Netherlands Red Cross - 510

This dataset has been consolidated from NDRMC / DROMIC reports from 2014 - 2020 to summarize the number of people affected and houses damaged in the Philippines as a result of typhoons.

Izabal department (GT-IZ), Guatemala: AI predictions of building footprint on Bing Maps images (approximately 2016-2019), see https://github.com/rodekruis/automated-building-detection. Produced in support to DRRT Guatemala for hurricane Eta and Iota. Coordinate reference system: WGS 84 / EPSG:4326

Southern Guatemala: AI predictions of building footprint on Bing Maps images (approximately 2016-2019), see https://github.com/rodekruis/automated-building-detection. Produced in support to DRRT Guatemala for hurricane Eta and Iota. Coordinate reference system: WGS 84 / EPSG:4326

OpenStreetMap buildings of Camarines Sur (as of 1-nov-2020) and AI predictions on Bing Maps images (approximately 2016-2019). Produced in support of Philippines Red Cross for typhoon Goni (1-nov-2020) Coordinate reference system: WGS 84 / EPSG:4326

For the floods in Southern Malawi of March 2019, we have combined flood extent maps (Sentinel) with HRSL settlement/population grid. This results in a calculation of # of affected buildings/people per district. The results is shared through maps and in a shapefile. 1. Data sources Sentinel 1 Imagery from 7th of March 2017 Sentinel 2 Imagery from 10th/12th/14th of March 2017 HRSL population data Facebook Connectivity Lab and Center for International Earth Science Information Network - CIESIN - Columbia University. 2016. High Resolution Settlement Layer (HRSL). Source imagery for HRSL © 2016 DigitalGlobe. Accessed 9 March 2019. 2. Good to know The flood extent for Nsanje district was separately added on March 14th, to the existing flood extent for the main area from March 12th. 3. Methodology A. Flood Extent Based on SAR The following steps were used to detect flood extent(water/no water). In SNAP tool the raw data downloaded from sci-hub Copernicus was processed to calibrate image for atmospheric correction, spike filter and terrain correction(This is mainly for Sentinel 1 data). Finally defining water no water based on a threshold applied on the corrected image. Defining a threshold is always a challenge in SAR image analysis for flood detection, we collected data from the field to define this threshold. For Sentinel 2 as a first step cloud filter was calculated by applying a combined threshold on Band 2 and Band 10. The cloud mask shown in the figure below didn’t capture shadows of clouds, these were miss interpreted by the flood algorithm as water/flood. To correct this areas with more cloud cover were clipped out with a polygon. To define water no water based on sentinel data we used NDWI index, the treshold is adjusted based on data collected from the field Validation points were collected by Field team tested different values and check if the threshold identified fits with observation. The complete methodology how to detect flooding based on Sentinel 1 data and SNAP toolbox is documented in ESA website. B. Affected People To calculate number of affected people per each admin level, flood extent map is combined with HRSL population data. This is done in two steps: First, in step 1, we calculate a raster, which multiplies the population grid with the flood grid, such that we are left with only "population in flooded area". This is done using raster calculator where population density raster was multiplied by flood extent raster, which has a value of 0 for no flood and 1 for flood areas. Note that the flood extent grid was first resampled to match it to the population grid. This whole exercise is repeated for settlement/buildings instead of population. Step 2: We apply zonal statistics per TA to calculate total number of buildings/people affected in each admin level. For each Admin level2 estimated number of affected people and affected houses are plotted in the map. The zonal statistics data used for plotting can be found in the shape file.

Update 15/09 (POST-EVENT) Now that the typhoon has passed the country, the model is not run with forecasted wind speeds and typhoon track any more, but with actual estimated wind speeds and typhoon track. They come from the same source (Tropical Storm Risk - UCL), and are of the exact same format. All output (map in PDF, data in Excel and in Shapefile) is of the exact same format and interpretation. Full methodology 1. Based on existing Priority Index model: 510 has previously developed the Priority Index model for typhoons in the Philippines One day after a typhoon has passed the Philippines .. .. the model predicts ‘% of completely damaged houses’ per municipality Based on 12 large typhoons in the last 5 years in the Philippines, for which detailed damage reports were available through NDRRMC (https://www.ndrrmc.gov.ph/) For these same events, we also collected possible explanatory indicators, such as wind speed (event-specific) and building materials of houses (PH national census). We built a statistical model, which could explain differences in damage on the basis of differences in wind speed and building materials (etc.) When dividing all municipalities in 5 equal damage classes (class 1 being the 20% municipalities with lowest damage; class 5 the 20% with highest damage) .. .. we found that in 73% of the cases we are at most 1 class off. 2. Mangkhut methodology: In the case of typhoon Mangkhut, we are dealing with an upcoming typhoon, which is still awaiting landfall on Saturday 15/09. This is a new situation, which requires the following noteworthy changes in methodology. Our wind speed source (UCL Tropical Storm Risk) has – in addition to post-event wind speed data as used above – also forecast wind speed data for 5 days ahead. This forecasted wind speed (and typhoon track) are plugged as input into the above-mentioned prediction model, which - still in combination with building materials - lead to the predicted damage class per municipality that can be seen in the map. Note that the results are strongly dependent on the input of windspeed, which is itself still an unknown. (see accuracy below). 3. How to use this product: The map contains damage classes (1-5) per municipality. As such, we advise to put priority on municipalities in damage class 5, and depending on available resources continue with class 4, etc. This damage class is based on ‘% of houses that are completely damaged’. As priority might also be based on exposure and vulnerability, we have added to the Excel a couple of relevant indicators, from the Community Risk Assessment dashboard. PRC can decide if and how to combine these various features. If needed, 510 can be asked for assistance of course. 4. Important notes: ACCURACY: it should be realized that during the course of the coming 3 days, the typhoon might change course, or increase/decrease in terms of strength. This will affect the quality of these predictions. The accuracy figure of 73% that is mentioned in the post-event case should be seen as an upper bound. Given the added inaccuracy of wind speed, the overall accuracy will be lower. This damage prediction is only about completely damaged houses, not about partially damaged houses. We only included municipalities that are within 100km of the forecasted typhoon track, as we have seen from previous typhoons (with comparable wind speeds) that damage figures outside of this area are low. 5. Sources The wind speed is provided by Tropical Storm Risk (University College London). It is the ‘maximum 1-minute sustained wind speed’. An average of this is calculated per municipality. (Latest forecast date: 2018-09-14 00:00 UT >> 7:00AM Manila time) Typhoon track (from which ‘distance to typhoon track’ per municipality is calculated), is provided by UCL as well. (Latest forecast date: 2018-09-14 00:00 UT >> 7:00AM Manila time) Additionally, various wall and roof type categories from the Philippines national census. The model uses 2010 census data, as it was developed using this data (2015 census data on municipality level only became available in 2018). The 2015 census data could not be easily plugged in, because of some differences in roof/wall categories. We believe that this would not change the result much though, as even if there are large differences from 2010 to 2015, these would still be dominated by wind speed effects in the model. All additional indicators, that are added to the Excel table (population, poverty) are derived from the Community Risk Assessment dashboard (Go to this link and click ‘Export to CSV’ on top-right.) The sources for these indicators can be found in the dashboard itself.

This dataset contains: windspeeds of Typhoon Nina rainfall of Typhoon Nina Priority Index of Typhoon Nina The predicted priority index of Typhoon Nina is produced by a machine learning algorithm that was trained on five past typhoons: Haiyan, Melor, Hagupit and Rammasun and Haima, It uses base line data for the whole country, combined with impact data of windspeeds and rains, and trained on counts by the Philippine government on houses damaged and completely destroyed. The output is a weighted index between partially damaged and completely damaged, where partially damaged is counted as 25% of the completely damaged. This has proven to give he highest accuracy. The absolute number of houses damaged / people affected is insufficiently validated at the moment, and should just be used for further trainng and ground-truthing. Scoring The model has an best r2 score of 0.794933727 and an accuracy of 0.699470899 Data sources: Administrative boundaries (P_Codes) - Philippines Government; Published by GADM and UN OCHA (HDX) Census 2015 (population) - Philippine Statistics Authority; received from UN OCHA (HDX) Avg. wind speed (mph) - University College London Typhoon path - University College London Houses damaged - NDRRMC Rainfall - GPM Poverty - Pantawid pamilyang pilipino program (aggregated) Roof and wall materials New geographical features All the columns with feat_ indicates the importance of that feature, if not present that feature was not used. learn_matrix name of the learning matrix with the 5 typhoons run_name unique run name (pickle files and csv files have this name for this model) typhoon_to_predict name of a new typhoon to predict val_accuracy accuracy based on 10 categories of damage 0% 10% 20% … val_perc_down perc of underpredicted categories val_perc_up perc of overpredicted categories Val_best_score best r2 score Val_stdev_best_score error on best score based on the CV Val_score_test r2 score on the test set (this should be around +- 5% of the previus number to not overfit Val_mean_error_num_houses average error on the number of houses val_median_error_num_houses median val_std_error_num_houses std deviation of the errors (lower is better) Algorithm developed by 510.global the data innovation initiative of the Netherlands Red Cross.

Datasets and map of dam break near solai in Kenya on May 9th 2018.

In this analysis we have combined several data sources around the floods in Bangladesh in August 2017. Visualization See attached map for a map visualization of this analysis. See http://bit.ly/2uFezkY for a more interactive visualization in Carto. Situation Currently, in Bangladesh many water level measuring stations measure water levels that are above danger levels. This sets in triggers in motion for the partnership of the 510 Data Intitiative and the Red Cross Climate Centre to get into action. Indicators and sources In the attached map, we combined several sources: Locations of waterlevel stations and their respective excess water levels (cms above danger level) at 14/08/2017 (Source: http://www.ffwc.gov.bd/index.php/googlemap?id=20) Population density in Bangladesh to quickly see where many people live in comaprison to these higher water-level stations. (Source: http://www.worldpop.org.uk/data/summary/?doi=10.5258/SOTON/WP00018 >> the People per hectare 2015 UN-adjusted totals file is used.) Vulnerability Index: we constructed a Vulnerability Index (0-10) based on two sources. First poverty incidence was collected from Worldpop (Source: http://www.worldpop.org.uk/data/summary/?doi=10.5258/SOTON/WP00020 >> The estimated likelihood of living below $2.50/day). Second, we used a Deprivation Index which is estimated in the report Lagging District Reports 2015 (Source: http://www.plancomm.gov.bd/wp-content/uploads/2015/02/15_Lagging-Regions-Study.pdf > Appendices > Table 20), which combines many socio-economic variables into one Deprivation Index through PCA analysis. Detailed methodology Vulnerability The above-mentioned poverty source file is on a raster level. This raster level poverty was transformed to admin-4 level geographic areas (source: https://data.humdata.org/dataset/bangladesh-admin-level-4-boundaries), by taking a population-weighted average. (Source population also Worldpop). The district-level PCA components from abovementioned reports were matched to the geodata based on district names, and thus joined to the admin-4 level areas, which now contain a poverty value as well as Deprivation Index value. Note that all admin-4 areas within one district (admin-2) obviously all have the same value. The poverty rates do differ between all admin-4 areas. Lastly, both variables were transformed to a 0-10 score (linearly), and a geomean was taken to calculate the final index of the two. A geomean (as opposed to an arithmetic mean) is often used in calculating composite risk indices, for example in the widely used INFORM-framework (www.inform-index.org).

A crude version of the INFORM risk-framework is applied to Enumeration Areas (which is unofficial, but is deeper than admin-3), in Southern Malawi. This is done specifically for area selection regarding the ECHO2 project in 3 TA's: Mwambo (Zomba district), Makhwira (Chikwawa district) and Ndamera (Nsanje district). Scope Enumeration areas are retrieved from http://www.masdap.mw/layers/geonode%3Aeas_bnd. These are used, because we want to prioritize on a deeper level than Traditional Authority (admin-3) level, and there are no other official boundaries available. The dataset in principle data for the whole of Malawi, but contains 4 filters, which can be applied, which are the following: Filter_south: this filters out only the South of Malawi, for which the drough and flood analysis has been carried out (see details below). Filter_district: contains all EA's from the 3 pre-identified districts Zomba, Chikwawa and Nsanje. Filter_TA: contains all EA's from the 3 pre-identified TAs Mwambo, Makhwira and Ndamera. Filter_GVH: there are also 44 Group Village Heads pre-identified for the project. As these GVH's are points on a map, all EA's are selected here which have a GVH within their boundaries or very close to their boundaries. INFORM risk-framework The INFORM framework (http://www.inform-index.org/) is applied to assess risk per community, which is considered the main criteria for prioritization within the project. Because of low data availability we apply a crude version for now, with only some important indicators of the framework actually used. Since we feel that these indicators (see below) still constitute together a current good assessment of risk, and we want to stimulate the use and acceptance of the INFORM-framework, we choose to use it anyway. The INFORM risk-score consists of 3 main components: hazards, vulnerability and coping capacity. Hazard: For hazard we focus - in line with the ECHO2 project - on floods and droughts only. Analysis has been carried out (see more details below), to determine flood and drought risk on a scale from 0-10 with a resolution of 250meter grid cells. This has subsequently been aggregated to Enumeration Areas, by taking a population-weighted average. Thereby taking into account where people actually live within the Enumeration Areas. (Population data source: Worldpop: http://www.worldpop.org.uk/data/summary/?doi=10.5258/SOTON/WP00155) Vulnerability: Vulnerability is operationalized here through poverty incidence. Poverty rate (living below $1.25/day) is retrieved from Worldpop (http://www.worldpop.org.uk/data/summary/?doi=10.5258/SOTON/WP00157) and again transformed from a 1km resolution grid to Enumeration Areas through a population-weighted average. Lack of Coping capacity: Coping capacity is measured through traveltimes to various facilities, namely traveltime to nearest hospistal, traveltime to nearest trading centre and traveltime to nearest secondary school. Together these are all proxies of being near/far to facilities, and thereby an indicator of having higher/lower coping capacity. See https://510.global/developing-and-field-testing-a-remoteness-indicator-in-malawi/ for more information on how these traveltimes were calculated and validated. Use All features are stored in a CSV, but can easily be joined to the geographic shapefile to make maps on EACODE. Flood and Drought calculations Drought layer The drought risk map was created by analyzing rainfall data in the past 20 years using standard precipitation index (SPI) , which is a widely used index in drought analysis. Based on SPI6 values for the period October-march, which is the main rainy season in Malawi. Each pixel is classified to drought or no drought for each year based on SPI6 values, drought year if SPI value for a pixel is less than -1. Next, relative frequency is calculated, the number of times drought has occurred in the considered 20 year period. This frequency is then converted to probability of drought occurrence in a given year. We validated our analysis by comparing NDVI values for the drought year against long term average values. Flood layer To identify flood moments in Malawi Landsat imagery was studied (1984-2017). Floods were clearly evidenced in 9 dates. For the clearest and most representative layers the mNDWI (modified Normalized Water Index) was calculated. The index mNDWI (McFeeters 1996; Xu 2006) for Landsat bands is calculated as follows: (b2GREEN-b7MIRSWIR/b2GREEN+b7MIRSWIR). In this variation of the index the higher values are the wettest. A threshold was applied to the mNDWI to separate flood from non-flood or water from non-water pixels. The resulting layers were aggregated and the final stretched from 0-10, where 0 are the pixels where no flood is expected while pixels with 10 are where most frequent flood has been evidenced and therefore expected. The largest flood was observed in 2015, as the scenes were cloudy the flood extent was manually interpreted from several scenes. The evidenced flood dates are: 29 Feb. 1988 low flood, 19 march 1989, 17 march 1997, Feb 1998, March 1999 low flood, 2001 since February 16 until end of April, 2007 17 February since early Feb., 2008 Feb. medium flood, 2015 January – March. The water bodies in this layer are not represented and have a value of 0 like the rest of land where flood is absent.

Product This priority index was derived by combining a detailed flood extent mapping with detailed human settlement geo-data. Both sources were combined to produce the location and magnitude of population living in flooded areas. This was subsequently aggregated to admin-4 areas (GND) as well as admin-3 areas (DS divisional). The flood extent mapping was derived in turn by combining two sources: Flood extent maps could be produced rather faster using satellite imageries captured by either optical sensors or Synthetic Aperture Radar (SAR) sensors. In most places flood is cause by heavy rainfall which means in most cases cloud is present, this is a limitation for optical sensors as they can’t penetrate clouds. Radar sensors are not affected by cloud, which make them more useful in presence of cloud. In This analysis we analyzed sentinel2 optical image from May 28th and Sentinel 1 SAR image from May 30th. Then we combine the two results adding up the flood extents. Main cloud covered areas and permanent water bodies are removed from the flood extent map using the Sentinel 2 cloud mask. The scale/resolution of the flood extent map is 30mts where as the permanent water body map has 250m scale resolution. This will introduce some discrepancy: part of flood extent map could be permanent water body. Scope Analysis focused on 4 districts in South-West Sri Lanka based on news reports (https://www.dropbox.com/s/n0qdqe7qfgq6fyv/special_situation.pdf?dl=0). Based on the admin-3-level analysis, highest percentages of population living in flooded areas were seen in Matara district. Admin-4 level analysis concentrated only on Matara district for that reason. Caveats The dataset is showing percentage flooded. The data has not yet been corrected for small populations. We believe the product is currently pointing to the high priority areas. In the shp or csv files the user of this data could easily correct for small populations, if there is a wish to target on the amount of people affected. Data used from partners The human settlement data was retrieved from http://ciesin.columbia.edu/data/hrsl/. Facebook Connectivity Lab and Center for International Earth Science Information Network - CIESIN - Columbia University. 2016. High Resolution Settlement Layer (HRSL). Source imagery for HRSL © 2016 DigitalGlobe. Accessed 01-06-2017. The Radar imagery analysis was done by NASA JPL, whose input in this product has been crucial. Visualization An example map is available here: http://bit.ly/SriLankaFloodMap Linked data Admin boundaries 3 and 4 can be found here (link on OBJECT_ID): https://data.humdata.org/group/lka?q=&ext_page_size=25&sort=score+desc%2C+metadata_modified+desc&tags=administrative+boundaries#dataset-filter-start How to use The ratio column in the SHPs or CSVs can be multiplied by 100 to get the percentage of flooding in the area.

Dataset contains windspeeds in miles per hour.

Blog post about this prediction can be found here: http://bit.ly/2fWF2jq The predicted priority index of Typhoon Haima is produced by a machine learning algorithm that was trained on four past typhoons: Haiyan, Melor, Hagupit and Rammasun. It uses base line data for the whole country, combined with impact data of windspeeds and rains, and trained on counts by the Philippine government on people affected and houses damaged. First run The Priority Index is a 1-5 classification that can be used to identify the worst hit areas: those that need to be visited for further assessments or support first. Second run The model now predicts two things: a weighted index between partially damaged and completely damaged, where partially damaged is counted as 25% of the completely damaged. This has proven to give he highest accuracy. the precentage of total damage (damaged houses versus all houses) The absolute number of houses damaged / people affected is insufficiently validated at the moment, and should just be used for further trainng and ground-truthing. Data sources: Administrative boundaries (P_Codes) - Philippines Government; Published by GADM and UN OCHA (HDX) Census 2015 (population) - Philippine Statistics Authority; received from UN OCHA (HDX) Avg. wind speed (km/h) - University College London Typhoon path - University College London Houses damaged - NDRRMC Rainfall - GPM Poverty - Pantawid pamilyang pilipino program (aggregated) For the second run of the algorithm we also included: Roof and wall materials New geographical features The result of different models can be found in the file 'Typhoon Haima - performance of different models - second run.csv' A note on how to interpret this. date running date alg_date same alg_model name of the algorithm used alg_predict_on name of the learning variable alg_use_log i s the learning variable transformed in log code_version version of the learn.py code All the columns with feat_ indicates the importance of that feature, if not present that feature was not used. learn_matrix name of the learning matrix with the 5 typhoons run_name unique run name (pickle files and csv files have this name for this model) typhoon_to_predict name of a new typhoon to predict val_accuracy accuracy based on 10 categories of damage 0% 10% 20% … val_perc_down perc of underpredicted categories val_perc_up perc of overpredicted categories Val_best_score best r2 score Val_stdev_best_score error on best score based on the CV Val_score_test r2 score on the test set (this should be around +- 5% of the previus number to not overfit Val_mean_error_num_houses average error on the number of houses val_median_error_num_houses median val_std_error_num_houses std deviation of the errors (lower is better) Algorithm developed by 510.global the data innovation initiative of the Netherlands Red Cross.

NDRRMC report of October 28th 2016