3.1. Catchment Level Peak Season Albedo (pSA), NPP, ET and NDVI
Figure 3 shows the spatial and statistical distribution of the PSA trend, computed as the pixel level slope of PSA regression over the study period for T35B and S50E for both MODIS (
Figure 3A,B,E,F) and Landsat (
Figure 3C,D,G,H).
Although similar spatial patterns are observed, it is clear from
Figure 3C,D, that there are some extreme changes that are not captured at coarser MODIS resolution. This is borne out by the larger range for Landsat displayed on the x-axes in
Figure 3G,H. The slope for MODIS pixels varied between −0.003 (blue pixels) in both catchments with maximum increase of 0.005 for S50E and 0.0026 for T35B (red pixels). Measured from Landsat PSA, greater variation of values between −0.01 (blue pixels) and 0.011 (red pixels) was calculated. Locations where Landsat PSA trend is either higher than the maximum MODIS trend or lower than the minimum trend are indicated with circles in
Figure 3C,D. At catchment scale the mean change (mpc) in PSA was less than one per cent ±10 standard deviations (sd) for MODIS and ±5 sd for Landsat.
Over the study period, mean MODIS PSA values of 0.145 ± 0.011 and 0.150 ± 0.014 were obtained for catchment T35B and S50E respectively, with mean Landsat PSA values significantly lower (
p < 0.05) at 0.143 ± 0.022 for T35B and 0.140 ± 0.022 for S50E. The boxplots in
Figure 4 illustrate mean annual PSA (
Figure 4A,B), NPP (
Figure 4D,E), ET (
Figure 4F,G) and NDVI (
Figure 4H,I) trends for the observed study period extracted from MODIS data. Mean annual rainfall (Agricultural Research Council weather station data, Tropical Rainfall Measuring Mission satellite data) is shown in the bar plot in
Figure 4C. WS 30388 represents the rainfall in S50E at Cala, while WS 30149 represents the rainfall for T35B at Ugie. The linear trend is shown with a dotted line while the LOESS curve indicates the local trend.
While similar spatial patterns were observed for mean MODIS PSA at coarser resolution and mean Landsat PSA, linear correlation between Landsat pixels, scaled to MODIS resolution, only shows an R
2 of 0.718 for T35B and 0.723 for S50E. In addition, the mean PSA in both S50E and T35B did not change significantly over the 15 year study period (
p > 0.05). However by fitting a median based linear model [
70,
71,
72], the S50E slope showed a slight increase (β
1M = 0.00023;
β1LS =
0.0003; p > 0.05), which would cause a net increase of 0.003 (
0.004) in PSA. In contrast, mean PSA trend in T35B was negative with MODIS (β
1M = −00009) but positive with Landsat (
β1LS =
0.0004), translating to PSA change of −0.001 (
+0.006). Non-significant trends at catchment scale were confirmed with a Mann-Kendall (MK) test (
p > 0.05) for both catchments. Mean albedo values and trend were also calculated from the 8-day MODIS product (T35B-σ = 0.135 ± 0.017, β
1M8 = 0.0001; S50E-σ = 0.146 ± 0.001, β
1M8 = 0.00004).
PSA generally followed an increasing trend in response to drop in rainfall, and a decreasing trend in response to increased rainfall, when comparing
Figure 4A,B with
Figure 4C. The high rainfall in 2006, categorised as a flood [
73], caused a drop in PSA reflected in 2006. Although a relationship between albedo and rainfall is suggested, neither the linear, nor non-linear trend (Theil-Sen slope, measured with MK-test) was significant (
p > 0.5) at catchment scale. NPP, ET and NDVI in T35B (
Figure 4) have higher mean values (0.892 kg.C.m
−2; 542 mm.yr
−1; 0.54) compared to S50E (0.802 kg.C.m
−2; 508 mm.yr
−1; 0.49) and are statistically different (
p < 0.05), measured with the Wilcoxon signed rank test for non-parametric data. Although the trends appear strongly related to that of the rainfall pattern in
Figure 4C, there is only a weak negative linear trend (
p > 0.1). Lower NPP, ET and NDVI were noted for 2003 in both catchments confirming the inflection point in 2004 indicated by [
2] associated with extreme low rainfall in 2003 (
Figure 4C). Even though the LOESS curve (in red) indicates a local downward trend, the linear trend is not significant (
p > 0.05) in any of the catchments.
The correlation between mean PSA, NPP, NDVI and ET is reported in
Table 3. Complete cases, where a value existed for each of the four datasets for the pixel in question, were extracted for every pixel within the two catchment extents for comparison. A positive correlation indicates the extent to which one variable e.g., PSA increases or decreases in parallel with another variable, while a negative correlation indicates the extent to which one variable increases as the other decreases.
In both the catchments, the strongest correlation was found between NPP and ET with 0.64 in T35B (n = 2162) and slightly higher at 0.71 for S50E (n = 2407). Correlation between NDVI and ET was ~0.6 in both catchments while NDVI showed a stronger relationship with NPP in S50E. A weak negative correlation was found between PSA, NPP and ET. In T35B, PSA had a weak positive correlation with NPP, but none in S50E. Detail of the correlations computed per land cover class and transition trajectory are provided in
Supplementary Material, Table S1. In contrast to the catchment results, at land cover class and transition level, the strongest correlation was between NDVI and ET (>0.79). Only persistent forest/plantation (n = 42; 0.55) and trajectory deforestation (n = 35; 0.75) in S50E showed a significant correlation between NPP and ET. Intensification of agriculture showed a similar response in both catchments, only the correlation between albedo and NDVI was stronger in T35B (n = 41; −0.54) as compared to S50E (n = 117; −0.45). Contrary to expectation, deforestation in T35B showed a positive correlation (n = 23; 0.7) between albedo and NPP. Afforestation in S50E (n = 6; −0.56) displayed a negative correlation between albedo and NPP, but a positive correlation in T35B (n = 60; 0.63). The aggregated catchment correlation masks some of the per class correlations, resulting in Simpson’s paradox where groups of data show one particular trend, which is reversed when groups are aggregated [
74]. Common in spatial analysis of heterogeneous landscapes, this is an example of MAUP [
28] where the sample size (n) is dictated by the arbitrary land cover aggregation units.
The spatial distribution of the correlation between PSA and each of the variables NPP, NDVI and ET are shown in
Figure 5 for T35B (top) and S50E (bottom). Only significant correlations (
p < 0.05) are symbolised, while p > 0.05 is shown in grey. “No data” values (white) are visible in
Figure 5D,F where the NPP and ET algorithms did not calculate a value for the Ncora dam in S50E. Negative values (brown) show negative correlation where one variable increases as the other decreases. Positive values (green) show positive correlation where variables increase in parallel. Pixels where all three variables are significantly correlated with PSA, are highlighted with blue (+PSA+ET+NDVI+NPP or –PSA-ET-NDVI-NPP) and red (+PSA-ET-NDVI-NPP or -PSA+ET+NDVI+NPP) buffers to indicate the direction of the correlation.
Labels 1, 2 and 3 in
Figure 5 indicate the spatial location of three points where pixel values were extracted to further illustrate the correlation between PSA, NPP, ET and NDVI at local scale, linked to specific land cover trajectories. Point 1 represents an area with high negative albedo trend (
Figure 5A), in contrast to point 3 with a high positive albedo trend (
Figure 5B). Point 2 was selected as the middle ground with almost no trend (
Figure 5B). In the case of points 1 and 3, negative correlation was noted while for point 2 positive correlation was measured between PSA and NPP, ET and NDVI. It is important to note that each of the variables (NPP, ET and NDVI) can show either positive or negative correlation with PSA at different spatial locations.
3.2. Land Cover Trajectories
Published albedo values are compared to similar land covers as those found in the study area (
Table 4).
No persistent bare soil was observed in T35B, while the extent of bare soils and water bodies was too small to extract mean MODIS PSA. Similarly, in S50E, mean MODIS PSA could not be evaluated for bare soils and wetlands. In this study, UG refers to herbaceous vegetation (grassland, savannas and degraded grassland), while in other databases found in literature, such as the CORINNE database [
75], grassland may refer to greener pastures with a lower albedo value. Similarly, in the case of shrublands it is probable that the albedo measured by [
48] are leafier thus having a higher LAI and lower albedo than in this study area. [
75] observed that class names used in land cover classification systems are often descriptive without providing detail on the criteria used to define these classes. Water bodies and croplands fall within the literature ranges, while forest/plantation lies within 0.01 of published values for this land cover class, although lower than reported by [
36].
The percentage area per catchment occupied by persistent land cover classes and transition trajectories and significant PSA change (trend slope p < 0.05), measured using both MODIS and Landsat, are summarised in
Table 5. Significant PSA change is divided into decrease in albedo (negative change) and increase in albedo (
positive change), given both in percentage of catchment area as well as PSA change. PSA change is calculated as the trend slope multiplied by the study period (15 years) to give the expected increase or decrease in PSA per land cover class or transition and is highlighted in light grey. Equally, the detail per land cover class is presented in
Supplementary Material, Table S2 and Table S3.
As expected, with persistent classes comprising 82% of T35B, the mean change (MODIS,
Landsat; −0.001,
0.004) for persistent classes only was similar to that of the entire catchment (−0.001,
0.003). Significant change (9%, 10%) was noted with similar trend directions. Negative trends amounted to a larger negative change to lower albedo values, however the positive change measured with Landsat covered a larger area. For S50E, persistent classes covered 75% of the landscape with a mean change in PSA over the study period of 0.004 measured by both MODIS and Landsat. Although the area mapped as persistent is almost the same among the data sources, the area of significant change (
p < 0.05) is almost double using Landsat to map the change.
Figure 6 illustrates the mean PSA for each persistent land cover class measured with MODIS and Landsat for T35B (A, C) and S50E (B, D) over the study period.
In S50E, persistent urban land cover displayed the highest PSA, measured with either sensor (
Figure 6C,D). In contrast, MODIS PSA in urban land cover (
Figure 6A), showed an anomalous result for T35B as a result of the fragmented nature of the urban class (n = 3;
Table S1), representing only 0.1% (n = 3) of the catchment area (
Tables S1 and S2). The urban sites in this catchment have a longer history of human occupation, and are considerably more woody than rural villages in S50E which are under communal tenure arrangements. Shrubland in T35B shows an unexplained trough between 2002–2006 and 2009–2011 in
Figure 6B. This could be related to variation in rainfall, IAP clearing activities and regrowth.
Transition classes (
Table 1) account for 18% in T35B and 21% in S50E [
4] at Landsat resolution. These transition classes measured with MODIS and
Landsat respectively showed smaller changes in T35B (−0.004,
0.001) compared to S50E (0.007,
0.009). Total area of transition in T35B is almost four per cent larger when measured with Landsat, while there is only two per cent difference in S50E, implying more local scale and fragmented transition in T35B. Between 2000 and 2014, gradual ecological change (woody encroachment, abandonment, degradation and reclamation) caused a positive significant increase in albedo for all Landsat-based classes (
Supplementary Material, Table S2 and S3), however the affected area covers less than 2% of the two catchments. In contrast, when MODIS data was used, only woody encroachment and reclamation caused increases in albedo. Therefore, it is clear that the detail of change in the landscape is not effectively captured using only MODIS data.
Figure 7 illustrates the relationship between the transition classes and PSA from MODIS and Landsat compared with the catchment average PSA (black line).
Degradation, urban intensification, increased cultivation and abandonment all have higher than catchment average PSA. These classes are all associated with increased bare surfaces with higher albedo. Increased cultivation also results in a higher albedo, due to clearing of existing vegetation to establish crops, the fraction of bare ground between standing rows or desiccation in fallow fields. In both catchments, the effect of degradation (De) is much larger when PSA is measured using Landsat, but the percentage is low (0.1% in both catchments). Deforestation (D) shows the expected increase in PSA in S50E, but not in T35B where it follows the afforestion (R) curve, possibly indicative of a classification error in the land cover products.
3.3. Season-Trend Model
The estimated trend and breakpoints from the deconstructed 8-day albedo time series using the STM method [
3], extracted for Points 1, 2 and 3 (
Figure 5) are depicted in
Figure 8. Significant structural breakpoints (95% CI) are indicated by red squares and horizontal red lines. The trend line on 8-day time series, between significant breaks, is added in blue. The significance of the trend line segments are indicated by blue stars to show the
p-value (***
p <= 0.001, **
p <= 0.01, *
p <= 0.05). The slope and significance of the trend line on annual aggregate is added in blue text, with the p-value illustrated with green stars on the trend line.
Trend for Point 1, with persistent forest/plantation (FP) and trajectory afforestation (Ra), shows a significant overall decrease of albedo (p <= 0.001 green *) with three significant breakpoints, each with significant trend (blue *). The overall slope indicates a small but significant negative change. Point 3 indicates the opposite trajectory with Da (deforestation) resulting in an increase of albedo (p <= 0.001). Two breakpoints are indicated with three significant segments (p <= 0.01). Point 2 is an example of persistent grassland (UG) where overall trend shows a very small, insignificant increase. Structural changes occurred at all three points in 2007.
Estimated inter-annual variability (IAV) (i.e., annual anomalies) and seasonality (i.e., mean seasonal cycle) are shown in
Figure 9 for all pixels in the catchments, not only those with significant change. In
Figure 9, the IAV is shown in the left panel, while the seasonal range is shown in the right panel for T35B (top; A, C) and S50E (bottom; B, D).
Over the study period of 15 years, albedo in S50E fluctuated annually with a mean of 0.0041, very similar to the mean of 0.0045 in T35B. However, the IAV for the two catchments were found to be significantly different (
p < 0.001; Wilcoxon rank sum test). The highest frequency of pixels varied with standard deviations (sd) between 0.003 and 0.005. Similarly, the mean seasonal cycle in the two catchments—based on 8-day MODIS albedo values—are significantly different (
p < 0.001; Mann-Whitney U test for non-parametric data). The albedo can vary between 0.01 and 0.08. Distinct spatial patterns are noted in the maps in
Figure 9.
3.4. Modelling ET and NPP
In
Table 6, the area percentage for modelled persistent land cover classes in 2030 are compared with the size of these land cover classes in 2014.
Table 6 also includes the net change, as well as the mean trend calculated from MODIS. Based on the mean MODIS PSA change and relationships with NPP and ET, three scenarios for future NEE and water use were calculated: (1) lower mean albedo indicating proliferation of woody vegetation; (2) mean albedo, the status quo persists; and (3) higher mean albedo, with conversion to agriculture and urban intensification dominating future transitions.
In the higher albedo scenario, the total modelled NEE in 2030 for persistent classes in T35B could reduce by 1% when compared with 2014. Should a low albedo scenario ensue, an increase of more than 80% could be obtained with a catchment mean of 3.2 × 10
6 kg C based on the mean time series NPP. Similarly, water use could decrease by almost three per cent or increase by up to 19% for persistent classes. In T35B, the total change (gain and loss) in the landscape over all land cover classes was 15.5% for modelled period 2014 to 2030 [
20], compared with 18.2% for the period between 2000 and 2014 [
4]. Trajectory labels indicating gradual and abrupt changes are responsible for the difference between persistence and the total modelled NEE and water use in the catchment. Trajectories abandonment, reclamation and degradation increase grasslands, woody encroachment boosts shrublands, increased cultivation, afforestation and urban expansion respectively result in higher croplands, forest/plantation and urban. Afforestation was the strongest modelled trajectory in T35B showing a net gain of 1.5% and a strong negative albedo trend. These changes could produce an additional 0.5–1.1 × 10
6 kg C and 0.3–0.4 Mm
3 ET.
For S50E, the total change over all land cover classes was 23% for the same modelled period [
20]. By comparison, the period between 2000 and 2014 exhibited 21% change [
4], assuming a similar map accuracy for the modelled map. The modelled NEE for persistent classes varies between 2.1 and 4.6 × 10
6 kg C, with modelled water use varying between 1.4 and 1.9 Mm
3. In 2014, these values were 2.5 × 10
6 kg C and 1.8 Mm
3 respectively (
Table 6). Changes to the landscape could account for NEE of 0.7–1.6 × 10
6 kg C and water use of 0.5–0.7 Mm
3. The expected scenario for S50E is increased PSA due to intensification of agriculture, lower NEE and water use depending on which land cover class is replaced.