All data used in this study were obtained from the ERA5-Land database , distributed by the Climate Data Store (CDS) of the Copernicus Climate Change Service . ERA5-Land is a high-resolution reanalysis dataset produced by the European Centre for Medium-Range Weather Forecasts (ECMWF). It offers detailed information on land surface variables, providing a consistent view of their evolution over several decades. ERA5-Land features a finer spatial resolution of approximately 9 km grid spacing, compared to ERA5 ∼25 km. It covers data from January 1950 to the present, with an hourly temporal resolution. The dataset includes various land surface parameters (such as soil moisture, soil temperature, snow cover, and surface runoff) to control the simulated land fields, ensuring the data remains accurate and consistent. ERA5-Land uses atmospheric variables from ERA5, like air temperature, wind speed and humidity, and presents hourly temporal resolution, thus providing the necessary input variables for the calculation of both reference and Proxy FWI versions, as well as the different predictor sets used in this study (Sect. ).

The reference noon LST FWI is calculated from ERA5-Land daily records of four near-surface meteorological variables measured at 12 UTC (∼ noon time in the Iberian Peninsula): 24 h accumulated precipitation, instantaneous wind speed, relative humidity, and temperature. These inputs are processed via a set of empirical equations that yield six intermediate components characterizing fuel moisture dynamics across different fuel layers , the influence of wind on fire spread, and the total fuel available for combustion. These components are then integrated to compute the FWI, a dimensionless index representing potential fire intensity under given meteorological conditions for a reference fuel type (mature pine stands).

Beyond the four meteorological inputs required at noon LST, the Canadian FWI System relies on six sequential sub-indices that describe fuel moisture conditions and potential fire behavior . Three of them (FFMC, DMC and DC) represent moisture contents at increasing depths of the forest floor and respond to atmospheric drivers at markedly different timescales.

The Fine Fuel Moisture Code (FFMC) characterizes the moisture content of surface litter and fine fuels, which react on hourly–daily timescales to changes in temperature, relative humidity and wind speed, as well as to even small rainfall amounts due to their limited water-holding capacity. FFMC largely governs ignition likelihood and short-term fluctuations in fire danger.

The Duff Moisture Code (DMC) represents the moisture content of moderately deep organic layers in the upper duff. Its evolution reflects medium-term drying (days to weeks), driven primarily by temperature and relative humidity, and decreases only when precipitation exceeds an effective infiltration threshold. The DMC modulates the potential for sustained fire spread once ignition occurs.

The Drought Code (DC) reflects the long-term moisture content of deep, compact organic horizons that respond over weeks to months. The DC accumulates persistent atmospheric moisture deficits and only decreases when substantial or prolonged precipitation is able to percolate into deeper layers. As such, it is a key indicator of seasonal-scale drought and deep-burning potential. These different memory timescales explain the unequal sensitivity of the moisture codes to the meteorological drivers .

The three moisture codes are combined into two intermediate fire-behavior indices. The Initial Spread Index (ISI) combines FFMC with wind speed to estimate the expected rate of fire spread immediately after ignition. The Buildup Index (BUI) integrates DMC and DC to represent the total amount of fuel available for combustion, accounting for both medium- and long-term moisture deficits. Finally, the Fire Weather Index (FWI) combines ISI and BUI through a nonlinear formulation to estimate potential fire intensity for a standard fuel type.

A practical consideration arises from the different response times and memory effects of these components. While the FWI system uses default initial values, leading to a short spin-up period, this transience is typically negligible during the fire season due to rapid stabilization and the minimal influence of snowmelt on moisture inputs .

The reference FWI is computed directly from instantaneous meteorological variables at 12:00 UTC (approximately noon local standard time in Iberia), as defined in the standard FWI formulation (see Sect. ). In contrast, the predictor sets used to train the DL emulators (P0, P1, P2) consist of different temporal aggregations and/or preprocessing of these same meteorological variables. Specifically, the emulators are provided with daily aggregated variables (daily means, daily minima/maxima, together with 24 h accumulated precipitation) rather than instantaneous noon values. This distinction is essential for interpreting emulator performance: the DL models learn to recover a noon-time fire weather signal from daily-aggregated inputs, which is an inherently different (and harder) task than the direct FWI computation. Reported errors reflect this added complexity and should not be interpreted as a simple approximation error.

To emulate the reference FWI, we consider several predictor sets derived from ERA5-Land, summarized in Table . The initial experiments use P0, which builds on the same set of variables used in the optimal FWI approach from (hereafter Proxy FWI). This allows us to assess whether the DL models can more accurately replicate the FWI transfer function when provided with the same predictors as the Proxy FWI. Accordingly, P0 includes 24 h accumulated precipitation, daily mean air temperature and wind speed, and minimum relative humidity. Proxy FWI is derived from the standard FWI formulation using the latter set of variables. The use of minimum relative humidity in P0 ensures the fairest possible comparison with Proxy FWI. However, in predictor sets P1 and P2, we intentionally replace minimum relative humidity with daily mean relative humidity. This is motivated by the fact that daily mean variables are more consistently provided by climate model outputs and reanalysis datasets, and we wanted to examine the performance of the emulator under such predictor-limited cases. Precipitation poses an additional challenge: it is one of the most difficult variables for climate models and reanalyses to represent reliably, due to its strong spatial and temporal variability, its dependence on complex physical processes, and the influence of multiple interacting factors. To examine whether robust performance can still be achieved in its absence, we exclude precipitation from predictor sets P1 and P2. Finally, in P2 we tried measuring the impact of using wind speed module versus wind speed zonal components, which could be relevant when determining fire danger. Other predictors sets have been tested, such as P1 using minimum relative humidity instead daily mean relative humidity or P0 including the 24 h precipitation, without any added value compared to the configurations mentioned in the manuscript. These experiment results are shown in Fig. .

In this section, we briefly describe the DL architectures trained in this study. A visual summary of their topologies is presented in Fig. . The selected architectures include the Fully Connected Dense (FCD) model, chosen as benchmark for its relative simplicity, as well as DeepESD and U-Net, which have been recently introduced in the literature for climate downscaling . Alternative architectures, such as Convolutional Long Short-Term Memory (ConvLSTM), were also evaluated. However, owing to their inferior performance relative to the selected architectures and their greater computational demands, they are not presented in this manuscript. The DL models are tasked to learn a mapping between different sets of FWI-related input variables (Table ) and the reference FWI over the Iberian Peninsula as output, by minimizing a loss function (Sect. ). The model output is an emulated version of the reference ERA5-Land FWI at the same spatial resolution than its inputs. All models are trained with the same optimization parameters independently of the architecture, using the Adam optimizer with a learning rate set to 0.0001, a batch size of 64 and a maximum of 1000 epochs. An early stopping mechanism is set to prevent overfitting, which stops the training process if there is no improvement in the validation loss after 30 consecutive epochs. Furthermore, the input data are standardized as part of the model preprocessing: 1x′i=(xi-μi)σi where x′ represents the standardized value, x represents the raw (unstandardized) value, and μi and σi represents the mean and standard deviation at gridpoint i. Parameters μ and σ have been computed relative to the training period 1979–2017 (per gridpoint).

In this study, the DL models are trained using daily data from the period 1979 to 2017. To evaluate the performance of the model, 20 % of these data is separated and treated as validation samples for model assessment during the training phase. The final results are presented for an independent test period spanning 2018–2021. We also evaluated longer test periods (2012–2021), which required shortening the training phase to 1979–2011. Since these tests produced robust and consistent results comparable to those for 2018–2021, they are not included in the text. To assess the performance of the DL models and the reproducibility of spatial and temporal patterns in the emulated results, several validation indices are used, each focusing on different spatial and temporal aspects of the predicted series (Table ).

At each grid point, validation indices are computed, and their spatial distributions are depicted in the maps presented in various figures in Sect. . These indices span different key characteristics of the predicted series to assess how models emulate FWI, such as the mean FWI and extreme FWI distribution, or the length of spells for extreme FWI events. Instead of using MAE (for FWI and FWI95) or bias (for Max Spell95), we assess relative bias with respect to the frequency of FWI95, given by the following formula: 3Relative bias=X-X^X×100 where X represents the observed and X^ the estimated value.

We also analyze the Receiver Operating Characteristic (ROC) curve, and the area Under the Curve (AUC), a graphical tool for evaluating the performance of binary classification models , in this case applied to assess the DL model's ability to discriminate dangerous FWI events. It plots the true positive rate (sensitivity) against the false positive rate at various decision thresholds and quantifies the overall discriminative ability of the model, with values ranging from 0.5 (random classification) to 1.0 (perfect classification). A higher AUC indicates better model performance in distinguishing between positive and negative cases, making the ROC AUC particularly useful in unbalanced classification problems where traditional accuracy measures may be misleading. In particular, the classification evaluated in the ROC AUC analysis pertains to FWI events classified as “extreme” (see Table ), defined as those exceeding the 95th percentile of the reference FWI, according to the danger levels established by the Spanish Meteorological Agency (AEMET, https://www.aemet.es/documentos/es/datos_abiertos/Estadisticas/IM_riesgo_incendios/eimri_generalidades.pdf, last access: 17 February 2025), being thus impact-relevant for wildfire risk assessment in the Iberian Peninsula. In addition to the validation indices listed in Table , we also use scatter plots the assessment of the emulated FWI 95th percentile against reference FWI climatologies, where the R2 coefficient of determination is used.

DL models have demonstrated remarkable predictive capabilities across various scientific fields, yet their complex architectures often limit interpretability. This challenge has motivated the growth of eXplainable Artificial Intelligence (XAI), which aims to provide tools for understanding the underlying relationships learned by machine learning models . In this work, we apply saliency-based methods to investigate the internal logic behind our DL predictions. Specifically, we employ a variant of the Integrated Gradients (IG) technique , which has been widely adopted in climate applications for its effectiveness and ease of interpretability.

Traditional IG computes feature attributions by integrating the gradients of the model output with respect to the input, along a linear path from a baseline input to the actual input. The method is formally expressed as: 4si(xi;xb)=(xi-xb)∫01∂fxb+t(xi-xb)∂xidt where xi is the feature value, xb is the baseline (typically zero), and f is the output of the model.

However, in this study we adopt a simplified approximation of IGs, where the integral is omitted and the feature relevance is estimated directly from the gradients at the input point. This results in a computationally simpler, yet informative, saliency map representation. The relevance of each input feature is therefore computed as: 5si(xi)=xi⋅∂f(x)∂xi This approach preserves the core idea of attributing the prediction to the input features based on the local sensitivity of the model, while avoiding the computational overhead of path integration. Although this approximation does not fully satisfy all axioms of the original IG method , it has been shown to yield stable and interpretable attributions in practical applications .

To address known XAI challenges , the raw saliency maps are post-processed by: (1) taking the absolute value to focus on magnitude only, (2) zeroing all values below 10 % of the per-sample maximum to suppress gradient noise , and (3) normalizing each map so its values sum to one, thus yielding relative contributions. These normalized maps are used for all subsequent explainability analyses.

This methodology enables us to quantify the relative importance of each predictor variable in the model’s decision-making process, providing valuable insights into the learned relationships and the adequacy of the different sets of predictors tested, aiding in the underlying physical interpretation of the results.

Both FWI calculation and emulation are carried out using the R-based climate4R framework . In particular, neural network emulators (Dense, DeepESD, U-Net) are defined and trained using the downscaleR.keras package , an extension of the downscaleR package that leverages Keras/TensorFlow.

Figure presents a multi-panel visualization, where the rows correspond to the validation measures shown in Table . From left to right, the first two columns represent the reference FWI and the Proxy FWI climatologies (P0, Table ). The last column indicates the MAE or bias between both – depending on the evaluated index, Table –. The FWI95 frequency for the reference FWI is not shown, because, by construction, each grid point in the reference data records exactly 5 % of days above its local 95th percentile threshold during the season. Therefore, the value in each grid across the spatial map is 0.05. It serves as the observed reference for error calculation: The 95th percentile of reference FWI is used as baseline to compute the frequency of Proxy FWI events exceeding this threshold (as well as emulated FWI in the following section).

Focusing on the differences between the Proxy and the reference FWI, we note that there exists greater MAE in Southern Portugal, North-western Spain and North-Eastern Spain, along the Ebro Valley. Conversely, the lowest MAE values (close to zero) are found along the Cantabrian and Mediterranean coasts, northwestern Portugal, the Pyrenees, and the Balearic Islands. These low-MAE areas generally align with the lowest climatological FWI values (as depicted in first column, first row), except for the Balearic Islands, where climatological JJAS FWI remains relatively high (FWI ≈ 35).

Regarding extreme FWI events (above reference FWI 95th percentile, FWI95), the Proxy FWI95 exhibits the highest MAE in the central-eastern and southeastern continental regions, coincident with very high reference FWI95 climatological values. Although the Proxy FWI95 accurately represents southern Portugal, revealing low MAE values, overall, the most critical FWI95 regions (mostly over Coastal Mediterranean and Central and Iberian Massifs) tend to be underrepresented by the Proxy FWI.

The Proxy FWI exhibits a relatively high bias in the frequency of FWI95 events, over continental and coastal Mediterranean regions (negative) as well as the western atlantic façade (positive). This behavior arises from structural differences between the Proxy formulation (based on daily aggregated variables) and the reference FWI, which depends on instantaneous noon-LST conditions. Because extreme fire-weather situations typically occur near midday, the use of daily means and daily minimum/maximum values smooths the sharp sub-daily variations that drive threshold exceedances, leading to systematic over- or under-counting of FWI95 days. Despite this limitation, the Proxy FWI reproduces the main spatial gradients and climatological patterns of mean FWI, FWI95 and Max Spell95, confirming that it remains a useful and physically consistent baseline against which to assess the added value of the deep-learning emulators. Regarding the temporal sequence of extreme–event occurrences, the relative bias for the FWI95 frequency displays a spatial pattern similar to that of the Max Spell95, with overestimation of FWI95 days in Portugal, north–western Spain (Galicia), the Cantabrian mountain range, the Pyrenees and the Ebro Valley, and substantial underestimation across the rest of the Iberian Peninsula. Consistently, the spatial distribution of Max Spell95 shows that the Proxy FWI overestimates spell lengths in Galicia and Portugal (by up to 4 d) while in the continental interior it tends to underestimate them, remaining close to the reference values over most of south–western Spain.

This section presents an intercomparison of the different DL models tested, by evaluating errors relative to the reference FWI and comparing them with the Proxy FWI results described in the previous section, which serves as a benchmark. For brevity, here we present the results using the predictor set P0 as input (Table ), considering that the overall intercomparison results are consistent regardless of the predictor set tested. Since P0 includes the variables necessary for computing the Proxy FWI using the standard FWI formulation, our goal is to assess whether different DL architectures are able to capture the reference FWI more accurately using the same input information.

Here, we analyze the saliency attributed to the predictor variables, grouped by the same climatic regions and fire danger categories defined in Appendix . The overall results indicate no major differences in the saliencies of the different DL models, underpinning their robustness and physical coherence. Therefore, here we present the explainability results of the U-Net model, since it outperforms Proxy FWI and attains similar validation results as DeepESD, while offering greater computational efficiency. To this aim, we compute the saliencies of each variable considering the time events corresponding to the different fire danger categories separately (Fig. ).

Focusing on the high, very high, and extreme fire danger categories, temperature and relative humidity emerge as the most influential variables in the model’s predictions. The relative importance of these two variables varies across climatic regions. In the Continental Mediterranean region, temperature and relative humidity show similar levels of saliency. For the high fire danger category, relative humidity slightly surpasses temperature, whereas for the very high and extreme categories, temperature becomes more dominant. Wind speed consistently ranks as the third most important variable in these categories, although its relevance diminishes as the fire danger increases. The relatively low saliency of wind speed for the most severe categories reflects both the structure of the training data and the role that wind plays within the Fire Weather Index (FWI) formulation at very high danger levels. In Iberia, extreme fire danger episodes are predominantly associated with persistent heat anomalies, very low relative humidity, and sustained fuel desiccation, rather than with short-lived wind maxima. Once fuels are critically dry, additional increases in wind speed exert a comparatively smaller influence on the final FWI class than thermodynamic controls, which leads the model to rely more strongly on temperature and relative humidity when discriminating between extreme and very extreme conditions. Moreover, wind exhibits higher temporal variability and weaker spatial coherence than the other predictors, particularly when daily aggregated inputs are considered.

This finding is consistent with the definition of the FWI itself, which relies primarily on temperature, relative humidity, and wind speed under dry conditions. The DL model thus reflects the structure and sensitivity of the FWI metric it is trained to emulate. Importantly, this does not mean that precipitation (or other inputs) is irrelevant for real-world fire risk; rather, it highlights that the predictand (FWI) gives limited weight to precipitation in high and extreme danger situations. Moreover, ERA5-Land predictor variables are not independent. Relative humidity, for instance, is derived from temperature and dew point, which are indirectly influenced by precipitation, while precipitation and wind are assimilated forcings. These inter-dependencies likely contribute to the model’s ability to achieve high predictive accuracy even when precipitation and wind receive lower attribution scores.

The Coastal Mediterranean region shows a similar pattern. However, in the high category, temperature becomes more salient than relative humidity. In contrast, for the extreme category, relative humidity overtakes temperature as the most relevant predictor.

Similarly, in the Atlantic region, temperature is the most significant variable for the high and very high categories. For the extreme category, however, relative humidity surpasses temperature. Notably, in this region, the difference in saliency between wind speed and the top-ranked variables is smaller compared to other regions, suggesting a higher reliance on this variable in this region. As with the other regions, the relevance of wind speed decreases with increasing fire danger, and precipitation remains consistently negligible.

For the medium and low fire danger categories, the importance of variables shifts more markedly and shows less consistency across climatic regions. In the Continental and Coastal Mediterranean regions, precipitation becomes the most relevant variable for the low category, followed by relative humidity, wind speed, and temperature. For the medium category, wind speed emerges as the most salient predictor, followed by temperature, relative humidity, and precipitation.

In contrast, in the Atlantic region, relative humidity is the most important variable for the low category, followed by precipitation, wind speed, and temperature. For the medium category, temperature takes the lead, followed by wind speed, relative humidity, and finally precipitation. These results suggest that precipitation is a key variable for low fire danger levels but loses significance as fire danger increases, becoming practically irrelevant in extreme conditions, when temperature and relative humidity gain prominence and become the most critical predictors. This is consistent with numerous studies that highlight the importance of high temperatures and low humidity in contributing to extreme fire danger .

Given that precipitation has negligible saliency in high percentile fire danger cases, which are precisely the most important from the point of view of impacts, the next section evaluates model performance after removing precipitation from the predictor set. We also examine alternative predictor sets, including replacements for precipitation and wind speed, such as eastward and northward wind components, under the hypothesis that these may carry additional useful information for the emulators.

In this section, we intercompare different alternative input predictor sets, summarized in Table . For simplicity, we focus on the U-Net DL method as in Sect. . The U-Net trained with P0 and P1 yields better results than when trained with P2, except for the MAE FWI (Fig. , analogous to Fig.  for comparability). The difference between U-Net P0, U-Net P1 and U-Net P2 is relatively small, with the most remarkable variations observed in the western Continental region. For the MAE of FWI95 events, U-Net P2 significantly overestimates across the entire domain, whereas U-Net P0 and U-Net P1 present similar MAE spatial distribution, standing P1 as the best approach. However, U-Net P1 shows a greater presence of errors in specific regions, such as north Portugal, the Pyrenees and the Cantabrian mountain ranges. Regarding the frequency of FWI95 events, U-Net P0 tends to overestimation across nearly the entire the north domain, while U-Net P1 tends to overestimate in the mid-south and east of the Peninsula and U-Net P2 exhibit a general underestimation. Among these, U-Net P0 appears to be the most effective in capturing this index. Examining the Max Spell95 results of the different experiments, U-Net P2 performs the worst, with overestimation up to 3 to 4 d in some areas, whereas the largest differences in P0 and P1 range between 2 and 3 d at most. U-Net P1 shows a slight improvement over U-Net P0, displaying more areas of overestimation, while U-Net P0 predominantly underestimates across most of the domain. Finally, the FWI95 scatter plots exhibit a high R2 fit for all three predictor sets, outperforming the Proxy FWI. In terms of overall performance, U-Net P1 produces the best results, followed by U-Net P0 and U-Net P2, though the differences among the latter remain relatively small.

Overall, in Fig. , U-Net P1 yields the best results as classifiers for extreme events. These experiments improve upon the median results of the Proxy FWI, with U-Net P1 emerging as the most effective model in this context.

In summary, the U-Net model demonstrates superior performance over the Proxy FWI, with U-Net P1 generally excelling in overall performance. The inclusion of wind components as additional predictors in P2 has a deleterious effect on the temporal and distributional similarity of emulated and reference FWI. This suggests that, for emulating the FWI index itself, wind speed magnitude already captures the relevant wind-related information, and the addition of directional components introduces redundant or noisy predictors that can hamper model performance, despite the internal feature selection of deep learning methods. Consequently, a prior predictor screening may still be necessary to ensure optimal results. Importantly, as seen through the current section, the scope of this work is restricted to the statistical reconstruction and characterization of fire-weather conditions as represented by the FWI, rather than the analysis of wildfire occurrence, behavior, or impacts. The DL-based results are validated considering the FWI95, FWI95 frequency and the Max Spell95 (Figs.  and ) and the AUC of the ROC curve for extreme fire danger classification as noted in Figs.  and . With this suite of validation metrics and experiments, we properly assess the ability of the DL-based results to statistically reproduce the most extreme events. Consequently, we do not investigate individual wildfire events or fuel-specific fire behavior, which lie outside the intended scope of this study.

We next present a summary of the validation results, including an intercomparison of alternative MSE and asymmetric loss functions in the DL models (Sect. ). Table presents the validation indices for different predictor sets and DL models, comparing their performance against the Proxy FWI baseline, where the ASYM loss function values are indicated in parentheses. For all indices, the DL models exhibit consistent improvements over the Proxy FWI, with at least one model outperforming it in each category.

Regarding the loss function, the ASYM loss generally improves performance in most validation indices. The ASYM-trained models tend to yield lower MAE than ASYM values, particularly for DeepESD in P1 (6.74 vs. 6.43) and P2 (6.24 vs. 7.12), showing better accuracy in estimating FWI. Similarly, in the FWI95 MAE, the ASYM-trained models consistently outperform MSE loss counterparts, as seen with DeepESD in P0 (9.28 vs. 6.82) and P2 (9.66 vs. 9.44). In terms of extreme event prediction, the Max Spell95 Bias shows inconclusive results. While the ASYM loss function improves the bias in some cases, such as U-Net in P2 (0.78 vs. 0.54), in other cases, it worsens bias, as observed with Dense in P2 (0.78 vs. 1.05). This suggests that ASYM does not consistently improve the temporal persistence of extreme events but may still provide benefits in specific cases. For the Relative Bias Frequency of FWI95, ASYM provides substantial benefits in reducing bias. The most notable case is U-Net in P1, where the ASYM-trained model achieves a significantly lower bias (0.358 vs. 0.269) than MSE, reinforcing ASYM's ability to enhance the reliability of extreme event frequency estimations.

Overall, ASYM appears to be particularly beneficial in reducing error in FWI estimations (MAE and MAE FWI95) and in improving the classification of extreme fire weather danger, as evidenced by the lower relative bias for FWI95 than MSE. However, its benefits on temporal aspects, such as Max Spell95 Bias, are not consistent. These findings suggest that while ASYM is generally advantageous, its effects are model- and index-dependent, although sometimes overestimates low and moderate values in the distribution (see Fig. ). Therefore, the ASYM loss function does not compromise the U-Net model performance in the lower quantiles, however for the Dense and DeepESD models low and moderate values for the FWI are worse than using the MSE as loss function. Therefore, the ASYM loss function can compromise the performance of low and moderate levels in some cases, and its selection should be carefully considered depending on the primary objective of the analysis.

The Canadian Forest Fire Weather Index System is structured around six interrelated components that describe both fuel moisture conditions and potential fire behavior. In particular, the three fuel moisture codes Fine Fuel Moisture Code (FFMC), Duff Moisture Code (DMC) and Drought Code (DC) represent distinct organic layers with characteristic response times ranging from hours (surface litter) to months (deep compacted layers). These components provide physically interpretable proxies of ignition potential, fire sustainability and drought build-up, and they underpin the behavior of the aggregated FWI. Given their distinct physical meaning and temporal memory, assessing the performance of the deep-learning emulators at the level of the individual moisture codes provides an additional and more stringent evaluation of the modeling framework. To explicitly evaluate rainfall sensitivity and temporal memory across scales, two predictor configurations are considered: P0, which includes precipitation, and P1, which excludes it. This design enables a controlled assessment of the extent to which daily precipitation information contributes to the emulation of fast-, intermediate-, and slow-response fuel layers. Beyond reproducing the aggregated FWI, this analysis examines whether the emulators are able to capture the multi-scale fuel-moisture dynamics embedded in the structure of the system. We emphasize that this analysis should be regarded as preliminary, and that a more comprehensive and detailed assessment is currently under development. Nevertheless, the present evaluation provides an initial characterization of the performance of the DL model originally designed to emulate FWI when the predictand is replaced by its individual fuel moisture components (DC, DMC, or FFMC) instead of the aggregated FWI index.

Table summaries the validation performance of the emulators for DC, DMC and FFMC under the two predictor configurations (P0, including precipitation; P1, excluding precipitation). The evaluation considers some of the validation metrics as the MAE, FWI95 MAE and relative bias in the frequency above the 95th percentile previously depicted in the manuscript. Moreover, we introduce an additional validation metric to explicitly assess the temporal consistency of the predicted subindex relative to the reference subindex. Specifically, we compute the Pearson correlation coefficient between the predicted and reference spatially aggregated time series. Prior to calculating the correlation, the seasonal cycle is removed from both time series to isolate anomalies and avoid inflating the correlation due to the shared annual cycle. To this end, the daily climatology (i.e., one value per Julian day of the year) is subtracted from each daily value in the time series. The daily climatological values are estimated over the full reference period using a centered moving average window of seven days, which smooths high-frequency noise while retaining the intra-seasonal structure of the climatology. This metric provides a domain-integrated measure of the emulator ability to reproduce the large-scale temporal variability of the fuel moisture codes, focusing explicitly on inter-daily to synoptic-scale fluctuations rather than on the mean seasonal evolution. As in the previous sections, these metrics are evaluated for the fire season (JJAS) in the test period (2018–2021), with the exception of the time series correlation, which is computed over the full annual cycle.

For the DC, which reflects long-term deep fuel moisture and cumulative drought effects, correlations are moderate (0.72 for P0 and 0.61 for P1). This suggests a greater challenge in fully capturing low-frequency variability and long-memory dynamics. Absolute errors are comparatively large in magnitude (MAE ∼ 106–118), which is consistent with the broader dynamic range of the DC. P0 achieves both lower MAE (105.62 vs. 118.17) and comparable performance under extreme conditions (FWI95: 227.51 vs. 225.74 are nearly identical, with a slight advantage for P1 in extremes but a higher overall MAE). The relative bias in high-danger frequency is very similar between predictor sets (0.81–0.82), indicating comparable representation of extreme drought-related conditions despite differences in correlation strength. The modest improvement obtained when precipitation is included is consistent with the DC formulation, where only substantial or sustained rainfall events effectively reduce deep-layer dryness; although such events are infrequent during JJAS, their accumulated impact contributes to improved interannual and low-frequency variability representation.

The temporal evolution shown in Fig.  provides additional context for interpreting the DC results. Despite the moderate correlations obtained for both predictor configurations, the DL reconstructions reproduce the main seasonal cycle and the timing of the major multi-month drought episodes throughout the 2018–2021 test period. Moreover, the two predictor sets generate remarkably similar DC trajectories, with only minor differences in amplitude and short-term fluctuations. This behavior is consistent with the validation metrics and indicates that the reduced skill observed for DC cannot be attributed solely to the exclusion of precipitation from the predictor set. Rather, it reflects the intrinsic difficulty of reproducing the long-memory dynamics and low-frequency variability embedded in the DC using architectures that do not explicitly model temporal dependencies, irrespective of whether precipitation is included. This limitation is further quantified in Appendix , where the analysis of temporal persistence reveals a pronounced underestimation of memory for the DC. Together, these results demonstrate that the challenge lies primarily in the representation of persistence rather than in the availability of predictor variables. We note that this limitation is not intrinsic to the emulation framework itself, but to the specific class of architectures employed. Models explicitly designed to capture temporal dependencies, such as ConvLSTM networks, have been shown to better reproduce long-memory processes in a fire-weather prediction context , albeit at a substantially higher computational cost, which limits their applicability in large-domain experiments such as the one presented here.

In the case of the DMC, representing intermediate-depth organic layers with response times of days to weeks, correlations are moderate to high (0.75 for P0 and 0.85 for P1). The removal of precipitation as a predictor in P1 markedly improves temporal correspondence. Overall MAE is slightly lower for P1 (48.23 vs. 50.63), and performance under extreme fire-weather conditions improves substantially (MAE FWI95: 139.81 vs. 173.32). The relative bias in high-danger frequency decreases from 0.93 (P0) to 0.77 (P1), indicating improved prediction of extreme-category occurrence under P1. This behavior indicates that during the Iberian fire season the DMC signal is largely governed by medium-term atmospheric drying driven by temperature and relative humidity, whereas rainfall events exceeding the effective infiltration threshold are comparatively rare at daily scale, thereby limiting the incremental value of explicit precipitation information for seasonal emulation.

Finally, for the FFMC, which governs short-term surface litter moisture and ignition potential, correlations are very high (0.98 for P0 and 0.97 for P1), demonstrating strong skill in reproducing rapid variability. Absolute errors are small (MAE ∼ 3.4–3.6; MAE FWI95: 2.16–2.91), with negligible differences in mean and extreme performance between predictor sets. The relative bias in extreme-day frequency decreases from 0.67 (P0) to 0.52 (P1), indicating improved representation of high-danger occurrences. Given the intrinsic sensitivity of FFMC to even light rainfall, the small differences between P0 and P1 suggest that the deep-learning architecture is able to reconstruct short-term surface moisture variability primarily from thermodynamic and wind predictors at daily scale, while the limited seasonal contribution of precipitation reflects its low frequency during JJAS rather than a lack of physical responsiveness.

In summary, the DL framework shows scale-dependent performance: very strong skill for the fast-response FFMC, moderate-to-high skill for the DMC (particularly under P1), and comparatively lower temporal fidelity for the slow-evolving DC. Despite these differences, biases in extreme-category frequency remain within a relatively narrow range, suggesting that the hierarchical structure of fuel-moisture dynamics is largely preserved. The differential sensitivity observed across predictor configurations is consistent with the physical formulation and temporal memory of each code, reinforcing the internal coherence of the emulation framework across fast, intermediate, and slow fuel-moisture regimes. Future work will also focus on the implementation of the emulator within the context of the PTI-Clima climate-service initiative (https://pti-clima.csic.es/, last access: 24 June 2026), alongside a more exhaustive validation against selected real-world wildfire events.