Tanzania, located in East Africa between 1˚S to 12˚S latitude and 29˚E to 41˚E longitude, exhibits complex topography ranging from coastal plains to the high-altitude regions of Mount Kilimanjaro and the Southern Highlands (Figure 1). This topographic diversity creates multiple climatic zones, from humid coastal areas to semi-arid interior regions, making it an ideal region for studying climate variability ([16]). The September-November (SON) season represents a critical transition period from the dry to wet season, characterized by increasing temperatures and the onset of seasonal rains. The study utilized surface maximum air temperature (Tmax) data from the Climate Prediction Center (CPC) during 1979-2024. This dataset provides comprehensive spatial coverage with sufficient temporal resolution for analyzing interannual variability ([37]). For investigating physical mechanisms, we employed ERA5 reanalysis data ([10]) with the same temporal coverage, including latent heat flux, sensible heat flux, surface longwave (LW) and shortwave (SW) radiation, cloud cover, and atmospheric variables such as wind, humidity, and vertical velocity. Reanalysis data have been widely validated for climate studies in tropical regions and provide physically consistent representations of atmospheric processes ([8]).

2.2.1. Analysis of Dominant Modes for SON-Mean Tmax

Empirical Orthogonal Function (EOF) analysis is employed to identify the dominant modes of interannual variability in SON Tmax across Tanzania, which has been widely used in climate studies ([4]; [35]). The EOF approach decomposes the spatio-temporal temperature field into orthogonal patterns that maximize the explained variance. Prior to the EOF analysis, the seasonal means are calculated from monthly Tmax data and the climatological annual cycle is removed to obtain anomalies. The covariance matrix is used for the EOF decomposition to preserve the physical units and magnitude of temperature variability. The significance of separated EOF modes is evaluated using [12] rule of thumb, ensuring that the reported modes represent distinct signals rather than sampling errors. The principal components (PCs) associated with each EOF pattern were subsequently analyzed to understand the temporal evolution of each mode, with a particular focus on the interannual variability in this study.

Figure 1. Shaded relief map of Tanzania, indicating variations in land surface elevation. The highest point is Mount Kilimanjaro at 5895 meters above sea level (Source: Author).

2.2.2. Composite Analysis of Atmospheric Circulation

Composite analysis is conducted to elucidate the atmospheric conditions associated with different phases of the identified Tmax variability modes. Years with PCs out of the range of ±0.5 standard deviation (s.d.) are selected to create composites for high (higher than +0.5 s.d.) and low (lower than −0.5 s.d.) PC years, following established methodologies in climate dynamics research ([9]). This threshold provides a balance between maintaining adequate sample size and ensuring clear separation between opposing phases. Statistical significance of composite anomalies or differences is assessed using Student’s t-test, with adjustments for effective sample size due to temporal autocorrelation ([36]).

This analysis of surface energy budget is carried out independently by assessing individual flux components to isolate their specific roles in driving temperature variability ([34]). Please note that all the fluxes and radiation in ERA5 are positive downward, which means a positive contribution signifies a process adding heat to the surface (warming) and a negative contribution signifies heat removal (cooling). Thus, net shortwave radiation remains a primary positive (warming) driver, while the typically upward net longwave radiation is explicitly isolated as a persistent negative (cooling) contributor. The turbulent fluxes are similarly interpreted upward sensible and latent heat fluxes (convection and evaporation) represent negative (cooling) contributions, with a reversal to downward fluxes indicating a positive (warming) effect. This sign-aware decomposition allows the attribution of temperature variability to distinct physical mechanisms radiative, turbulent, latent and conductive providing a clearer process-based diagnosis than analyzing the net energy balance alone.

The spatial analysis of monthly mean Tmax across Tanzania from 1979 to 2024 reveals a distinct and dynamic climatological pattern (Figure 2). The central plateau region, encompassing much of the Dodoma and Singida, consistently exhibits the highest temperatures throughout the year, with mean Tmax values frequently exceeding 32˚C. The hottest conditions are observed in the months leading up to the primary rainy season, particularly from October to December, where the central interior experience mean Tmax values reaching 34˚C or higher (Figures 2(j)-(l)). In contrast, the eastern and southern coastal zones, including Dar es Salaam and Mtwara, along with the offshore islands of Zanzibar and Mafia, display a more moderated and stable thermal regime with mean Tmax values ranging between 28˚C and 31˚C across all months. The mountainous regions in the northeast (Kilimanjaro) and southwest (Mbeya) persistently show the coolest conditions, with mean Tmax consistently below 26˚C. During the peak of the dry season, from June to September, the band of highest temperatures expands, covering much of the central and southern inland regions (Figures 2(f)-(i)). Conversely, the central rainy season from March to May brings a noticeable cooling effect to the interior (Figures 2(c)-(e)). However, this cooling is less pronounced along the coast, which maintains its relatively stable temperatures year-round.

This annual cycle ends in the SON season, which represents the transitional period from the cool and dry season to the hot and wet season (Figure 3). The spatial average for this season consolidates the patterns showing a strong thermal gradient from the hot central plateau to the cooler coastal and highland fringes. The central region remains the thermal center of the country during SON, with spatially averaged Tmax values firmly in the range of 32˚C - 34˚C. This confirms that SON is the hottest quarter of the year for the majority of Tanzania’s interior, setting the stage for the onset of seasonal rains.

Figure 2. Spatial distribution of monthly mean Tmax (in ˚C) across Tanzania during 1979 to 2024 based on CPC from (a) January to (l) December.

Figure 3. Seasonal distribution of Tmax (in ˚C) over Tanzania averaged from September to November (SON) during 1979 to 2024.

The first leading EOF mode (EOF1) accounts for 41.1% of the total variance, which demonstrates that the most significant driver of interannual temperature anomalies is a uniform and nationwide warming or cooling (Figure 4(a)). However, the second EOF mode (EOF2) accounts for 21.4% of the total variance. Interestingly, EOF2 shows a dipole pattern of Tmax anomalies, with a strong contrast between the central interior and the northeast/coast (Figure 4(b)). Clearly, this EOF2 mode spatially agree with the climatology pattern of SON Tmax, especially for the Tmax center over the central interior region (comparing Figure 4(b) to Figure 3). This spatial agreement would allow us to explore the interannual variability of Tmax over the central plateau based on the EOF2 mode, rather than EOF1 with anomalous Tmax center over the mountain region in the southwest. Thus, our focus is on the EOF2 pattern (i.e., the dipole pattern) and its associated atmospheric anomaly conditions in the following.

Figure 4. The leading Empirical Orthogonal Function (EOF) modes for the SON-mean Tmax from 1979 to 2024: (a) the first EOF mode (EOF1) and (b) the second EOF mode (EOF2).

Figure 5 shows the principal component time series of the dipole mode (i.e., PC2). Clearly, there exist certain inter-decadal variations (Figure 5(a), red line) of this mode, suggesting possible low-frequency oscillations that may modulate the frequency or intensity of this dipole pattern over multi-year periods. More critically, the inter-annual component (Figure 5(b), green line) allows for the identification of specific years where this mode is strongly active. According to interannual time series of PC2, ten high years are identified as 1980, 1981, 1987, 1993, 2003, 2005, 2016, 2017, 2018 and 2021, and nine low years are identified as 1982, 1984, 1989, 1994, 1997, 1999, 2006, 2019 and 2020.

The physical reality and impacts of this dipole can be confirmed by the composite analysis (Figure 6). During years with a strong positive PC2, the central and southern interior experiences significant positive Tmax anomalies mostly exceeding +1.0˚C, while the southwest and coastal regions show weak warming or even slight cooling (Figure 6(a)). Conversely, during strong negative PC2 years, the pattern is opposite with the interior cooling and the coastal/southwestern regions warming (Figure 6(b)). The difference map highlights this dipole pattern and confirms that the contrasting anomalies between the two regions are statistically significant (Figure 6(c)). As shown later, this seesaw effect is likely driven by regional anomaly patterns of atmospheric circulation and clouds.

Figure 5. Principal component of the EOF2 (PC2): (a) total variability (blue line) and inter-decadal variability (9-year moving average, red line) and (b) inter-annual variability (calculated as total minus inter-decadal components, green line; referred to as PC2).

Figure 6. Composites of the interannual anomalies of SON-mean Tmax based on extreme years of PC2 for (a) high years, (b) low years, and (c) their difference (i.e., high years minus low years; (a) minus (b)). The stippling indicates the difference is statistically significant with p < 0.05 based on Student’s t test.

3.3.1. The Correlation Analysis between the Interannual Component of PC2 and Standard Climate Indices

The results of the analysis of the SON PC2 time series reveals a climate mode with distinct extreme events but weak linear forcing from common tropical indices (Figure 7). The standardized time series in panel (a) identifies specific extreme years where PC2 exceeded ±1.0 standard deviations, with notable positive peaks including 1998, 2010, and 2020 (Figure 7(a)); these eight highs and seven low extremes form the basis for the composite analysis (Figure 7(h)). Despite of the clarity of these events, a multivariate linear regression using six indices (DMI, NINO12, NINO34, NINO4, IOD, SOI) explains only 7.8% of the total variance in PC2, indicating a very weak collective linear influence (Figure 7(d)). The strongest individual contemporaneous teleconnection is with the NINO12 index, as shown by the correlation analysis and the scatter plot (Figure 7(b), Figure 7(c)). This is quantified where NINO12 contributes the largest portion of the individually explained variance (Figure 7(i)). However, the 15-year running correlation demonstrates that this relationship with NINO12 is highly non-stationary, shifting markedly over recent decades and suggesting that the identified extreme years are not consistently tied to a single ENSO phase (Figure 7(g)). Furthermore, the lagged correlation diagnostics in panel (f) reveal that indices such as the DMI/IOD exhibit significant correlations when leading PC2 by several months, indicating they may act as precursor forcing mechanisms for extreme PC2 events (Figure 7(f)). The SON PC2 mode is characterized by strong episodic extremes but appears to be driven by a complex, evolving interplay of tropical signals, where direct contemporaneous influences are modest and often supplemented or preconditioned by lagged interactions from basins like the Indian Ocean.

3.3.2. Surface Energy Fluxes and Cloud Characteristics

Figure 8, Figure 9 show the composites of interannual anomalies of surface energy fluxes and clouds, which provide the crucial physical mechanisms that explain the distinct regional dipole pattern (EOF2) of SON Tmax variability identified above.

The key initiator of the surface heating pattern over the central interior is a significant increase in surface net LW radiation (LW_net) during the positive PC2 phase (Figures 8(d)-(f)). This large positive LW_net anomaly (exceeding +20 W·m⁻²) indicates substantially enhanced downwelling LW radiation at the surface. This is primarily due to increased downwelling longwave radiation (Figure not shown), likely from a warmer and more humid atmospheric boundary layer under increased cloud cover. On the other hand, surface net SW radiation (SW_net) increases (decreases) under high (low) PC2 years (Figures 8(a)-(c)), indicating that the Tmax change is also partially induced by the SW_net. Clearly, the anomaly magnitude of SW_net is much smaller than that of LW_net, suggesting a bigger role of LW_net in driving SON Tmax changes.

The combined increase in LW_net and SW_net provides abundant net radiative energy at the surface. This energy is partitioned primarily into sensible heat flux, which shows a large and statistically significant increase (over +10 W·m⁻²; Figures 8(g)-(i)). The warmed surface heats the near-surface air, and this warmer air then further warms the surface through enhanced downward sensible heat flux, creating positive feedback that amplifies maximum temperatures. Simultaneously, the latent heat flux in the interior shows a slight decrease or minimal change (Figures 8(j)-(l)), indicating that the available energy is channeled preferentially into sensible heating rather than evaporation.

Figure 7. Analysis of the second principal component (PC2) time series for September-October-November (SON) from 1979 to 2023 quantitatively correlated with climate indices. Panels shows (a) Standardized PC2 time series, highlighting extreme years (b) Correlations between PC2 and key climate indices (c) Scatter plot of the strongest correlation (d) multivariate regression coefficients (total explained variance R² = 0.078 R² = 0.078) (e) Time series of key climate Indices (f) Multifaceted diagnostics including lagged correlations (positive lag = index leads PC2), (g) 15-years running correlation between PC2 and NINO12 (h) Composite analysis for high/low PC2 years and (i) Individual explained variances.

Figure 9 shows the spatial distribution of composite of cloud cover. During high PC2 years with central warming, the central interior experiences a substantial increase of cloud cover (Figure 9(a), Figure 9(d), Figure 9(g)), which directly explains the concurrent significant increase in LW_net (over +20 W·m⁻²; Figure 8(d)). The spatial pattern of enhanced longwave cloud forcing aligns perfectly with the positive lobe of the Tmax dipole. This further confirms that the initial and dominant radiative control mainly comes from the larger LW_net anomaly due to cloud changes, rather than SW_net.

Figure 8. Composite analysis of the interannual anomalies of SON-mean (a-c) surface shortwave radiation (SWnet), (d-f) surface net longwave radiation (LWnet), (g-i) surface sensible heat flux (SSHF), (j-l) surface latent heat flux (SLHF) based on extreme years of PC2 for (a, d, g, j) high years, (b, e, h, k) low years, and (c, f, I, l) their differences (i.e., high years minus low years). All the surface fluxes and radiation are in W/m2 with positive values downward. The stippling indicates the differences are statistically significant withp < 0.05 based Student’s t test.

Please note that the initial radiative warming in the interior can be further amplified by the land-atmosphere feedbacks. Accelerated soil drying shifts the surface energy balance from latent to sensible heat flux, a globally recognized accelerator of heat extremes ([32]). This feedback is particularly effective in transitional climate zones like the Tanzanian interior, as demonstrated in studies on heatwave amplification ([20]) and the global sensitivity of afternoon temperatures to surface moisture ([19]). Furthermore, the resulting warm, dry boundary layer enhances downwelling longwave radiation, creating a localized “greenhouse” effect that reinforces nighttime warming and event persistence, a mechanism detailed in heatwave analyses ([29]; [28]).

Figure 9. Composite analysis of the interannual anomalies of SON-mean (a-c) low cloud cover (in % of sky), (d-f) high cloud cover (in % of sky), (g-i) total cloud cover (in % of sky) based on extreme years of PC2 for (a, d, g) high years, (b, e, h) low years, and (c, f, i) their differences (i.e., high years minus low years). The stippling indicates the differences are statistically significant with p < 0.05 based Student’s t test.

3.3.3. Atmospheric Circulation, Vertical Velocity and Moisture Flux Divergence

The analysis of atmospheric circulation and moisture dynamics provides the fundamental dynamics for the cloud cover and surface energy flux patterns that drive the SON Tmax dipole pattern (Figure 10, Figure 11). During the positive PC2 years, remarkable anomalous easterly and northeasterly winds occur over most areas of Tanzania, which advect a warm and moist oceanic air mass from the northeast into the central and southern interior regions (Figure 10(a), Figure 10(c)). Accordingly, the lower troposphere becomes moister with increased relative humidity (RH). However, at approximately 850 hPa, the central interior experiences a significant increase in moisture flux divergence during high PC2 years, where the wind field actively transports moisture away from the near-surface layer, thus desiccating the very low-level atmosphere. This low-level divergence is vertically decoupled from the dynamical forcing above it (Figure 11(a)), anomalous ascent dominates the layer from 850 hPa to 700 hPa. This ascent lifts the advected moist air, facilitating cloud formation at these higher altitudes despite the near-surface divergence. In contrast, the coastal zone shows a tendency towards moisture flux convergence at 850 hPa, consistent with the direct advection of maritime moisture and more clouds there (Figure 9(a)). As a result, there is a large increase in total cloud cover in the central interior and coastal regions, while clouds increase less in the western inland (Figure 9(g)).

Figure 10. Composites analysis of the interannual anomalies of SON-mean (a-c) relative humidity (RH, shading, in %) and (d-f) moisture flux divergence (shading, in 1/s) at 850hPa based on extreme years of PC2 for (a & d) high years, (b & e) low PC2-IV years, and (c & f) their differences (i.e., high years minus low years), overlaid with 850-hPa wind anomalies (vectors, in m/s). The stippling indicates the differences are statistically significant with p < 0.05 based Student’s t test.

Figure 11. Analysis of the Pressure-latitude cross-sections for the composites interannual anomalies of vertical velocity (ω, in Pa/s) averaged over the 25˚ to 42˚E based on extreme years of PC2 for (a) high years, (b) low years, and (c) their difference (i.e., high years minus low years). The stippling indicates the differences are statistically significant with p < 0.05 based Student’s t test.

3.3.4. The Potential Drivers of the Uniform Warming/Cooling Pattern of EOF1

The uniform warming/cooling pattern of EOF1, which explains the dominant share (41.1%) of SON Tmax variance is not analyzed in detail here but is a characteristic of large-scale coherent mode typically associated with global background warming and pervasive sea surface temperature forcings ([32]). This pattern is most consistent with forcings that operate synchronously across the region, primarily the long-term trend of global climate change and basin-wide sea surface temperature (SST) anomalies, such as uniform Indian Ocean basin warming or the canonical, broad-scale teleconnection from Pacific ENSO events, which can modulate atmospheric stability and cloud cover across East Africa. While EOF1 captures the most overall variance, the we focus mechanistic analysis on EOF2 (21.4% of variance) because of its dipole structure aligns with the core climatological heat center over the central plateau. This allows for an investigation of the regionally specific dynamics moisture advection, vertical motion and the cloud-radiative feedbacks that interact with powerful land-atmosphere coupling which are critical for understanding and predicting impactful temperature extremes in Tanzania’s interior, a region highly sensitive to soil moisture temperature feedbacks ([20]; [18]; [29]).