV fraction and NDVI to stem structure (r² < 0.70). These results indicate a strong potential for the use of non-green land cover indices, together with greenness-based land cover metrics, to model tree basal area and biomass across African TDF landscapes. Other recent studies have shown the value of using not only multiple non-green and green land cover metrics, but also ancillary data on landscape variables to map aspects of tree structure across African forests (Hansen et al 2016, Halperin et al 2016). Strong total canopy cover relationships to basal area and biomass also suggest that emerging radar sensors, such as Sentinel-1, may hold promise for mapping miombo and other African TDF biomass in combination with optical satellite data.

All senesced and green vegetation metric-based maps accurately estimated landscape-scale ranges of tree basal area and biomass in comparison to published miombo woodland tree inventories (basal area 0–35 m² ha⁻¹; biomass 0–150 Mg biomass ha⁻¹ or 0–70 Mg ha⁻¹ biomass carbon (C) ha⁻¹) (Frost 1996, Williams et al 2008, Kalaba et al 2013a). All metrics also captured large differences in tree structure between forest and non-forest land cover, >10 m² ha⁻¹ in basal area and > 30 Mg biomass ha⁻¹. These results indicate that SMA-NPV, SMA-GV fraction images, or NDVI-based maps would all be capable to tracking miombo tree structure trends associated with large changes (e.g. deforestation) landscape scales. Similar approaches may be transferrable for coarse landscape-scale forest structure monitoring in other global TDFs, or using other optical remote sensing data sources such as the new Sentinel-2 sensors.

Together, the first map assessment of back-comparison to field site measurements and the third map assessment involving image validation points both suggested the SMA-NPV metric estimated above-ground tree biomass values more accurately and with greater sensitivity to variability across different tree cover types than the greenness metrics, particularly for low-biomass (< 30 Mg ha⁻¹) regions. Using a guideline of only estimating biomass variability or change at twice or more than the magnitude of a metric's error or uncertainty, the SMA-NPV metric could in theory resolve biomass variability or change of about 25 Mg ha⁻¹, given its overall remotely sensed-field biomass mean absolute value differences of 12 Mg ha⁻¹ across all sites. With mean absolute value differences in map-field biomass estimates of > 18 Mg ha⁻¹, the greenness measurements would have a resolving power of detecting biomass variability or change at rates closer to 40 Mg ha⁻¹. The image validation point analyses further demonstrated the advantages SMA-NPV metrics showed over greenness metrics. From past miombo woodland research, sparse woody biomass at the periphery of mbuga regions in miombo landscapes has been measured at 15–30 Mg ha⁻¹(Frost 1996, Woollen et al 2012). Those values are intermediate between the 10–24 yr (10 Mg ha⁻¹) and 30–40 yr regrowth forest sites (40 Mg ha⁻¹) we measured. In both national reserve and village-managed landscapes, the SMA-NPV metric estimated similar mean biomass around mbuga peripheries at 24 Mg ha⁻¹, significantly different from mature forests (>80 Mg biomass ha⁻¹) and cleared areas (<5 Mg ha⁻¹). In contrast, the greenness metrics underestimated mbuga periphery woody biomass of about 5–10 Mg ha⁻¹, which does not differentiate mbuga from cleared area biomass. These comparisons indicate that the SMA-NPV senesced vegetation metric showed better accuracy and precision for modeling tree biomass relative to the green vegetation metrics, particularly in intermediate to low-biomass areas.

Given these performance assessments, the Landsat-scale tree structure maps we present may be useful for a number of forest and tree structure monitoring applications. These include monitoring biomass changes not only from clearing, but also from degradation of mature to 30–40 yr regrowth forests (biomass > 40 Mg ha⁻¹) by harvests for timber or wood-fuel production, which can occur at scales of > 25 Mg ha⁻¹ biomass removals (Luoga et al 2002). Another application is studying how governance regimes, land tenure structures or management plans relate to tree structure. For example, the scales of difference in tree structure variability we measured on the ground between village and nationally-managed forest areas (biomass > 20 Mg ha⁻¹; basal area >10 m² ha⁻¹) were detectable with the Landsat SMA-NPV indicators. The SMA-NPV indices also show promise over greenness metrics for monitoring biomass changes of smaller magnitudes in lower to moderate tree biomass areas (20–30 Mg ha⁻¹) associated with agroforestry or forest restoration projects. Given the uncertainties, however, none of the remote sensing approaches could be considered accurate for estimating stem structure variability or change below or between ranges of the 3–4 and 10–24 yr regrowth field sites (<10 Mg ha⁻¹ or basal area < 10 m² ha⁻¹).

This study highlights three key opportunities for further work in quantitative optical remote sensing to monitor tree cover in TDF landscapes. First, there is need for more field data on tree structure and land cover in African dry tropical landscapes to enable more rigorous calibration and accuracy assessment of remote sensing products. Second, there is a need to refine remote sensing methods to monitor smaller biomass changes and tree structure variability in low-biomass regions. Other SMA approaches that characterize land cover at sub-Landsat pixel scales, such as multiple-endmember spectral mixture analyses or use of automated SMA methods such as the Carnegie Landsat Analysis System are promising approaches (Roberts et al 1998, Asner et al 2009). Data-fusion approaches merging optical with active remote sensing data (e.g. radar, LiDAR) have proved effective for mapping tree structure in African TDFs and across the global tropics for larger areas (Ribeiro et al 2008, Baccini et al 2012, Saatchi et al 2011), and may also be useful. The availability of Sentinel-1 radar data at 10 m globally since 2014 presents strong opportunities for integrative analyses with optical image archives. High resolution (< 1 m) remote sensing observations of tree structure, such as those from small drones, may also prove useful for improved scaling between field and satellite data (Tang and Shao 2015, Paneque-Galvez et al 2014). Three, further integration of repeat local community monitoring and scientific research on land cover-forest structure relationships is an opportunity to improve forest monitoring capabilities and inform assessment of local management effects on forest and tree cover condition.