Using Machine Learning Models for Predicting Culex Mosquito Habitats and Breeding Patterns in Washington D.C.

doi:10.21203/rs.3.rs-1969585/v1

Download PDF

Research Article

Using Machine Learning Models for Predicting Culex Mosquito Habitats and Breeding Patterns in Washington D.C.

https://doi.org/10.21203/rs.3.rs-1969585/v1

This work is licensed under a CC BY 4.0 License

Version 1

posted

You are reading this older preprint version

Read the latest preprint version →

Culex mosquitoes pose a significant threat to humans and other species due to their ability to carry deadly viruses such as the West Nile and Zika. Washington D.C., in particular, has a humid subtropical climate that makes it ideal as a habitat for mosquito breeding. Thus, tracking mosquitoes’ habitats and breeding patterns in Washington D.C. is crucial for addressing local public health concerns. Although fieldwork techniques have improved over the years, monitoring and analyzing mosquitoes is still difficult, dangerous, and time-consuming. In this work, we propose a solution by creating a Culex mosquito abundance predictor using machine learning techniques to determine under which conditions Culex mosquitoes thrive and reproduce. We used four environmental variables to conduct this experiment: precipitation, specific humidity, enhanced vegetation index (EVI), and surface skin temperature. We obtained sample data of these variables in the Washington D.C. areas from the NASA Giovanni Earth Science Data system, as well as mosquito abundance data collected by the D.C. government. Using these data, we created and compared four machine learning regression models: Random Forest, Decision Tree, Support Vector Machine, and Multi-Layer Perceptron. We searched for the optimal configurations for each model to get the best fitting possible. Random Forest Regressor produced the most accurate prediction of mosquito abundance in an area with the four environment variables, achieving a mean average error of 3.3. EVI was the most significant factor in determining mosquito abundance. Models and findings from this research can be utilized by public health programs for mosquito-related disease observations and predictions.

Artificial Intelligence and Machine Learning

Environmental Engineering

mosquito breeding patterns

machine learning techniques

Culex mosquitoes

environmental variables

Culex mosquitoes are some of the most common species of mosquitoes worldwide and can carry many types of diseases, including the West Nile and Zika viruses (Omodior et al. 2018). At present, there are no licensed vaccines or medicines for such diseases. Culex sp. are native to Africa, Europe, and Asia, and are found worldwide (Soh and Aik 2021). Therefore, a large number of societies are prone to these deadly diseases. Past studies have been limited in their noticeable effects on human populations; however, very little has been done due to the lack of understanding of how these viruses affect the human body (Center of Disease Control 2020). Furthermore, there are currently no established systems that predict mosquito-borne disease outbreaks.

Washington D.C. is known to have a humid subtropical climate which tends to be ideal for mosquito breeding habitats. With an annual rainfall average of 42 inches per year, Washington D.C.’s rainfall is 12 inches higher than the nationwide yearly average of 30.2 inches. Thus, Washington D.C. is prone to abundant mosquito populations because they thrive in relatively wet places (National Centers for Environmental Information 2022). Washington D.C. has a unique environment, with the Potomac River, a freshwater stream running through Washington D.C., serving this area and the areas around it as habitable mosquito microenvironments. Washington D.C.’s natural topography varies, as the areas to the east of the Potomac River range from 100-140m above sea level, while the west of the Potomac River ranges from 40-80m (Topographic-Maps 2022, Fig. 1). The first human case of West Nile Virus in the United States was reported in the District of Columbia (DC Health 2022). In 2018, there were 13 reported human cases, and there were 11 cases in 2019, in a mere 68.34 mi² area (Center of Disease Control 2020). This is evidence that Washington D.C. has a significant public health threat from mosquitos and the West Nile virus, as well as other mosquito-transmitted diseases.

Historically, machine learning models have proven to be valuable tools for predicting trends and patterns. Furthermore, machine learning is significantly helpful in predicting breeding patterns of various animals, including mosquitoes. In this work, we chose four primary machine learning models: Random Forest, Decision Tree, Support Vector Machine, and Multilayer Perceptron. Random Forest regression models were recently used in predicting West Nile Virus positivity rates and abundance in the city of Chicago (Schneider et al. 2021). Random Forest models are constructed with a collection of decision trees to perform classification and regression tasks (Schonlau and Zou 2020). As Random Forest has been heavily used specifically in geology and earth science, we expect that it would also achieve the best performance in our tests. Decision Tree regression, on the other hand, features only one decision tree analysis feature by partitioning data and fitting a simple model for each partition and exists as a special case of Random Forest (Loh 2011). The Support Vector regression model is based on a linear regression that fits its data by a hyperplane in a higher dimension, allowing it to recognize subtle patterns (Basak and Pal 2007). In addition, Multilayer Perceptron regression is a neural network algorithm that consists of layers of nodes and can learn non-linear relations, which could prove to be helpful (Murtagh 1991).

A previous study by Drakou et al. (2020) revealed that the amount of Culex pipiens, Aedes detritus, and Aedes caspius mosquitoes grew due to precipitation increase. Francisco et al. (2021) found connections between the Dengue virus, another disease transmitted by mosquitoes, and environmental and landscape factors, including precipitation, land surface temperature, the normalized difference of vegetation index, and impervious surfaces, including roads. Other studies, such as Madewell et al. (2019), have also discovered connections of mosquito habitat with urbanization. In this project, we found that studying landcover GLOBE data was useful, together with the various environmental factors available from the NASA Giovanni Earth datasets; and our quest was to use machine learning techniques to predict Culex mosquito breeding patterns in Washington D.C. with GLOBE and open-sourced data.

We chose Washington D.C. as our primary area of interest (AOI) due to its high accessibility of open-sourced mosquito data and environmental data, as well as its historical abundance of mosquitoes. We obtained data from the NASA Giovanni Earth science remote sensing data system, the GLOBE database (GLOBE 2022), and the Washington D.C. government data. The data we collected spans from April 2016 to October 2018. Data sets included the daily data of the Average Surface Temperature, Specific Humidity, Precipitation, and enhanced vegetation index (EVI) (NASA Giovanni 2022). The Washington D.C. government provided open-sourced quantitative mosquito data in our respective AOI (Open Data DC 2021).

Initially, when trying to extract data from GLOBE, we ran into the issue of getting inconsistent data, specifically because of the opportunistic nature of this database. Then, instead of investigating the GLOBE Observer Mosquito Habitat Mapper data, we started looking at GLOBE’s land cover data. This helped us immensely by providing the perspective on how the land cover pictures around Washington D.C. correlate to the environmental factors we used in our experiment. The GLOBE data was collected using land cover measurements that were updated daily, including green-down, green-up, tree height, and land cover classifications. The qualitative GLOBE open-source data exhibited that the land cover points were generally concentrated in Ward 3 of Washington D.C., with many GLOBE citizen science users. Using GLOBE and Washington D.C. open data, Ward 3 displayed relatively low container removals (Fig. 2). Therefore, we speculate that Ward 3 will result in somewhat greater mosquito breeding as there are lower container removals which are an ideal breeding habitat for Culex mosquitoes.

2.1 Data Features

We calculated the correlation between mosquito breeding patterns and the environmental variables: precipitation, EVI, specific humidity, and surface skin temperature. We chose these four environmental factors as we hypothesized that they would significantly affect mosquito breeding patterns. We decided to calculate this relationship because it allowed our machine learning models to predict Culex mosquito breeding patterns given certain environmental conditions in our AOI location. The visual derivatives of the mosquito abundance in our AOI location showed high variability over our selected time and were utilized in our models to identify the specific relationships further. The means and ranges of all the data used are shown in Table 1.

Table 1

The mean and ranges of inputs from NASA Giovanni Data and Washington D.C. Open Data.
Factor	Mean	Range
Average Mosquitoes	8.3600	37.720
Precipitation (mm/day)	2.5939	34.623
Specific Humidity (kg/kg)	0.0113	0.0150
Average Skin Surface Temperature (℃)	22.352	25.639
EVI (Spectral Index)	0.3279	0.1882

2.2 Analyzing Trends and Lag

We aligned the ecological variables with the daily mosquito data taken from the government of Washington D.C. We displayed the graphical relationships between each environmental variable and mosquito abundance in Fig. 3. The correlation p-value was calculated by using the Ordinary Least Square Regression model. For a factor to be considered as statistically significant, it needs to have a p-value < 0.05. We had previously hypothesized that there would be a lag in the data due to the lengthy incubation period of mosquito eggs, which would lead ecological variables to only show an effect in the population several days after (Environmental Protection Agency 2022).

In Fig. 3, the top left scatter plot shows the relationship between the precipitation measured in mm/day and the average mosquito populations. The precipitation has a p-value of 0.0023. The top right scatter plot shows the relationship between the specific humidity measured in (kg/kg) and the average mosquito populations. The specific humidity has a p-value of 0.0487. The bottom left scatter plot shows the relationship between the average surface skin temperature measured in ℃ and the average mosquito populations. The average surface skin temperature has a p-value of 0.0393. The bottom right scatter plot shows the relationship between the enhanced vegetation index (EVI) measured in Spectral Index (Band Ratio) and the average mosquito populations. The EVI has a p-value of 5e-6.

However, when we compared the p-values between with a one-week lag and with no lag, we found that most of the variables were statistically more significant (with smaller p-values) with no lag. Temperature was the only one that was more statistically significant with a one-week lag. Peaks in the data also matched better with no lag, as shown in Fig. 4. Certain factors such as humidity and precipitation would not have been statistically significant with the one-week lag. We compared the differences between the p-values for each ecological variable with vs. without the data lag in Table 2.

Table 2

The differences between the p-values of precipitation, temperature, humidity, and enhanced vegetation index (EVI) when with vs. without the data lag. A statistical significance threshold is 0.05.
Variables	p-value: 1 Week Lag	p-value: No Lag
Precipitation	0.349566	0.002316
Temperature	0.021255	0.039277
Humidity	0.581798	0.048652
EVI	0.000020	0.000005

As shown in Table 2, the p-value is less than 0.05 for every environmental variable, verifying that they all significantly affect mosquito populations. This rejects the null hypothesis stating that none of the environmental variables has a statistically significant relationship with the mosquito populations. In particular, the EVI has a much smaller p-value than the other variables, indicating that it is the most correlated factor with mosquito abundance.

2.3 Data Preprocessing

Data used in this study came from different sources and various satellites, thus it was necessary to do plenty of data cleaning. The Washington D.C. government-collected mosquito data was measured twice a week for most of 2016 and once a week for the rest of the time. The data contained the numbers of both females and males of various types of mosquitoes. Because our focus was on predicting mosquito population growth, we only kept data for female Culex mosquitoes. Due to the varying amount of mosquito traps per day, we took the average number of mosquitoes per trap per day for the dates on which mosquito traps were set. We had 108 days from May to October in the years 2016–2018. For each of these days, we collected data for EVI, average surface skin temperature, specific humidity, and precipitation. We found a couple of holes in some of the environmental factor data. For these, we used SciPy’s interpolation method ‘interp1d’ to fill in the gaps, which we found to be the most accurate.

2.4 Training the Models

We trained a variety of models to find the best possible solution. All models were implemented based on the SciKit-Learn python package, and we tested model hyperparameters using the Grid Search Cross Validation tool in SciKit-Learn. We studied four types of models in particular: the Random Forest Regressor model, the Decision Tree model, the Multilayer Perceptron model, and the Support Vector Regression model. The hyperparameters tested and chosen for Random Forest Regressor and Decision Tree Regressor are listed in Tables 3 and 4, respectively. We adopted a 7:3 ratio for training vs. testing data size, using SciKit-Learn’s function “train_test_split” to split them up randomly.

Table 3

Hyperparameters for the Random Forest regressor: values tested and the values chosen with Grid Search.
Hyperparameter	Values Tested	Value Chosen
‘bootstrap’	True, False	True
‘max_depth’	10, 20, 30, 40, 50, 60, 70, 80, 90, 100, 110, None	60
‘max_features’	'auto', 'sqrt'	‘sqrt’
‘n_estimators’	100, 300, 500, 700, 900, 1100, 1300, 1500, 1700, 1900	300
‘min_samples_leaf’	1, 2, 4	1
‘min_samples_split’	2, 5, 10	2

Table 4

Hyperparameters for the Decision Tree regressor: the values tested and the values chosen with Grid Search.
Hyperparameters	Tested Values	Chosen Value
‘splitter’	‘best’, ‘random’	‘random’
‘max_depth’	1, 3, 5, 7, 9	9
‘min_samples_leaf’	1, 2, 3, 4, 5, 6, 7	4
‘min_weight_fraction_leaf’	0.1, 0.2, 0.3, 0.4, 0.5	0.1
‘max_features’	‘auto’, ‘log2’, ‘sqrt’, None	‘auto’
‘max_leaf_nodes’	10, 20, 30, 40, 50, 60, None	60

When testing on the regression models, we measured two metrics to evaluate the prediction accuracy: the mean absolute error (MAE) and the root mean square error (RMSE). These two measurements are both critical in different ways. MAE measures the average magnitude of the error (i.e., the difference between the predicted value and the true value) without caring about the direction. RMSE, on the other hand, measures the square root of the average of the squared differences. MAE is more often used, primarily for comparing model statistics; but it has been found that RMSE works better to represent the model performance when the error is in Gaussian distribution. We believe it is necessary to measure both metrics to understand which model has the best performance (Chai and Draxler 2014). When comparing all four models, we discovered that the Random Forest regressor performed better in both MAE and RMSE than other models. Support Vector Machine performed the worst in RMSE, and Multi-Layer Perceptron performed the worst in MAE. Table 5 details the scores for each model in MAE and RMSE. However, all models did perform similarly.

Table 5

The performance metrics for the Random Forest, Decision Tree, Support Vector Machine, and Multi-Layer Perceptron regression models.
Model	Mean Absolute Error (MAE)	Root Mean Square Error (RMSE)
Random Forest Regressor	3.27046	5.14630
Decision Tree Regressor	3.40613	5.30988
Support Vector Regressor	3.51837	6.08155
Multi-Layer Perceptron Regressor	3.92544	5.40554

We plotted the comparison between the predictions and the actual mosquito abundance values for the four models in Figs. 5, 6, 7, and 8, respectively. We noticed that all models struggled to correctly predict the values for the two days July 24th, 2018, which had an average of 29.45 mosquitoes, and June 28th, 2016, which had an average of 25.27 mosquitoes. Most models predicted that July 24th, 2018, would have about an average of 10 mosquitoes, and June 28th, 2016, would have an average of 15 mosquitoes. These two points had the highest values among all data points, and are about ten mosquitoes more than the next largest. Furthermore, when these two points were removed, the Random Forest Regressor accuracy metrics significantly improved, as shown in Table 6. The MAE was able to improve by 0.65, while the RMSE was able to improve by 1.71.

Table 6

The scores for Random Forest Regressor with and without the outliers.
Condition	Mean Absolute Error	Root Mean Square Error
With Outliers	3.27046	5.14630
Without Outliers	2.62273	3.43861

When looking at the environmental factors, as shown in Fig. 9, we found that these two data points had higher values in most of the environmental factors, especially for June 28th, 2016, which had the highest value for two of the environmental factors. This may explain why while the model didn’t give a close enough prediction for this point, it still predicted it to be a much higher value than most of the other points. However, July 24th, 2018 only had the highest specific humidity level and was in the medium range for the other factors, which is most likely why it was not predicted to have a very high value. Nevertheless, this case represents how the amount of change in environmental factors is not always comparable to the change in the average number of mosquitoes, especially when the average number of mosquitoes is extremely high. Therefore, it is necessary to have a significantly larger amount of data, especially more data to represent the high mosquito abundance cases, to train the models in order for them to handle such cases well.

Our research presents an analytical viewpoint on the relationships between mosquito abundance and environmental factors. Our machine learning models produced highly accurate predictions by correlating our selected factors with increases or decreases in mosquito abundance. Based on the results, we found that EVI has the most significant effect on mosquito abundance populations in our respective AOI. This finding can be applied virtually anywhere across the globe. However, we acknowledge that there are infinitely many anthropogenic and non-anthropogenic factors, including the variables we used, that could affect mosquito abundance in any specific area. Several examples of literature have claimed that temperature and EVI significantly affect mosquito population growth, which partially supports the results from our machine learning models.

Our research could have been improved in several ways. All of our mosquito data was taken from Washington D.C.’s public database, dating from 2016 to 2018. Due to climate change, environmental factors have changed drastically over the years, and our machine learning predictors are more suitable for analyzing the trends a few years ago. Also, due to the opportunistic nature of recording mosquito abundance in an AOI, we did not have complete mosquito abundance data when we ran our machine learning models, leaving us with relatively low amount of data to use. Our models would have created more accurate predictions about isolated areas around Washington D.C. if more consistent mosquito data and additional environmental variables are available.

As we anticipated, the Random Forest regressor provided the best prediction result for mosquito abundance, with EVI as the most influential factor. However, we did find that all models had more or less similar results. All four machine learning models have proven to be useful predictors in analyzing mosquito abundance patterns. Our developed model provides a thorough, easily examinable way to understand the correlations between environmental factors and mosquito populations. These machine learning models can be used in the future for predicting mosquito breeding patterns across various areas of interest, including but not limited to Washington D.C., to help public health organizations in predicting mosquito habitats and mosquito-borne diseases.

Future research may expand the areas of interest to locations outside Washington D.C. and the United States, as training models with data in diverse areas can yield better predictors. Moreover, we hope more ecological variables are available to be used in future research projects, which may lead to more accurate results.

Crediting Contributors

Author Contributions: Conceptualization, I.X., N.S., L.H., and J.W.; data collection, I.X., N.S., L.H., and J.W.; formal analysis, I.X. and N.S.; investigation, L.H., J.W. and M.H.; methodology, I.X.; software, I.X., writing—original draft, I.X., N.S., L.H., J.W., and M.H.; writing—review and editing, I.X., N.S., L.H., J.W., and M.H. All authors have read and agreed to the submitted version of the manuscript.

Data Availability

GLOBE Observer data were obtained from NASA and the GLOBE Program and are freely available for use in research, publications, and commercial applications. GLOBE Observer data analyzed in this project are publicly available at globe.gov/globe-data (accessed on July 5, 2022). Giovanni Earth data were obtained from NASA and are publicly available at https://giovanni.gsfc.nasa.gov/giovanni/ (accessed on July 5, 2022). Washington D.C. mosquito data was provided by the government of the District of Columbia and is publicly available at https://opendata.dc.gov/datasets/DCGIS::mosquito-trap-sites/about (accessed on July 5, 2022).

Acknowledgements

The authors would like to acknowledge the support of the 2022 Earth System Explorers Research Team, NASA STEM Enhancement in Earth Science (SEES) Internship program. Special thanks to Dr. Rusanne Low, Ms. Cassie Soeffing, Mr. Peder Nelson, Dr. Erika Podest, Andrew Clark, and Alexander Greco for their guidance when preparing this work. The NASA Earth Science Education Collaborative leads Earth Explorers through an award to the Institute for Global Environmental Strategies, Arlington, VA (NASA Award NNX6AE28A). The SEES High School Summer Intern Program is led by the Texas Space Grant Consortium at the University of Texas at Austin (NASA Award NNX16AB89A).

Competing interests: The authors declare no competing interests.

Basak, D., & Pal, S. (2007). Support Vector Regression. Statistics and Computing, 11(10), 203–224.

Celestin, M. N., & Musteata, F. M. (2021). Impact of Changes in Free Concentrations and Drug-Protein Binding on Drug Dosing Regimens in Special Populations and Disease States. Journal of pharmaceutical sciences, 110(10), 3331–3344. https://doi.org/10.1016/j.xphs.2021.05.018.

Centers for Disease Control and Prevention (2021, Dec. 17). Final Cumulative Maps and Data. Centers for Disease Control and Prevention. Retrieved July 22, 2022, from https://www.cdc.gov/westnile/statsmaps/cumMapsData.html.

Centers for Disease Control and Prevention (2020, Dec. 7). Prevention. Centers for Disease Control and Prevention. Retrieved July 22, 2022, from https://www.cdc.gov/westnile/prevention/index.html

Chai, T., & Draxler, R. R. (2014). Root Mean Square Error (RMSE) or Mean Absolute Error (MAE)? – Arguments against Avoiding RMSE in the Literature. Geoscientific Model Development, 7(3), 1247–1250. https://doi.org/10.5194/gmd-7-1247-2014.

DC Health (2022). https://dchealth.dc.gov/service/west-nile-virus.

Drakou, K., et al. (2020). The Effect of Weather Variables on Mosquito Activity: A Snapshot of the Main Point of Entry of Cyprus. Int J Environ Res Public Health, 17(4): 1403.

Environmental Protection Agency. (n.d.). EPA. Retrieved July 22, 2022, from https://www.epa.gov/mosquitocontrol/mosquito-life-cycle.

Francisco, M. E., Carvajal, T. M., Ryo, M., Nukazawa, K., Amalin, D. M., & Watanabe, K. (2021). Dengue Disease Dynamics Are Modulated by the Combined Influences of Precipitation and Landscape: A Machine Learning Approach. Science of The Total Environment, 792, 148406. https://doi.org/10.1016/j.scitotenv.2021.148406.

Global Learning to Benefit the Environment (GLOBE) Data User Guide (2019). version 1.0, www.globe.gov.

Loh, W. Y. (2011). Classification and Regression Trees. WIREs Data Mining and Knowledge Discovery, 1(1), 14–23. https://doi.org/10.1002/widm.8.

Madewell, Z. (2019). Comparing Vector and Human Surveillance Strategies to Detect Arbovirus Transmission: A Simulation Study for Zika Virus Detection in Puerto Rico. PLOS Neglected Tropical Diseases.

Murtagh, F. (1991). Multilayer Perceptrons for Classification and Regression. Neurocomputing, 2(5-6), 183–197. https://doi.org/10.1016/0925-2312(91)90023-5.

NASA. (n.d.). Giovanni. NASA. Retrieved July 22, 2022, from https://giovanni.gsfc.nasa.gov/giovanni/.

National Centers for Environmental Information (2022, May 11). Washington D.C. Precipitation. Retrieved July 27, 2022, from https://www.weather.gov/media/lwx/climate/dcaprecip.pdf.

Omodior, O., Luetke, M. C., Nelson, E. J. (2018). Mosquito-Borne Infectious Disease, Risk-Perceptions, and Personal Protective Behavior among U.S. International Travelers. Prev Med Rep, 12: 336–342.

Open Data DC (2021, December 8). Mosquito Trap Sites. Retrieved July 22, 2022, from https://opendata.dc.gov/datasets/DCGIS::mosquito-trap-sites/about.

Schneider, J., Greco, A., Chang, J., Molchanova, M., & Shao, L. (2021). Predicting West Nile Virus Mosquito Positivity Rates and Abundance: A Comparative Evaluation of Machine Learning Methods for Epidemiological Applications. AGU 2021 Fall Meeting. https://doi.org/10.1002/essoar.10509422.1.

Schonlau, M., & Zou, R. Y. (2020). The Random Forest Algorithm for Statistical Learning. The Stata Journal: Promoting Communications on Statistics and Stata, 20(1), 3–29. https://doi.org/10.1177/1536867x20909688.

Soh, S., & Aik, J. (2021). The Abundance of Culex Mosquito Vectors for West Nile Virus and Other Flaviviruses: A Time-series Analysis of Rainfall and Temperature Dependence in Singapore. Science of The Total Environment, 754, 142420. https://doi.org/10.1016/j.scitotenv.2020.142420.

Washington, D.C. Topographic Map, Elevation, Relief. Topographic. (n.d.). Retrieved July 22, 2022, from https://en-nz.topographic-map.com/maps/sqll/Washington-D-C/.

West Nile Virus. West Nile Virus. (n.d.). Retrieved July 22, 2022, from https://dchealth.dc.gov/service/west-nile-virus.

Download PDF

Version 1

posted

You are reading this older preprint version

Read the latest preprint version →

Using Machine Learning Models for Predicting Culex Mosquito Habitats and Breeding Patterns in Washington D.C.

Status:

Version 1

Abstract

Figures

1. Introduction

2. Methods

2.1 Data Features

2.2 Analyzing Trends and Lag

2.3 Data Preprocessing

2.4 Training the Models

3. Results

4. Discussion

5. Conclusion

Declarations

References

Status:

Version 1