**3. Results**

#### *3.1. Seasonal Profiles and Phenology of SDTF Studied Sites*

The interannual variability did not differ among sites for the SOS (X<sup>2</sup> = 1.0; df = 2; *p*-value = 0.606) nor the EOS (X<sup>2</sup> = 1.3; df = 2; *p*-value = 0.520). The variability was higher for the SOS than the EOS at MSD (X<sup>2</sup> = 10.6; df = 1; *p*-value = 0.001) and BP (X<sup>2</sup> = 11.1; df = 1; *p*-value = 0.0008) but not for SD (X<sup>2</sup> = 0.8; df = 1; *p*-value = 0.363). The LOS differed among sites (F2,54 = 4.1; *p*-value = 0.02), with SD presenting a shorter LOS than BP (*p*-value = 0.01) but with no differences between MSD vs. BP (*p*-value = 0.37) and MSD vs. SD (*p*-value = 0.28). The amplitude of EVI also differed among sites (F2,54 = 5.2; *p*-value = 0.008), with MSD presenting lower amplitude than BP (*p*-value = 0.006) but with no differences between SD vs. BP (*p*-value = 0.52) and MSD vs. SD (*p*-value = 0.10).

The monthly values of the seasonal EVI profile for each ecoregion studied using 20 years of time series are presented in the box plot of Figure 2. The continuous line on the boxplot indicates the median of the monthly values for 20 years of the EVI time series. It can be seen in Figure 2 that the median values are always between 0.2 and 0.5. Despite the similarity in the amplitude of the EVI values, the graph of the median of the monthly values has different temporal behaviour for the different ecoregions studied. The maximum values on the MSD experimental site are similar in February and March. The monthly median presents a well-defined maximum value for the SD experimental site, occurring in April. The maximum EVI values occur in May and June at the BP experimental site. For MSD and SD studied ecoregions, the minimum values occur between six and seven months after the maximum values are observed, while for BP, it happens after four to five months.

Although the graphs with the monthly EVI values presented in Figure 2 understand the temporal behaviour, they do not reveal as many characteristics about the vegetation as the graphs with the phenological metrics presented in Figure 3. When analysing the SOS, it is noticed that the BP and MSD experimental sites present a greater interquartile range when compared to the SD experimental site. For most observations, the SOS from the SD experimental site has taken place between January and February. On the MSD experimental site, it is observed that SOS occurs most frequently between November and January. At BP, it is observed that the highest frequency of SOS occurs between January and May (117 days).

Interestingly, the interquartile range observed for the SOS is not for the EOS and LOS metrics for the MSD and BP experimental sites. For the EOS and LOS, it was observed that the MSD and BP sites have a smaller interquartile range than the SD experimental site, with interquartile range values of less than 30 days for both situations. For the SD experimental site, the variation in the interquartile range in the EOS is similar to that observed for SOS, with a variation slightly greater than 30 days interval and the LOS reaching an interquartile range of 60 days. A relationship of the amplitude with the other metrics, SOS, EOS, and LOS, was not observed. The amplitude's highest values were observed for the BP and SD experimental sites, with 25% of the observations above 0.46. The amplitude values for the MSD experimental area had 75% of the observations below 0.31.

**Figure 2.** Monthly EVI values for 20 years of time series in (**A**) MSD ecoregion experimental, (**B**) SD ecoregion experimental, and (**C**) BP ecoregion experimental.

**Figure 3.** Caatinga phenological metrics of the three ecoregions studied, (**A**) SOS, (**B**) EOS, (**C**) LOS, and (**D**) Amplitude.

## *3.2. Environmental Drivers*

Due to the high number of deciduous species, the SDTF, such as the Caatinga, presents high variability in plant biomass in the annual cycle. Figures 4–6 present the environmental drivers' boxplots that may influence the triggers of phenological changes in the Caatinga. When observing the shape of the graphs of the median behaviour of each environmental driver, it can be seen that there is a similarity in the water deficit, precipitation, and soil moisture behaviour. Pearson's correlation coefficient for the environmental drivers and time series EVI is presented in Figure 7. Figure 7 only presents the lag (0–3 months) environmental drivers with the highest Pearson's correlation. With the application of Pearson's correlation, this study allowed for analysis of each environmental driver and identified their action times on vegetation. These results reinforce the importance of precipitation as the environmental driver that best reflects plant biomass production in the Caatinga. Thus, it is the environmental driver that presents the highest coefficient correlation among all: MSD (r = 0.7258; *p* < 0.05, lag = 1), SD (r = 0.8267; *p* < 0.05, lag = 1), and BP (r = 0.7546; *p* < 0.05, lag = 1). Subsequently, water deficit and soil moisture had the highest correlation values. The water deficit's correlations were: MSD (r = −0.6, *p* < 0.01, lag = 0); SD (r = −0.79, *p* < 0.01, lag = 0); BP (r = −0.69, *p* < 0.01, lag = 1). Soil moisture is one of the main environmental triggers of the Caatinga, and in shallow soils, there is a tendency to saturate and also dry out more quickly, not allowing, in many cases, a long-term response from the vegetation: MSD (r = 0.52; *p* < 0.05, lag = 0), SD (r = 0.69; *p* < 0.05, lag = 0) and BP (r = 0.52; *p* < 0.05, lag = 0). Pearson's correlation coefficient between the EVI and air temperature time series showed the weakest but most significant correlation. While MSD and SD presented positive relations between EVI and temperature (r = 0.6; *p* < 0.05, lag = 2) and (r = 0.59; *p* < 0.05, lag = 3), BP (r = −0.56; *p* < 0.05, lag = 0) showed a negative association.

**Figure 4.** Environment drivers for MSD in ( **A**) Water Deficit, (**B**) Precipitation, ( **C**) Air Temperature, and ( **D**) Soil moisture.

**Figure 5.** Environment drivers for SD in (**A**) Water Deficit, (**B**) Precipitation, (**C**) Air Temperature, and (**D**) Soil moisture.

**Figure 6.** Environment drivers for BP in (**A**) Water Deficit, (**B**) Precipitation, (**C**) Air Temperature, and (**D**) Soil moisture.

**Figure 7.** Correlation between EVI and environment drivers for MSD, SD, and BP in water deficit (**A**–**C**), precipitation (**D**–**F**), air temperature (**G**–**I**), and (**D**) Soil moisture (**J**–**L**). The lag months have the highest (r2) between the environmental drivers and the EVI monthly time series.

In Pearson's correlation, environmental drivers were analysed one by one. However, the observed behaviour may not reflect the actual effects on vegetation represented by the EVI time series. There may be an association of impact between the analysed environmental drivers. Partial correlation analysis is a way to solve this problem, allowing the analysis of multiple variables. The partial correlation analysis between the environmental drivers and EVI is presented in Table 1 with five scenarios. Scenario 1 considers all environmental drivers to calculate partial correlations, and in the other scenarios, the effect of one of the drivers is retained. Precipitation and soil moisture always showed a positive partial correlation. In contrast, the water deficit showed negative correlations, with r values ranging from −0.25 to −0.62. Only at the BP site, the partial correlation with temperature was negative. In scenario 1, with all environmental drivers, precipitation was the environmental driver that presented the highest correlation, with r values ranging

from 0.48 to 0.51. The other environmental drivers are better perceived when precipitation is removed from the analysis (scenario 2). In scenario 2, the highest correlations were for temperature in MSD (0.46), water deficit and soil moisture for SD (0.40), and water deficit for BP ( −0.62). This dependence relationship becomes evident when the drivers with the highest partial correlation identified in scenario 2 are removed from the analysis. Thus, the highest values of r for precipitation are seen in scenarios 5, 3, and 4 for MSD, SD, and BP sites, respectively.

**Table 1.** Partial correlation coefficients between the EVI and environment drivers across MSD, SD, and BP sites. Scenario 1 considers all environmental drivers to calculate partial correlations. For the other scenarios, the effect of one of the variables is removed: scenarios 2 (without precipitation), 3 (without soil moisture), 4 (without water deficit), and 5 (without temperature). "-" indicates that the environment driver was not used to calculate the partial correlation. Only for statistical significance (*p* < 0.05). Missing estimates (NS) are not significant.

