The pairwise IOAs
) testifies that seasonal variations agree reasonably well across all MODIS products. Peak values are reached when combining different collections of one and the same sensor, indicating that the seasonal component remained largely untouched by the recent version update to Collection 6. In our opinion, the slightly larger deviations between Terra-MODIS and Aqua-MODIS originate from the different times of satellite overpass associated with markedly different atmospheric conditions. Duane et al.
], for instance, emphasize that cloud formation in the area is subject to pronounced diurnal forcing, with cloud dissipation during the morning (Terra overpass; i.e.
, higher chance of a cloud-free view) and reformation in the afternoon (Aqua overpass).
The considerable decline in IOAs
when in conjunction with NDVI3g
can best be explained on the basis of Figure 1
b,c. The depicted 8-km datasets agree reasonably well across the savanna plains in north-eastern and south-western direction but reveal increasing dissimilarities when approaching the mountain. While the latter phenomenon seems to be of particular severity along the southern and western mountainside, the underlying reasons for the identified spatial gradient are possibly twofold.
Firstly, the degree of association per grid cell depends on an unambiguous seasonal signal. The savanna plains reveal such a clear phenological cycle since they are strongly linked with the bimodal annual rainfall distribution [51
]. For the entire north-eastern quarter of the study area, for instance, Detsch et al.
] reported long-term seasonal amplitudes in the range of
between the long rains and the pronounced dry season. Our interpretation is thereby well in line with Fensholt and Proud [6
] who, in a global context, confirmed a good agreement between Terra-MODIS and GIMMS NDVI in areas with a distinct phenological cycle. The montane forest belt, by contrast, shows a constantly high intra-annual NDVI with small-scale variations being attributable to signal noise rather than natural fluctuations [52
], and thus has a dampening effect on the level of association.
Secondly, we argue that the IOAs
is associated with the degree of sub-pixel heterogeneities. In the north-eastern and south-western savanna plains, narrow quantile bands indicate a mostly homogeneous pixel composition. The NDVI signal of both datasets is thus fed by one dominant land cover type which results in the highest IOAs
observable in the study area. In the north-western lowlands and when approaching the mountain, however, the landscape becomes more heterogeneous as a consequence of intense land management in the foothills of Mt. Kilimanjaro encompassing e.g., widespread coffee plantations [53
], intensely managed Chagga homegardens in the South [18
] and broad agroforestry systems in the West [11
]. When moving further uphill along the climatic gradient [14
], the mountain’s geometry leads to an under-representation of the upper, spatially less extended regions in the coarse 8-km grid. Considering such small-scale transitions between predominant land covers, which particularly applies to the fragmented foothill areas, spatial resampling of the 1.1-km LAC input data to the regular 8-km GIMMS grid leads to a dilution of the unique, yet small-scale NDVI signals and the seasonality associated therewith. The same applies for spatially resampled NDVIAqua-C5
CMG data which, possibly due to the differing input data and aggregation method, considerably deviates from NDVI3g
in the more heterogeneous areas. The 250-m MODIS products, on the other hand, preserve such small-scale signals and hence range on a considerably higher level of association.
4.2. Long-Term Monotonic Trends
The largest differences between “significant” and “conclusive” trends become evident from NDVI3g
applies, trends largely occur rather isolated in the outer rim of the study area, where none of the other products indicates relevant results. Moreover, the sharply outlined patches in the north-eastern and western foothills (Figure 2
c–f) are not reflected by NDVI3g
at all. Considering our tentative example calculation of the FDR depicted above (Section 3.2.3
), we argue that more than half of all “significant” trends derived from NDVI3g
can be expected to represent false alarms. Our interpretation is supported by the non-existing “conclusive” NDVI3g
trends where, according to our calculations, the presence of false positives should be significantly reduced. Therefore, we conclude that (a) “significant” NDVI3g
trends in the study area primarily originate from random chance; and (b) small-scale trends as documented by MODIS cannot be adequately captured by NDVI3g
. Evidence for the latter finding comes e.g., from Yin et al.
] who attributed differences between GIMMS and Satellite Pour l’Observation de la Terre (SPOT) Vegetation long-term trends to the GIMMS-specific resampling procedure from 1.1-km LAC input data to a regular 8-km grid [55
]. Indeed, it seems reasonable to assume that the rather small sample size of
pixels covering the study area cannot adequately capture long-term vegetation dynamics in such a highly heterogeneous landscape [14
]. Nonetheless, it has to be emphasized that such regional-scale findings do not necessarily apply to larger and more homogeneous areas with a smaller amount of sub-pixel heterogeneities associated with NDVI3g
The most evident finding from the remaining MODIS products is the enhanced proportion of browning observable from NDVITerra-C5
. This is not only testified by the depicted relative trend amounts (Table 2
and Table 3
), but also by the calculated MDτ
that indicates a negative shift of all trends identified from NDVITerra-C5
, particularly when
applies (Figure 5
). As mentioned at the beginning, this finding can be attributed to sensor degradation heavily impacting the red and near infra-red Terra-MODIS bands. Analyzing annual NDVI trends over North America, Wang et al.
] demonstrated a very similar predominance of browning derived from NDVITerra-C5
when compared with NDVIAqua-C5
. The same phenomenon has only recently been evidenced for the global scale and with particular severity identified over humid and dry-subhumid zones [56
Our interpretation is further supported by the markedly smaller portion of browning trends in NDVITerra-C6
that possibly results from the adjusted band ageing calibration approach taken by the new Collection 6 product algorithm [9
]. It remains beyond the scope of this study, however, whether the massive greening peaks found for NDVITerra-C6
(as seen from the density distributions in Figure 2
and Figure 4
) indicate a possible overcompensation of Terra-MODIS sensor degradation in Collection 6. The finding that comparably large deviations do not become evident from the two Aqua-MODIS products might reinforce such an “overcompensation theory” impacting—and markedly raising—the number of Terra-based greening trends.
Apart from such tentative considerations and while the determined greening-to-browning ratios remain constant across the applied significance levels, the absolute amounts of MODIS-based trends sharply decline from the “significant” to the “conclusive” domain. Considering the above FDR estimates of
associated therewith, we argue that “significant” trends cannot be unambiguously divided into real discoveries (i.e.
, true positives) or random chance (i.e.
, false positives). Such a differentiation is far more reliable when dealing with “conclusive” trends of which an estimated proportion of
can be assumed to reflect real conditions. Therefore, we recommend future studies to employ a significance level of
in order to deduce reliable long-term trends from spatially discrete time series based upon the Mann–Kendall test or, to stick with Colquhoun [44
] (p. 12),
“If you want to avoid making a fool of yourself very often, do not regard anything greater than as a demonstration that you have discovered something. Or, slightly less stringently, use a three-sigma rule.”
As an alternative, a sufficiently reliable distinction between real and randomly introduced trends can also be achieved by involving contextual information from neighboring locations, which is particularly valid for shorter time scales and noise-prone data. Neeti and Eastman [36
], who developed a so-called “Contextual Mann–Kendall test”, pointed out that for a given trend location, a higher degree of confidence could be achieved when similar trends occurred in adjacent locations—or, to put it the other way round, a spatially isolated trend is more likely to originate from either random chance or the effects of insufficiently masked cloud remnants when no neighboring trends are observable. Since such contaminants have been sufficiently accounted for (Section 2.2.1
) and, more importantly, the applied Whittaker smoother [26
] largely eliminates the effects of non-captured clouds by downweighting explicitly small values, we argue that the chosen Mann–Kendall approach, which takes into account one-dimensional pixel time series only, proves sufficient for the purposes of this study.