Convective weather characterization and prediction using Machine Learning algorithms: analysis for Amazon Region

doi:10.21203/rs.3.rs-5314505/v1

This study presents an objective tool for predicting convective storms (CS) in the Terminal Maneuverer Area of Manaus (TMA-Manaus), Amazon, Brazil, using Machine Learning (ML) algorithms. The occurrence and severity of CS events were characterized by atmospheric discharges (AD) using thresholds. The prediction models leverage AD thresholds as target and 12Z radiosonde data as predictors. The AD climatology between 2012 and 2017 for TMA-Manaus revealed that AD occurs in every month of the year, in this period there were only in 21 days without lightning. The analysis of feature importance for classifying CS stages revealed that the Showalter index, Bulk Richardson Number, Convective Available Potential Energy, Lifted index, and Equilibrium Level are the most relevant thermodynamic indices for classifying the convective state in the region. Results indicate that, for a small amount of AD (69/day), the mean POD and FAR for the ten selected models were 0.92±0.06 and 0.19±0.01, respectively. The QDA algorithm showed the best performance with a POD of 0.99 and a FAR of 0.19. However, as the AD threshold increased up to 5,000 AD per day, a decrease in model performance was observed, because SevereCS is rare compared to CS. The findings suggest that ML models using radiosonde data as predictors are only capable of predicting the occurrence or not of CS with relative accuracy, but the models are not capable of classifying whether it will be severe. The development of this tool marks a significant step towards improving the accuracy and timeliness of CS forecasts, thereby enhancing safety and efficiency of flights in the region.

convective storm

atmospheric discharges

thermodynamic instability indices

machine learning

Aviation impact

Amazon Region

The focus of Aeronautical Meteorology is to improve flight safety and efficiency. As a member of the International Civil Aviation Organization (ICAO), Brazil addresses the recommendations in Annex 3 to the Convention on International Civil Aviation - Meteorological Service for International Air Navigation, through the Department of Airspace Control (DECEA). Meteorological monitoring and forecasting for aviation are managed by the Integrated Aeronautical Meteorology Center (CIMAER), which collaborates with the Aeronautics Command (COMAER) and the Air Navigation Management Center (CGNA), both based in Rio de Janeiro-RJ, along with Brazilian airlines representatives (ICAO, 2018).

With the purpose to address challenges in aeronautical meteorology, DECEA and the Federal University of Rio de Janeiro (UFRJ), specifically the Applied Meteorology Laboratory (LMA), have collaborated through the Cathedra of Aeronautical Meteorology Project. This partnership aims to develop studies and implement tools for short- to very short-term forecasts, improving the ability of aeronautical meteorology services to provide timely and accurate forecasts for flight operations. Several outcomes of this collaboration are documented in França, Albuquerque Neto, and Campos Velho (2023).

Various meteorological phenomena can impact the safety and economics of air operations, including fog, turbulence, ice formation, low cloud cover, atmospheric pressure changes, and especially Convective Storms (CS). The latter can significantly affect all phases of the flight, from takeoff (causing delays or cancellations), cruising altitude (causing need to deviate from convective clouds), to landing (resulting in delays or changes in landing location).

In the Northern Region of Brazil, particularly in the Amazon area, two key factors significantly influence flight operations in terms of safety and fuel economy: the distances between airports and the impact of meteorological conditions. Eduardo Gomes International Airport (SBEG) in Manaus is notably distant from other major urban airports, with approximate distances to Belém-PA (1,300 km), Boa Vista-RR (660 km), Porto Velho-RO (760 km), and Brasília-DF (1,930 km).

Moreover, the importance of studies in the subject of aeronautical meteorology for this region is underlined by the military operations based in Manaus, including supersonic aircraft operations at Ponta Pelada Airport (SBMN). The Manaus Air Base (BAMN), established in 1970, played a crucial role in the region’s development, providing humanitarian support and contributing to national defense strategies. Experience from pilots and meteorologists at BAMN indicate that military aviation is highly sensitive to adverse weather conditions, particularly CS.

Precipitation variability over the Amazon depends on local factors, such as the forest itself, and remote factors, such as the atmospheric and oceanic circulations of the Tropical Atlantic Ocean (e.g., the migration of the InterTropical Convergence Zone) and the Tropical Pacific Ocean (i.e., the El Niño Southern Oscillation phenomenon) (Marengo & Fisch, 2021). Anthropogenic factors must also be considered. Furthermore, Squall Lines often develop on the north-northeast (N-NE) coast of South America and propagate inland towards the Central Amazon, causing appreciable amounts of precipitation (Ferreira & Cavalcanti, 2022).

The prevalence of vertical wind shear and other tropical weather phenomena, such as the South American Monsoon System (SAMS), significantly influence convective activity and storm development. SAMS is a significant climatic phenomenon that affects weather patterns over much of South America, particularly in tropical and subtropical regions.

Currently, forecasting CS in the Terminal Maneuver Area (TMA) of Manaus (TMA-Manaus) involves a certain level of subjectivity on the part of meteorologists, who rely on atmospheric profiles, satellite data and numerical models, due to the lack of specific and objective tools for this purpose.

This study aims to develop an objective tool to predict the occurrence or not of CS for TMA-Manaus using Machine Learning (ML) models. The ground-based Atmospheric Discharges (AD) detected through antennas installed in the area are used to determine the occurrence and severity of CS, and the thermodynamic indices from 12Z radiosonde data are used as input. We expect this tool will help to increase the accuracy and timeliness of forecasts of CS, improving the safety and efficiency of aviation operations in the region.

House and Beebe (1955) described general methods for forecasting CS, aiming to familiarize forecasters with graphs, tools, methods, parameters, definitions, and techniques used in weather forecasting operations. Mueller et al. (2003) proposed a forecasting system of up to 1-hour lead time for locating CS, utilizing surface observations, weather radar data, satellite data, and numerical weather prediction (NWP) models, along with the stages of storm development.

Nascimento (2005) presented a methodology for predicting severe weather conditions, including CS, in Brazil by studying thermodynamic variables from radiosondes and outputs from a mesoscale NWP model. This study showed that certain parameters used in mid-latitude forecasts for the northern hemisphere can be applied to forecast systems in Brazil. Isaac et al. (2006) developed a forecasting system for aviation in Canada, integrating data from NWP models, surface observations, weather radar and satellite data, and a microwave radiometer to generate meteorological forecasts up to 6-hour lead time for major airports.

Queiroz (2009) created a severity index for short-term forecast (nowcasting) of CS using weather radar data and the FORTRACC system. França, Almeida, and Rossete (2016) developed a neural network model for predicting locations of significant atmospheric instability in TMA-Rio de Janeiro, Brazil, utilizing surface observations, radiosondes, rainfall data, and AD.

Paulucci (2017) analyzed the spatial and temporal distribution of cloud-to-ground AD in the Metropolitan Region of Rio de Janeiro (MRRJ) from 2000 to 2016, identifying a peak concentration between 6 PM and 7 PM local time. Hermsdorff (2018) attempted to increase the predictability of CS in MMRJ by studying AD prediction using decision trees and data from multiple sources, including the RINDAT network, surface observations at Rio de Janeiro International Airport (SBGL), and thermodynamic variables from radiosondes.

Gultepe et al. (2019) highlighted the recognized impact of atmospheric processes on aviation since 1900. Bonnet, Evsukoff, and Morales Rodriguez (2020) explored the usability of a weather radar image prediction method using deep learning algorithms to assist in qualitative precipitation nowcasting with up to 1-hour lead time. Soares et al. (2021) proposed two 6-8-hour lead time forecast models of CS using GOES- R thermodynamic indices and ML models for the air route between the states of Rio de Janeiro and São Paulo.

De Castro et al. (2022) developed a tool using AD and ML models to predict CS in TMA-São Paulo up to 5 hours in advance, determining the location, period, and severity of CS. Almeida, França, and Campos Velho (2020) and Nunes, França and Almeida (2023) used AD and radiosonde data to predict CS in TMA-Rio de Janeiro and TMA-Brasília, respectively, using ML models. Building on this research, the present study uses AD to characterize CS and radiosonde data from Manaus to train ML algorithms.

3.1 Studied Area

The TMA-Manaus area is a circle with approximately 100 km (or 50 nautical miles) in radius as shown in Figure 3. This area covers the city of Manaus, including the Eduardo Gomes International Airport (SBEG) and the Ponta Pelada Military Airport (SBMN). The considered AD area is represented by the square circumscribed around the TMA circle (Figure 3), between latitudes of 02º to 04° S, and longitudes of 59º to 61º W.

3.2 Data

The data used in this work comprises two sources (Table 1):

1) AD detected by the network Sferics Timing and Ranging NETwork (STARNET) (Morales et al., 2002), available at www.starnet.iag.usp.br. The STARNET system consists of a set of ground-based VLF (7 - 15 kHz) receiving antennas that detect radio noise emitted by AD in the atmosphere. From this, the system locates AD within the antenna coverage area using the time difference of arrival method. The STARNET network has receiving antennas installed in several regions of Brazil, Central America and African countries. In this study, the Amazon-based antenna was used. The data has a variable frequency as it depends on the occurrence of lightning, so detection can occur on a millisecond scale. In order to identify the occurrence and classify the severity of CS, AD data was accumulated per day, named as Daily Rate of AD (DR-AD);

2) The radiosonde data from the altitude station of SBMN is available twice a day, 00Z and 12Z. In order to evaluate the weather in a pre-convective scenario, only the 12Z radiosonde was considered. We hypothesize that the thermodynamic indices can capture the local atmospheric state and, thus, can be used as predictors for ML algorithms to generate CS forecasts for TMA-Manaus. Therefore, 24 thermodynamic indices from the 12Z radiosonde were used: Showalter index (SHOW), Lifted index (LIFT), LIFT computed using virtual temperature (LFTV), SWEAT index (SWET), K index (KINX), Cross totals index (CTOT), Vertical totals index (VTOT), Totals totals index (TTOT), Convective Available Potential Energy (CAPE), CAPE using virtual temperature (CAPV), Convective Inhibition (CINS), CINS using virtual temperature (CINV), Equilibrium Level (EQLV), Equilibrium Level using virtual temperature (EQTV), Level of Free Convection (LFCT), LFCT using virtual temperature (LFCV), Bulk Richardson Number (BRCH), Bulk Richardson Number using CAPV (BRCV), Temp [K] of the Lifted Condensation Level (LCLT), Pres [hPa] of the Lifted Condensation Level (LCLP), Mean mixed layer potential temperature (MLTH), Mean mixed layer mixing ratio (MLMR), 1000 hPa to 500 hPa thickness (THTK), and Precipitable water [mm] for the entire sounding (PWAT).

The two sources of data cover different periods: STARNET comprises Jan-2012 to Dec-2017, totalizing 2,192 days; SBMN radiosonde comprises Jun-2012 to Jul-2016, totalizing 1,216 days due to absence of radiosondes on several days and in the period from Feb-2015 to Jun-2015. So, for use in ML models, the data period was equalized. Table 1 summarizes the used datasets.

Table 1 Data used

Source

Frequency

Qty.

Data

Focus

Period

Area

STARNET (www.starnet.iag.usp.br)

variable (~ms)

2,422,173

Location (lat., lon.), time occurrence and quantity of AD

Determine the occurrence of CS and study its climatology

2012-2017

Square circumscribed around the TMA circle in Fig. 3

Radiosonde SBMN (weather.uwyo.edu/)

once a day at 12Z

1,216

24 thermod. indices

Predictor variables of CS of ML algorithms

2012-2016

Location: 59.98°W, 3.15°S

3.3 Method

The steps of the method used are summarized as follows:

a) Climatology of AD: Analyze AD on annual, monthly, and hourly scales to identify the period with the highest frequency of occurrence in TMA-Manaus using the entire AD dataset.

b) Daily Rate of AD (DR-AD): Accumulate AD data per day and identify the occurrence of CS and define severity thresholds.

c) Data Equalization: Equalize time series by identifying and removing days with missing SBMN 12Z radiosonde data from the DR-AD dataset.

d) Binary Classification Variables: Create binary variables for the occurrence or non-occurrence of CS (YES/NO) for each severity threshold obtained in step (b), resulting in a binary variable for each threshold.

e) Predictor Variable Selection: Select predictor variables from the 24 thermodynamic indices to be used as input to the ML models through an analysis of inter-feature correlations and percentage discrepancy intra-feature. The intra-feature analysis is computed as, according to the binary variables obtained in step (d).

f) Training and testing samples: For each threshold (obtained in step (b)), randomly split the data into training and testing samples (50-50), based on the variables chosen in step (e) and using the binary variables created in step (d) as the target.

g) Normalization: For each threshold (obtained in step (b)), normalize the training set and apply the same normalization coefficients to the test set.

h) Class Balancing: For each threshold (obtained in step (b)), balance classes by applying the random oversampling technique to the training set. The YES class was randomly replicated, multiplied 1 to 5 times, in other words, the YES class quantity of the smaller threshold was maintained and the YES class quantity of the upper threshold was replicated 5 times.

i) Implementation of ML Algorithms: Implement 10 ML categorical algorithms: Logistic Regression, Random Forest, Extra Trees, Decision Tree, Gradient Boosting (GBoosting), Bagging, Adaptive Boosting (AdaBoost), K-Nearest Neighbors (KNN), Linear Discriminant Analysis (LDA), and Quadratic Discriminant Analysis (QDA). These algorithms were selected for their effectiveness in handling categorical data and their capability to capture intricate patterns within the dataset. The algorithms were implemented and run through the Google Colab virtual environment using the Python package Scikit-Learn (Pedregosa et al., 2011), version 1.2.2, with default hyper parameters. Further details on each algorithm can be found in Annex 1.

j) Training and Testing: For each threshold (obtained in step (b)), train and test the algorithms.

k) Evaluation and Discussion: Evaluate and discuss results using metrics such as Probability of Detection (POD), False Alarm Rate (FAR). A positive event implies the occurrence of CS (or class YES) and a negative event implies the non-occurrence of CS (or class NO).

This section is structured into two distinct parts: the first provides a diagnosis of the spatiotemporal analysis of AD within TMA-Manaus, while the second focuses on the evaluation of the results of a set of ML algorithms in forecasting CS in TMA-Manaus.

4.1 Atmospheric Discharges

To evaluate the distribution of the quantity of AD and establish the severity thresholds of CS, the boxplot and histogram of the amount of DR-AD within TMA-Manaus are presented in Figure 1. The distribution of quartiles was as follows: the 1st Quartile (Q1) at 69.0, the median (Q2) at 283.5, the 3rd Quartile (Q3) at 1063.5, and an Upper Fence [Q3 + (1.5 * IQR)] at 2555.75. These quantities define the severity thresholds. From 2012 to 2017, there were only 21 days with DR-AD = 0, while 248 days exceeded this upper fence value, indicating significant events. We determined that a rate of 5000 AD is related to the occurrence of more severe CS, with considerable potential impacts on air traffic within TMA-Manaus. There were 109 days exceeding 5000 DR-AD.

Figure 2 displays the percentage distribution of AD events from two datasets: the complete set of AD from 2012 to 2017, consisting of 2,422,173 events (labeled as "All"), and the dataset obtained after step 3.3c of the methodology, which aligns AD data with radiosonde data, consisting of 1,692,333 events (labeled as "Used") that were used to train and test the ML algorithms. Figure 2 includes (a) the hourly distribution of AD, (b) the monthly distribution, and (c) the yearly distribution. We include the “Used” distribution to demonstrate its similarity to the “All” distribution.

From Figure 2a, we see that 73.68% of AD events occur between 15Z and 21Z, with 51.2% occurring between 16Z (12:00 Local Time) and 19Z (15:00 Local Time), indicating that the afternoon period is the most impactful for aviation, as expected. Since the input data is acquired at 12Z, the forecast lead time is set to 4 hours. Figure 2b shows that the highest AD occurrences are during the spring months, from August to November. Figure 2c indicates that 56.1% of AD events occurred between 2013 and 2014. No correlation with the El Niño Southern Oscillation phenomenon was observed; other climate variability parameters, such as Sea Surface Temperature and variation in the position of the ITCZ, were not evaluated.

Figure 3 illustrates the monthly spatial distribution of AD density in the studied area in the period of 2012 to 2016. As seen in Figure 2b, months with higher AD present a more homogeneous spatial distribution, while the other months present cores of AD. But it is interesting to note that AD occurs in every month of the year.

4.2. ML Algorithms

From the SBMN 12Z radiosonde, each of the 24 thermodynamic indices encapsulates critical aspects of atmospheric conditions, offering valuable insights into the potential for classifying convective activity. To optimize the efficiency of the ML models while minimizing redundancy and computational complexity, we conducted an analysis of inter-feature correlations (methodology step 3.3e). This procedure improves the interpretability of the trained model and enhances its predictive performance by mitigating multicollinearity and over fitting.

By identifying and eliminating highly correlated features, the dataset was streamlined to focus on the most informative and discriminative predictors. The resulting subset of features represents a balanced selection that captures essential atmospheric thermodynamics relevant to CS prediction. After this processing, 16 features remained out of the initial 24: SWET, BRCH, CAPE, LIFT, SHOW, KINX, EQLV, CINS, CINV, LFCT, VTOT, LFCV, MLTH, PWAT, THTK, and LCLT.

Table 2 presents the percentage discrepancy among the 16 features for CS and non-CS events according to the defined severity threshold, with DR-AD > k identified as YES where k = 63, 283.5, 1063.5, 2555.75, 5000, and NO otherwise. It is expected that the features with the greatest discrepancy between YES and NO events are particularly influential. The highest discrepancy values are linked to the following 5 features: SHOW, BRCH, CAPE, LIFT, and EQLV. These 5 indices are pivotal indicators of CS occurrence over TMA-Manaus, given their capacity to assess atmospheric stability and wind shear.

For example, the SHOW index evaluates thunderstorm potential in weather forecasting, with positive values indicating stable atmospheric conditions less conducive to thunderstorm formation, while negative values suggest increased instability, raising the likelihood of convective activity and thunderstorms (Doswell III, Davies-Jones & Keller, 1993). Similarly, BRCH values denote heightened instability, promoting storm development through convective cloud formation and heavy precipitation (Brooks & Craven, 2002). Moreover, BRCH considers vertical wind shear (Holton, 2004), prevalent in tropical regions like Manaus, influencing storm structure and intensity.

CAPE represents the available energy driving convective cloud development, with elevated values indicating heightened convective potential due to factors such as humidity, warm temperatures, and solar heating, facilitating atmospheric instability and convective cloud formation (Moncrieff, 2010). LIFT assesses atmospheric stability by comparing the temperature of a lifted air parcel to its environment at a specified altitude, typically 500 hPa above ground level. Negative LIFT values suggest warmer, less dense parcels, indicating instability favorable to convective activity, while positive values imply stability, inhibiting convective development (Emanuel, 1994). The EQLV marks the altitude where a lifted air parcel becomes cooler than its surrounding environment, ceasing to ascend freely (Wallace & Hobbs, 2006).

Together, these features offer crucial insights into atmospheric thermodynamics, essential for comprehending and forecasting CS in TMA-Manaus. The following discussion specifically addresses the relevance of these features during ML training.

Table 2 Percentage discrepancy intra-feature for DR-AD > k, identified as YES where k = 63, 283.5, 1063.5, 2555.75, 5000 and NO otherwise

		Percentual discrepancy according to DR-AD threshold
Feature		69	283.5	1063.5	2555.75	5000
1	SHOW	114.59	139.03	171.85	173.26	258.96
2	BRCH	42.90	4.98	39.08	39.44	50.11
3	CAPE	34.89	33.29	26.37	26.59	30.60
4	LIFT	47.18	37.84	27.64	19.63	20.90
5	EQLV	18.85	15.49	12.35	11.74	14.12
6	CINS	27.24	21.95	7.89	9.34	5.79
7	KINX	8.75	7.00	6.20	5.45	5.91
8	CINV	20.61	14.17	0.53	0.11	2.95
9	SWET	2.09	2.31	1.21	1.16	1.69
10	LFCT	3.57	2.79	1.65	1.89	1.38
11	VTOT	0.80	0.81	1.32	0.71	0.90
12	LFCV	2.38	1.94	0.81	0.96	0.55
13	MLTH	0.12	0.05	0.00	0.05	0.12
14	PWAT	6.12	4.33	1.90	0.24	0.07
15	THTK	0.06	0.04	0.08	0.11	0.07
16	LCLT	0.28	0.21	0.08	0.02	0.03

Each ML algorithm (methodology step 3.3i) was executed three times, using 5, 16, and 24 features as input, according to the order of relevance established in the index of Table 2. Figure 4 presents five scatterplots of POD vs. FAR for the ML algorithms for the 5 DR-AD thresholds. Figure 5 summarizes the mean behavior of POD and FAR results using the 16 selected features. To show the CS prediction performance for the lowest DR-AD threshold (69), Table 3 presents the results of POD and FAR, for each model and the number of input features used in the run, ranked by POD value.

From Figure 4, we can observe the decrease in model performance and the increase in model dispersion as the DR-AD threshold increases. Lower thresholds provide the best results, probably due to a greater number of YES events. Also, there is a high false positive (false negative) rate for low (high) thresholds, since YES (NO) events have a higher number (Figure 1b).

Evaluating the results of DR-AD=69, the set of ML models demonstrates quite similar and acceptable performances, with mean POD and FAR values of 0.92±0.06 and 0.19±0.01, respectively, as illustrated in Figures 4a and 5. From Table 3, the best performance was achieved by the QDA algorithm, with a POD of 0.99 and FAR of 0.19. Still from Table 3, the models LDA and Logistic Regression present comparative results to QDA for a smaller number of features. The Decision Tree presented the worst results.

Table 3 Performance of the 10 ML algorithms for DR-AD=69 using 5, 16, and 24 features. The results are ranked by POD value

Model	# Features	POD	FAR
QDA	24	0.99	0.19
LDA	5	0.99	0.20
Logistic	5	0.99	0.20
ExtraTrees	24	0.96	0.18
ExtraTrees	16	0.96	0.18
Logistic	24	0.96	0.18
QDA	5	0.96	0.19
Logistic	16	0.96	0.19
RF	24	0.95	0.18
LDA	24	0.95	0.18
LDA	16	0.95	0.18
RF	16	0.94	0.18
KNN	24	0.94	0.19
GBoosting	24	0.93	0.17
ExtraTrees	5	0.93	0.19
KNN	5	0.93	0.19
KNN	16	0.93	0.19
GBoosting	5	0.93	0.20
GBoosting	16	0.92	0.18
AdaBoost	16	0.92	0.19
RF	5	0.92	0.19
AdaBoost	5	0.91	0.18
QDA	16	0.90	0.18
AdaBoost	24	0.89	0.18
Bagging	16	0.87	0.17
Bagging	5	0.86	0.18
Bagging	24	0.85	0.17
DecisionTree	16	0.80	0.19
DecisionTree	5	0.80	0.20
DecisionTree	24	0.76	0.18

This study evaluated the performance of 10 ML models in predicting the occurrence of CS using thermodynamic indices derived from the SBMN 12Z radiosonde data as input and ground-based AD to characterize CS. The AD climatology between 2012 and 2017 revealed that AD occurs in every month of the year, in this period there were only in 21 days without lightning. By employing a systematic feature selection process, we identified a subset of 16 out of the original 24 indices that were most informative for CS prediction, thereby reducing redundancy and computational complexity. The analysis revealed that 5 indices, SHOW, BRCH, CAPE, LIFT, and EQLV were particularly influential in distinguishing between CS and non-CS events, highlighting their importance in assessing atmospheric stability and wind shear.

The evaluation of the ML models, which included multiple runs with varying numbers of selected features and different thresholds of DR-AD, demonstrated that these models are only capable of predicting the occurrence or not of CS with relative accuracy. The QDA algorithm exhibited the best performance with a POD of 0.99 and FAR of 0.19 for DR-AD = 69, using the 24 features as input. The models LDA and Logistic Regression present comparative results to QDA for a smaller number of features. The decrease in performance as the DR-AD threshold increases, because Severe CS is rare compared to CS, indicates challenges in accurately predicting Severe CS events, which is a critical limitation for practical applications in regions like TMA-Manaus.

In conclusion, our findings highlight the potential of ML models in enhancing CS predictions using radiosonde data and AD. However, to improve their practical utility, especially to use in operational meteorology, future research should focus on integrating additional data sources, such as high-frequency atmospheric infrared profiler, and developing to enhance the models' ability to predict event severity.

Author Contributions: All authors contributed to the study's conception and design. Material preparation, data collection, pre-processing, and analysis were performed by H.C.B and S.M.B; S.M.B. and G.B.F made the experiments. G.B.F. wrote the first draft of the manuscript and all authors commented on previous versions of the manuscript. All authors read and approved the final manuscript.

Funding: This study was funded by the Department of Airspace Control via the Brazilian Organization for Scientific and Technological Development of Airspace Control (CTCEA) (SCHOLARSHIP GRANT: 002-018/COPPETEC_CTCEA)

Data Availability Statement: The datasets presented in this study are available on request from the corresponding author.

Acknowledgments: This article comprises part of the master dissertation of the first author. The authors are grateful to Professor Carlos Augusto Morales Rodriguez, from USP, for providing the atmospheric discharge dataset used in this research.

Conflicts of Interest: The authors declare no conflict of interest.

Almeida, V. A. de, França, G. B., & Campos Velho, H. F. de (2020). Short-range forecasting system for meteorological convective events in Rio de Janeiro using remote sensing of atmospheric discharges. International Journal of Remote Sensing, 41(11), 4372–4388. https://doi.org/10.1080/01431161.2020.1717669
Bonnet, S. M., Evsukoff, A., & Morales Rodriguez, C. A. (2020). Precipitation Nowcasting with Weather Radar Images and Deep Learning in São Paulo, Brasil. Atmosphere, 11, 1157. https://doi.org/10.3390/atmos11111157
Breiman, L. (1996). Bagging predictors. Machine Learning, 24(2), 123-140.
Breiman, L. (2001). Random forests. Machine Learning, 45(1), 5-32. https://doi.org/10.1023/A:1010933404324
Breiman, L., Friedman, J. H., Olshen, R. A., & Stone, C. J. (1984). Classification and regression trees. Wadsworth International Group.
Brooks, H. E., & Craven, J. P. (2002). A database of proximity soundings for significant severe thunderstorms, 1957–1993. Bulletin of the American Meteorological Society, 83(1), 88-100.
Cover, T. M., & Hart, P. E. (1967). Nearest neighbor pattern classification. IEEE Transactions on Information Theory, 13(1), 21-27. https://doi.org/10.1109/TIT.1967.1053964
Cox, D. R. (1958). The regression analysis of binary sequences. Journal of the Royal Statistical Society: Series B (Methodological), 20(2), 215-232. https://doi.org/10.1111/j.2517-6161.1958.tb00292.x
De Castro, J. N., França, G. B., Almeida, V. A. de, & Almeida, M. V. de (2022). Severe convective weather forecast using machine learning models. Pure and Applied Geophysics, 179(8), 2945-2955. https://doi.org/10.1007/s00024-022-03088-8.
Doswell III, C. A., Davies-Jones, R., & Keller, D. L. (1993). Thunderstorm forecasting: A review. Meteorological Monographs, 21, 27-59.
Emanuel, K. A. (1994). Atmospheric Convection. Oxford University Press.
Ferreira, N., & Cavalcanti, I. (2022). Sistemas meteorológicos atuantes no Brasil. São Paulo: Oficina de Textos. ISBN 978-65-86235-52-4.
Fisher, R. A. (1936). The use of multiple measurements in taxonomic problems. In Annals of Eugenics, 7(2), 179-188. https://doi.org/10.1111/j.1469-1809.1936.tb02137.x
França, G. B., Albuquerque Neto, F. L., & Campos Velho, H. F. (2023). Nowcasting using machine learning and deterministic models: a Brazilian initiative to improve aviation meteorology. Rio de Janeiro, RJ: Editora da Universidade da Força Aérea. 280 p. https://hdl.handle.net/11422/22155
França, G. B., Almeida, M. V. de, & Rossete, A. (2016). An Automated Nowcasting Model of Significant Instability Events in the Flight Terminal Area of Rio De Janeiro, Brazil. Atmospheric Measurement Techniques. 9, 2335–2344. https://doi.org/10.5194/amt-9-2335-2016.
Freund, Y., & Schapire, R. E. (1997). A decision-theoretic generalization of on-line learning and an application to boosting. Journal of Computer and System Sciences, 55(1), 119-139. https://doi.org/10.1006/jcss.1997.1504
Friedman, J. H. (2001). Greedy function approximation: A gradient boosting machine. In Annals of Statistics, 29(5), 1189-1232. https://doi.org/10.1214/aos/1013203451
Geurts, P., D. Ernst, & Wehenkel, L. (2006). Extremely randomized trees. Machine Learning, 63(1), 3-42. https://doi.org/10.1007/s10994-006-6226-1
Gultepe, I., Sharman, R., Williams, P. D., Zhou, B., Ellrod, G., Minnis, P., Trier, S., Griffin, S., Yum, S. S., Gharabaghi, B., Feltz, W., Temimi, M., Pu, Z., Storer, L. N., Kneringer, P., Weston, M. J., Chuang, H., Thobois, L., Dimri, A. P., Dietz, S. J., França, G. B., Almeida, M. V. de & Albquerque Neto, F. L. (2019). A review of high impact weather for aviation meteorology. Pure and applied geophysics, 176, 1869-1921. https://doi.org/10.1007/s00024-019-02168-6
Hastie, T., Tibshirani, R., & Friedman, J. (2009). The elements of statistical learning: Data mining, inference, and prediction. Springer. https://doi.org/10.1007/978-0-387-84858-7
Hermsdorff, J. (2018). Previsão de instabilidade atmosférica significativa usando árvore de decisão na região metropolitana do Rio de Janeiro. Dissertation, Universidade Federal do Rio de Janeiro, 109 f.
Holton, J. R. (2004). An Introduction to Dynamic Meteorology. Elsevier Academic Press.
House, D.C., & Beebe, R.G. (1955). SELS Forecasting Procedures. US Weather Bureau.
ICAO (International Civil Aviation Organization) (2018). Annex 3 to the Convention on International Civil Aviation, Meteorological Service for International Air Navigation. 20th Edition. ICAO.
Isaac, G., Bailey, M., Cober, S., Donaldson, N., Driedger, N., Glazer, A., Gultepe, I., Hudak, D., Korolev, A., Reid, J., Rodriguez, P., Strapp, J. & Fabry, F. (2006). Airport Vicinity Icing and Snow Advisor (AVISA). In 44th AIAA Aerospace Sciences Meeting and Exhibit. https://doi.org/10.2514/6.2006-1219
Marengo, J., & Fisch, G. (2021). Clima da Região Amazônica. In Cavalcanti, I. F. A.; Ferreira, N. J. Clima das regiões brasileiras e variabilidade climática (1st ed., pp. 9-24). São Paulo: Oficina de Textos. ISBN 978-65-86235-24-1.
Moncrieff, M. W. (2010). The multiscale organization of moist convection and the intersection of weather and climate. Meteorological Monographs, 37, 407-418.
Morales, C. A., Weinman, J. A., Branch, M. S., Anagnostou, E. N., Goodman, S., Center, G. H., & Williams, E. (2002). Continuous Long-Range Thunderstorm Monitoring by a VLF Receiver Network. Part II: Cloud-to-Ground and Intra-Cloud Detection Efficiency. Colorado State University.
Mueller, C., Saxen, T., Roberts, R., Wilson, J., Betancourt, T., Dettling, S., Oien, N., & Yee, J. (2003). NCAR auto-nowcast system. Weather Forecast,18, 545–561.
Nascimento, E. L. (2005). Previsão de tempestades severas utilizando-se parâmetros convectivos e modelos de mesoescala: uma estratégia operacional adotável no Brasil? Revista Brasileira de Meteorologia, 20, 121-140.
Nunes, M. P., França, G. B., & Almeida, V. A. (2023). Convective Meteorological Events Forecast using Machine Learning models. PREPRINT (Version 1) available at Research Square. https://doi.org/10.21203/rs.3.rs-2802341/v1
Paulucci, T. B. (2017). Caracterização Espaço-Temporal de Descargas Atmosféricas e Tempestades Elétricas na Região Metropolitana do Rio de Janeiro entre 2001 e 2016. Monography, Universidade Federal do Rio de Janeiro, 73p.
Pedregosa, F., Varoquaux, G., Gramfort, A., Michel, V., Thirion, B., Grisel, O., Blondel, M., Prettenhofer, P., Weiss, R., Dubourg, V., Vanderplas, J., Passos, A., Cournapeau, D., Brucher, M., Perrot, M. & Duchesnay, É. (2011). Scikit-learn: Machine Learning in Python. Journal of Machine Learning Research, 12, 2825-2830.
Queiroz, A. P. (2009). Monitoramento e previsão imediata de tempestades severas usando dados de radar. Dissertation, Instituto Nacional de Pesquisas Espaciais: São José dos Campos.
Soares, C. M., França, G. B., Valdonel, M. A. de, & Almeida, V. A. (2021). Study of GOES- R Thermodynamic Indices for Short-Term Forecasting of Convective Weather Events Using Machine Learning. Pure and Applied Geophysics, 178, 4651-4663. https://doi.org/10.1007/s00024-021-02889-7.
Wallace, J. M., & Hobbs, P. V. (2006). Atmospheric science: An introductory survey. Academic Press.

No competing interests reported.

Convective weather characterization and prediction using Machine Learning algorithms: analysis for Amazon Region

Status:

Version 1

Abstract

Figures

1. Introduction

2. Literature review

3. Data and Method

4. Results

5. Conclusion

Declarations

References

Additional Declarations

Status:

Version 1