Fig 1.
Common outdoor dengue vector breeding sites in Thailand (from left to right): large jar, bucket, old tire, potted plant, bin, ceramic bowl, cup, vase.
Fig 2.
Performing transfer learning on a pre-trained model by replacing the output layer with new target classes.
Table 1.
Number and percentage of images and containers from each source used for training and testing.
Fig 3.
Examples of containers detected by using Faster R-CNN with new transferred categories.
Fig 4.
Information visualization dashboard.
The choropleth map displays container densities for all sub-districts in Nakhon Si Thammarat province. The top chart on the right shows relative percentages of container types in the whole province. The second and third charts show statistics for the selected sub-district, in this case Krung Ching. When hovering over a subdistrict the data for the subdistrict is displayed. The choropleth map in this figure was produced using ArcGIS version 10.4 (Esri, Redlands, CA, USA). Source of shapefile: United Nations Office for the Coordination of Humanitarian Affairs https://data.humdata.org/dataset/thailand-administrative-boundaries.
Table 2.
Object recognition accuracy at 0.5 recognition confidence threshold for each category of container and grouping all container types.
The average precision is calculated from the precision/recall curve by taking the average over all recall levels.
Fig 5.
Image coverage in each province.
Choropleth map produced using ArcGIS version 10.4 (Esri, Redlands, CA, USA). Source of shapefile: United Nations Office for the Coordination of Humanitarian Affairs https://data.humdata.org/dataset/thailand-administrative-boundaries). Note: White color means no image coverage.
Fig 6.
Container counts vs area covered by GSV images (km2) in a) Bangkok, b) Krabi, and c) Nakhon Si Thammarat.
Fig 7.
Container count vs population at the district level, color coded by province.
Each datum is sized according to population density. (Note: NST = Nakhon Si Thammarat, Pop Density = population density).
Fig 8.
Distribution of relative prevalence of five most common container types (bin, bucket, jar, potted plant, tire) over sub-districts of (a) Bangkok, (b) Krabi and(c) Nakhon Si Thammarat provinces. Kernel density estimation was applied to smooth the values. (Note: Differences in bin widths are due to use of the Freedman-Diaconis rule for automatic binning used in plotting the distributions).
Fig 9.
The relative numbers of containers in Lansaka District of Nakhon Si Thammarat from analysis of GSV images and from manual survey [57].
Values are shown relative to the highest count over the sub-districts for each study (95% confidence interval).
Fig 10.
The relative numbers of containers in five sub-districts of Nakhon Si Thammarat from analysis of GSV images and from manual survey data of outdoor containers obtained from the Thai Ministry of Public Health.
Values are shown relative to the highest count over the sub-districts for each study (95% confidence interval).
Table 3.
Description of detected containers used in comparison with larval surveys for entire year, dengue season and non-dengue season.
Table 4.
Description of Breteau Index data for the entire year used in analyses: Number of surveys per sub-district (N), mean value of BI, and SD.
Table 5.
Description of Breteau Index data for the dengue season used in analyses: Number of surveys per sub-district (N), mean value of BI, and SD.
Table 6.
Description of Breteau Index data for the non-dengue Season used in analyses: Number of surveys per sub-district (N), mean value of BI, and SD.
Fig 11.
Correlation between container density by land area and BI for (A) entire year, (B) dengue season, and (C) non-dengue season, and predicted vs actual values of BI for multivariate linear regression model for (D) entire year, and (E) dengue season, and (F) non-dengue season. The solid line is a linear trendline which is an indication of the linear (Pearson) correlation between the two variables. (Note: shading shows the 99% confidence interval).
Table 7.
Absolute standardized coefficients and p-values from linear regression for dengue season.
The largest absolute values are the most important variables in the regression model.
Fig 12.
Choropleth maps of SAPE values for the multivariate linear models for (A) entire year, and (B) dengue season, where A.1, B.1 are gradient colormaps, and A.2, B.2 are thresholded colormaps using the 25% and 75% quantiles as threshold values. The dashed circle and solid circle delineate the clusters where the model fit are good and poor, respectively. White color represents subdistricts with no data. Choropleth map produced using ArcGIS version 10.4 (Esri, Redlands, CA, USA). Source of shapefile: United Nations Office for the Coordination of Humanitarian Affairs https://data.humdata.org/dataset/thailand-administrative-boundarieson) correlation between the two variables.
Fig 13.
Scatter plots of (a) SAPE residual and (b) AE residual values of sub-district predictions versus Breteau index. The 25% and 75% quantiles are used as thresholds for the categorization into Good, Average, Poor.