To carry out the experiments, the WRF (Weather Research and Forecasting) mesoscale model is used, with its ARW (Advanced Research WRF) dynamic core in version 3.8.1. This constitutes the SisPI’s core [6] [14] . The SisPI design has two-way nested domains of 27 and 9 km (kilometers) of spatial resolution, respectively (d01: 140 × 78; d02: 199 × 112). It also has a high spatial resolution domain of 3 km (d03: 421 × 184) one-way nested through the down tool (Figure 1).

The 27 and 9 km domains output is every three hours, the high resolution domain prints the solutions hourly basis. The model is initialized with data from the GFS (Global Forecast System) with 0.5˚ spatial resolution. These characteristics related to the execution and initialization of the operational version of SisPI were respected in the experiments described in this document to evaluate the operating system as homogeneous as possible.

The configuration proposed by SisPI includes 28 vertical levels, the Mellor-Yamada-Nakanishi and Niino2.5 PBL scheme, and the RRTM longwave

parameterization for all domains. In the case of low-resolution domains, it contains the microphysics of WSM5, the Grell-Freitas cumulus parameterization, and the shortwave Dudhia scheme. For the high resolution domain, the Morrison double-moment microphysics is used, the cumulus parameterization is deactivated and the Goddard shortwave radiation scheme is used. These differences have been supported by sensitivity studies made in the project development [6] [14] .

To develop the assimilation experiments, the WRFDA (WRF Data Assimilation) was used in its version 3.9, available free of charge at https://github.com/wrf-model/WRF/releases?page=3. The selection of this slightly higher version than the WRF version used in SisPI responds to the fact that it incorporates the 4DEnVAR method for the first time, in addition to updating and improving the assimilation of satellite data compared to previous versions.

Assimilation was executed only on the high resolution domain, a decision that responds to several factors. In the first place, the fact that data assimilation can be carried out only in one domain at a time would force us to modify the execution philosophy of SisPI with the consequent increase in computational cost by repeating the process in each of the domains. On the other hand, one-way nesting gives the high-resolution grid the possibility of having initial and boundary conditions adjusted to it, which allows assimilation to be carried out in this domain by adding only the computational cost of the assimilation process itself.

The main disadvantage of this design is that the spatial coverage is limited and this can produce fluctuations concerning the availability of data to assimilate, especially with radiances, which come from polar-orbiting satellites.

Data is assimilated in prepbufr format, which contains information from weather stations in FM12 and METAR format, as data from ships, buoys, and soundings. The radiances are assimilated from information from polar-orbiting satellites with information from microwave channels in bufr format, the sensors AMSU-A (NOAA-15/16/18/19), MHS (NOAA-18/19), SSMIS (DMSP-16) and ATMS (Suomi-NPP) were used, in all cases downloaded free of charge from the site: https://rda.ucar.edu/datasets. The set of assimilable data to make the experiments was the same in all cases.

The hybrid methods, as well as the purely variational case, can be summarized as the iterative solution to find the state x that minimizes the cost function J(x) [5] [15] . This solution represents the maximum probability (least variance) estimate of the true state of the atmosphere.

Hybrid schemes combine two ways of obtaining the BEC. On the one hand, they use a BEC generated for a variational process. In this case, the differences or perturbations between valid forecasts for the same time but with different initialization are calculated, generally following the scheme (T1 + 24) minus (T2 + 12), where T is the instant of initialization in each run [15] [16] . These differences are usually constructed using different times of the day to remove the diurnal cycle, typically using 00:00 and 12:00 UTC (United Time Coordinated).

The differences obtained are averaged over a period that must range from 15 to 30 days before the initialization of interest at least [17] , which provides the system with a mean background error. The BECs obtained in this way tend to be synoptically dependent and time invariant, which acts as an encapsulated climatology, which can provide poor results when the flow of the day differs from the one fixed in the covariance matrix [2] [18] . For these experiments, static BECs built with 15 days are used, taking into account that previous results obtained with SisPI [8] , do not show statistically significant differences between the results obtained with BECs with 15 and 30 days.

The other way to obtain the BEC is by calculating the perturbations from a previous ensemble, where all the members contain the initialization time of the experiment to be carried out. For this, the SisPI solutions corresponding to two days before the initialization time are used. This feature responds to the fact that SisPI is executed 4 times a day and has a forecast extension of 36 hours. This allows to generation a small ensemble of 6 members (Figure 2).

The ensemble size can be a point of discussion because several authors work with ensembles larger than 20 members ( [19] Kong et al. 2017), others authors like Tong et al. [20] suggest that with small ensemble a satisfactory analyses can be achieved. By the other hand, the adjustment of the ensemble size responds to the technological limitations because its work with a design that can be used operationally.

This way of calculating the covariance matrix generates a flow-dependent BEC similar to that used in sequential assimilation schemes such as EnKF (Ensemble Kalman Filter). It is based on the fact that the covariances that weight the errors of the model extracted from the set, being dependent on the flow, can better represent the error of the day, as opposed to the isotropic and static characteristics in purely variational schemes [21] .

The same authors suggest that because the purely variational BEC is flow-independent may not adequately represent the structure of a cyclonic vortex, since the not geostrophic and vortex motion of tropical cyclones implies that this

background error may vary with the flow of the system, making it dependent on the characteristics of the flow where it is embedded.

Therefore, the hybrid methods proposed in this research use a combination of both ways to calculate the BEC. The control variables associated with the BECs obtained for this research are detailed in Table 1.

The main advantage of using a hybrid scheme is the fact that it uses small ensembles to obtain flow-dependent perturbations, in these cases the use of static BEC can be beneficial to obtain a high covariance index [18] [20] [21] . On the other hand, the main disadvantage is precise that the ensemble used must well represent the flow of the day, otherwise, the result of the assimilation may not be the most appropriate.

In 3DEnVAR formulation (Equation (1)) the ensemble is valid only for the analysis time [10] [11] :

J ( x ; α ) = B s 1 2 ( x 1 − x b ) T B − 1 ( x 1 − x b ) + B e 1 2 ∑ i = 1 n ( a | i T C − 1 | a i )     + 1 2 [ y − H ( x 1 + x e ) ] T R − 1 [ y − H ( x 1 + x e ) ] (1)

Here J(x,α) is the cost function, x is the analysis field, x_b the background field, C is the diagonal matrix that controls the spatial correlation of α, which in turn represents the control variable used, B_s covariance matrix static, B_e the flow-dependent covariance matrix, B the weight that is established between the matrices, H the observational operator contained within the WRFDA module, which interpolates the grid point values at the location of the observations and transforms the variables predicted by the model in observational quantities, finally R is the observational covariance error also contained within the assimilation package. B_s, B_e and R matrices assuming a zero-mean Gaussian error distribution and are not correlated.

According to [21] and [22] the calculation of the increments is rewritten taking into account the contribution of both matrices (Equation (2)).

x − x b = x 1 + 1 N − 1 ∑ i = 1 N α i x i (2)

In this case, x₁ represents the increase associated with the static BEC, while the expression on the right represents the increase associated with the ensemble perturbation. On the other hand, x_i represents the average perturbation of the

ensemble members, whose size is indicated by N (for this research N = 6).

The weight between the ensemble and static BEC must follow the relationship (Equation (2a)):

1 B s + 1 B e = 1 (2a)

4DEnVAR is very similar, the only difference is that it requires flow-dependent perturbations and observations for multiple time steps (Equation (3)). The weight assigned to the covariance matrices does not vary over time and in this case, a time window of 6 hours is used, 3 before and 3 after the initialization time.

J ( x ; α ) = B s 1 2 ( x 1 − x b ) T B − 1 ( x 1 − x b ) + B e 1 2 ∑ i = 1 n ( a | i T C − 1 | a i )       + 1 2 [ y − H ( x 1 + x e ; k ) ] T R − 1 [ y − H ( x 1 + x e ; k ) ] (3)

where k represents the multiple perturbations used in the algorithm.

For this research, it was decided to study 3 possible combinations of relationships between the static BEC and the flow-dependent BEC. These combinations were established arbitrarily, being, in terms of percentage, 75/25, 50/50, and 25/75, always placing first the value of the variational BEC and in second place the one assigned to the ensemble BEC. The purpose of this design is to give greater importance to one or the other matrix over the background error, as well as equalize the impact of the matrices. From now on, for synthesizing the explanations, the weight value relative to the flow-dependent BEC will be taken as a reference. The nomenclature assigned to the different experiments is reflected in Table 2.

Because the evaluation is made on landfall of tropical storm Eta, it focused on 3 fundamental parameters of the storm, trajectory, intensity, and precipitation.

For this evaluation, continuous verification statistics are used, such as the mean absolute error (MAE), (Equation (4)), the root mean square error (RSME)

(Equation (5)), and Pearson’s correlation (Pearson) (Equation (6)) [23] . Where, in all cases, O represents the observations and P the experiment’s solutions.

MAE = ∑ i − 1 n O i − P i n (4)

RSME = 1 n ∑ i − 1 n ( O i − P i ) 2 (5)

Pearson = ∑ i = 1 n [ ( O i − O ¯ ) ( P i − P ¯ ) ] ∑ i = 1 n ( O i − O ¯ ) 2 ∑ i = 1 n ( P i − P ¯ ) 2 (6)

The evaluation also includes a dichotomous analysis with the use of statisticians calculated through a contingency table (Table 3), which was applied mainly in the evaluation of precipitation. This type of analysis allowed a more detailed description of the behavior of the error of the precipitation forecasted by the different experiments for different thresholds.

The calculated statistics allow for obtaining an estimate of the occurrences that are correctly predicted (Equation (7)) and the proportion of non-occurrences that were incorrectly predicted (Equation (8)) [23] .

H = a a + c (7)

F = b b + d (8)

The critical success index (CSI) is calculated too. However, despite this statistician is used for low-frequency events since it does not take into account the correct negatives, the calculation of the extreme dependency index (EDI) is also introduced (Equation (10)), which can offers a better characterization than the CSI, given that it tends to give very low values when there are significant differences between the frequencies of manifestation of an event, which can lead to difficulties when making an adequate interpretation. This often happens with precipitation because large accumulations tend to occur in small areas and are less frequent.

CSI = a a + b + c (9)

EDI = log ( F ) − log ( H ) log ( F ) + log ( H ) (10)

In the particular case of precipitation, the accumulated every 6 hours measurements by conventional meteorological stations distributed throughout the country was used for evaluation purposes (Figure 3).

Incremental analysis is carried out to measure the impact of each of the first guess proposed designs. The value of the increases or corrections is obtained by subtracting the values of the variables of the analysis field (assimilation result) and the background field (initial condition without assimilation). This step allows understanding of the reason for the model solutions, providing exact values of how the observations and the background error modify the first guess. Higher values of increments suggest a greater impact on the assimilation process.

Since the objective of SisPI is short and very short-term forecasting, the evaluation is made within the first 24 forecasting hours after the initialization.

After reacquiring tropical depression status on November 6, located east of Belize, the system turned toward the east-northeast, a movement favored by the flow resulting from the combination of a mid- and upper-level trough with the circulation of the subtropical anticyclone (Figure 4) [24] .

At the beginning of day 7th, Eta once again reaches the category of a tropical storm, maintaining course, but temporarily accelerating its forward speed during that day to estimated speeds of 28 km/h. With this movement, it made landfall on the southern coast of the central region of Cuba, specifically through the province of Ciego de Ávila, at approximately 09:00 UTC on day 8th, with maximum sustained winds around 100 km/h, according to data from the reconnaissance plane.

The erratic nature of Eta’s track after emerging into the Caribbean Sea and its subsequent interaction with an upper-low pressure system made it very difficult to forecast the track of the tropical system. On the other hand, over the national territory, Eta experienced a reduction in its movement speed and an inflection in its trajectory, changing its course towards the northwest, aspects that tend to complicate this forecast in practice.

The runs initialized on November 7th at 12:00 UTC show great uncertainty during the first 6 to 9 hours (Figure 5(a)). This fact is related to the low ability of the model to locate the center of circulation, a usual difficulty in very weak systems such as the case of Eta at that time. Subsequently, the errors of all methods

are similar, although, in the intermediate terms, the designs with a greater influence of the static BEC tend to lead to less effective solutions than SisPI.

On the other hand, in the experiments started on day 8 at 00:00 UTC, a decrease in the trajectory error and a lower dispersion of the solutions can be seen (Figure 5(b)).

Common points are observed in both initializations. At the first place, the impact of data assimilation is evident, since the initial location of the cyclonic circulation center is more consistent with reality than in the case of SisPI, decreasing the model’s location error. On the other hand, the 4DEnVAR_75 scheme showed the best performance, generating the lowest forecast errors about the other designs.

These results are directly related to the forecast environment surrounding the storm. In this sense, the direction and speed of mean flow in the layer of isobaric surfaces that are between 700 and 400 hPa are decisive for the movement of the storm.

Concerning November 7th initialization, a tendency of the experiments to displace the center of circulation of the low-pressure zone in the height towards the southeast, where the latitudinal error is more significant, is observed (Figure 6).

The schemes that give greater weight to the flow-dependent BEC tend to generate an average circulation associated with this incipient low, a result that is more consistent with reality, which indicates that the set better represents the flow conditions where the storm is embedded than the static BEC.

It can be seen that a further eastward localization of the upper low core leads to solutions that turn more northward to Eta’s best track, which increases the trajectory error. About November 8th initialization, it can be observed that the solutions of the 4DEnVAR_50 and 4DEnVAR_75 experiments better represent the mean circulation of the layer selected, although generating a dipole with a northeast-southeast orientation as a result of a deepening of Eta, indicating an overestimation of storm intensity. In the case of the 3DEnVAR method, the 3DEnVAR_75 experiment better represents these characteristics, coinciding with a better representation of those experiments where the flow-dependent BEC has greater weight (Figure 7).

The intensity forecast presented several difficulties. SisPI generated an overestimation of this indicator, forecasting minimum pressures within category 1 hurricane thresholds, something that never happened. A probable cause of these results lies precisely in the trajectory error since advancing the turn northward and therefore moving the system further westward, kept it longer over the sea, a factor that usually favors the intensification processes.

In both initializations, it is observed how the error is considerably reduced during the first 9 forecasting hours and, after this time, the experiments solutions converge to SisPI (Figure 8). This indicates that the impact of assimilation reaches its maximum effectiveness precisely at this time threshold.

The experiments related to 4DEnVAR method show the lowest errors, as a result of a more realistic reproduction of the synoptic environment unlike 3DEnVAR. In this sense, the 4DEnVAR_50 and 4DEnVAR_75 designs exhibit

the best results, coinciding with the most accurate track forecasts.

These differences between both methods are also related to the volume of assimilated observations, which is much greater in the case of 4DEnVAR since it includes data contained in a 6 hours window time, in contrast to 3DEnVAR, which only assimilates the observations present in the environment of the initialization time. This allows 4DEnVAR to have information, not only about the real state of the atmosphere over the high-resolution domain at the initialization instant but also about the behavior of the real gradients in the assimilation window used.

The way of different experiments solves the interaction of the storm with the upper low significantly influences the intensity forecasts. Eta was subjected to dry and cooler air intrusion from the third and fourth quadrants, this affected the storm’s core, which is usually warm and humid, limiting the intensification processes and leading Eta to remain in a kind of transition between a classic tropical system and a subtropical storm.

The cold, dry air intrusion mechanism works by generating undersaturated convective downdrafts that reduce the wet static energy in the boundary layer, thus limiting the energy available to the storm. In situations like this, it takes several hours for evaporation to recover moisture from the surface boundary layer before the intensification process can resume.

Therefore, this interaction affects the inertial stability of the storm, which is nothing more than its resistance to external forcing [2] . This is due to the inertial stability tends to decrease with height, due to the weakening of the rotational wind and the absolute vorticity of the upper anticyclonic outflow, which makes Eta susceptible to fluctuations of the upper low, a process that is reflected in the experiments, with different accuracy degrees.

The designs that make a better forecast of this extremely complex process were the most successful in predicting intensity. Figure 9 illustrates this by comparing temperature and relative humidity advection forecasts between the 3DEnVAR_25 (low performance) and 4DEnVAR_50 (high performance) designs.

In this example, the 3DEnVAR_25 solution shows cold advection around the core but without interrupting the warm advection generated by deep convection growing around the center of the cyclonic vortex (Figure 9(a)), which is consistent with the limited dry advection around the center of Eta (Figure 9(c)).

On the opposite, 4DEnVAR_50 shows an interruption of the warm advection generated by the growth of deep convection clouds and a rearrangement of the greater convection, expressed in a greater warm advection, to the left side of the circulation center, one of the characteristics that they usually show subtropical cyclones or in transition to extratropicalization. This also coincides with a greater entry of dry air (Figure 9(b) and Figure 9(d)), which limits the intensification of the system and, by affecting the growth of deep convection cells due to interruptions in the flow of humidity, also limits the ability of Eta to produce significant precipitation over the central region.

In general, it can be summarized that all the proposed experiments lead to a reduction in the intensity forecast error, mainly during the first 9 forecasting hours. However, both hybrid schemes lead to an underestimation of the intensity of the storm, as opposed to SisPI which overestimates this feature, mainly at November 8th initialization (Figure 8).

Once again, it is observed that when weighting the background errors fundamentally using the flow-dependent BEC, more realistic solutions are achieved, although, in the case of 4DEnVAR, the variant that gives equal weight to both BECs also generates satisfactory results.

As Eta is a relatively weak system and moves very close to the Guamuhaya mountains, crossing extensive areas of agricultural importance in the country, precipitation is precisely the most important associated phenomena for forecasting purposes, whose forecast, as expected, is more or less accurate in correspondence with the ability to forecast the track and intensity of the storm.

An increase in the ability to forecast precipitation is observed in all the assimilation schemes, fundamentally in the initialization corresponding to November 7th, being somewhat more discreet in those experiments initialized on November 8th (Figure 10).

In the experiments initialized on November 7th, 4DEnVAR_50 and 4DEnVAR_ 75 designs showed the best performance indices for accumulates exceeding 10 mm/6 hours (Figure 10(a)). On the opposite, the experiments carried out with the 3DEnVAR method led to an underestimation of the accumulated precipitation, where the 3DEnVAR_50 design showed slightly superior results to the rest. In this sense, 3DEnVAR_75, with an increase in false alarms, and 3DEnVAR_25, with failures, exhibit a marked underestimation of the precipitation forecast.

The experiments initialized on November 8th show more modest results. In this sense, the highlight is a substantial increase in false alarms, as a consequence

of an overestimation of the humid advection towards the west of the country (Figure 11), together with a greater volume of faults towards the central region due to an increase in dry advection associated with a greater intrusion of dry air from the upper low.

This dry advection overestimation is directly related to the forecast of the approximation of the center of the upper low to Eta core, since those solutions where both circulations are closer, the more depressed the precipitation will be in the center of the storm, which responds, as stated above, to the fact that dry

air weakens the convection, creating undersaturated downdrafts that subsequently cool the surface boundary layer and reduce its moisture.

Analyzing the results taking into account different precipitation thresholds, it is observed that these differ to a certain extent from those shown up to now (Figure 12). This behavior is directly related to the fact that the Eta forecast does not manage to generate enough deep convection around the core and extensive areas of stratiform precipitation predominate, particularly towards the western and central portions of the country.

The experiments reflect this environment, however, it overestimates the spatial distribution of stratiform precipitation while underestimating deep convection zones. These differences are related to the dry advection around the storm core, something that, as has been seen, has a significant weight in the increase of errors in the designs where the influence of the static BEC predominates.

In general, it can be seen that 4DEnVAR method loses effectiveness as the value of the precipitation threshold increases. This means that the method predicts spatially limited deep convection zones, mainly towards the storm center, where the largest accumulated were generated. In this case, the 4DEnVAR_25 experiments, in both initializations, are the ones with the best performance with this method. This result indicates that when these experiments overestimate the convection around the center of Eta, not disrupting the center of the vortex, they improve the correct detections towards the central region and compensate, to some extent, the failures towards other regions of the country.

Related to 3DEnVAR method, the 3DEnVAR_75 experiments show the best results in both initializations. The main difficulties are presented towards the weakest precipitation thresholds, approximately from 0 to 20 mm/6hours. However, for heavy rain (≥50 mm/6hours) it exhibited EDI values above 0.5 in both initializations.