Data Exploration and Conditioning

We developed a monitoring protocol that generates monthly estimates of precipitation for 23 inhabited atolls and islands using Integrated Multi-satellitE Retrievals for Global Precipitation Measurement (IMERG) version 6 satellite data (National Aeronautics and Space Administration, 2023). IMERG uses pixels that are larger than atoll land masses (0.1 by 0.1 degree, or roughly 10 by 10 kilometers) and thereby cover land and water. One or more pixels were selected for each location using a global shoreline dataset adapted from Sayre and others (2019). Any point within a pixel is assumed to have received the same amount of precipitation, but IMERG provides a more-seamless spatial coverage than a rain gage. Both datasets are important for developing the monitoring protocol because rain gages provide measured data that serve as a ground truth against which the satellite estimates can be compared.

The National Aeronautics and Space Administration provides IMERG precipitation in three different runs. The IMERG Early Run (that is, the signal processing steps that create IMERG Early Run data) is near real time with low latency (hours) that is available first. The IMERG Late Run (that is, Late Run data) is near real time (1-day lag) with quasi-Lagrangian time interpolation that uses all satellite microwave precipitation estimates, together with microwave-calibrated infrared satellite estimates, and rain gage analyses improving the estimate. The IMERG Final Run (that is, Final Run data) is considered the final (1- to 6-month lag) product and has monthly gage data added to create estimates.

We used 20 years (2001–20) of IMERG version 6 Final Run data (Senay and others 2022) to define baseline statistics; in other words, central tendency, and dispersion of the satellite data for each of the 23 locations. This 20-year period included a roughly equal number of El Niño and La Niña events (Golden Gate Weather Services, 2023; Physical Sciences Laboratory, 2023; Webb, 2023), which means the baseline statistics were representative of the range of ENSO effects on precipitation patterns. IMERG Late Run data are available on a daily basis and thus were selected because of their near real-time (1-day lag) availability to provide timely precipitation reporting.

To demonstrate representativeness of the IMERG data, we compared the IMERG Final Run data for the 20-year period with the measured rain gage data from 7 sites (Ailinglaplap, Jaluit, Kwajalein, Majuro, Mili, Utirik, and Wotje; fig. 1). Before the comparisons, the datasets were explored and quality controlled. The IMERG Final Run data were complete for all months and locations during the 20-year period. In contrast, the rain gage data were not complete and had periods of missing data in the record. Months with partial data were excluded if less than 28 days of data were in any given month. If data were missing for entire months, we manually entered precipitation reported in Pacific ENSO Applications Climate Center (PEAC) update reports if available (see all reports at https://www.weather.gov/peac/update). The final database was relatively complete, especially for the years from 2012 to 2020. The 2001–11 period was more variable and had notable data gaps for 2004 and for the Jaluit site.

We used the Pearson’s correlation coefficient (r) as the basis of comparison (Welkowitz and others, 2011). For each location, daily IMERG Final Run data were plotted against the rain gage data. A few bivariate outliers were identified (Welkowitz and others, 2011; PennState Eberly College of Science, 2023); thus, we filtered out two standard deviation outliers from the daily records before summing the daily data into monthly totals. The pattern of outliers was not in particular months of the year or years in the dataset, and no month had less than 28 days per month after filtering outliers. We computed monthly sums for each site. The IMERG Final Run monthly sums for each location were positively skewed (ranging from 0.57 to 2.01). Separate regressions of IMERG monthly estimates on the rain gage data were performed for each location; intercepts were forced through zero assuming and observing that generally zero rainfall conditions from a rain gage would correspond to IMERG estimates of zero rainfall. For all seven sites, Pearson’s r values were high, ranging from a low of 0.81 for Wotje Atoll to a high of 0.95 for the rain gage at the Majuro Atoll airport.

In a second evaluation step, we compared the 2001–20 dataset of IMERG Final Run data (used to establish the baseline statistics) with the IMERG Late Run dataset (Senay and others, 2022). Regressions were performed on monthly sums for each of the 23 sites. Pearson’s r values were uniformly high, ranging from a low of 0.97 for Wotho Atoll to a high of 0.99 for the Mili Atoll. The results of these two evaluations demonstrated that using IMERG Late Run data for timely monthly reporting is reasonable because it is well correlated with IMERG Final Run data and could be used as reliable indices with confidence.

Developing the Report Content

The content of the monthly reports consists of descriptive statistics assembled from the IMERG data plus visual indices of those statistics. The report that came out each month was a summary of the estimated precipitation of the previous month. Because the monthly reports were designed to supplement existing precipitation, drought, and agroforest management products, the reports do not include predictive statistics or interpretation of potential risk to or effect of the estimated precipitation on vegetation health or crop production.

The team discussed which approach was more important for the monthly reports; monitoring typical annual trends (that is, usual weather patterns) or detecting unusually dry or wet conditions. To monitor typical annual trends, the data from all locations and months in the 20-year database would be pooled to compute descriptive statistics that would be used to assess how typical any given month’s precipitation was. Instead, the team decided detecting unusually dry or wet conditions was more important; therefore, any given month’s precipitation should be compared against the long-term (2001–20) data from that location and month. Thus, a set of descriptive statistics was calculated for each month at each of the 23 locations using the 20-year database. This means that dry months and (or) dry locations may receive little precipitation but that precipitation might still be considered typical based on the range of precipitation in the 20-year record. Importantly, low precipitation may still affect freshwater availability or vegetation health even if it is typical. That said, connecting satellite-estimated precipitation with risk exposure of specific crops in specific locations would require thorough understanding of the substrate, groundwater, crop agronomy, and distribution of vegetation types at each of the locations; this knowledge does not currently exist in systematic form for all 23 sites.

For the report, the team agreed to have two types of rainfall indicators that would serve as a frame of reference for how typical or unusual the monthly precipitation was:

Importantly, the two indices may not always agree. The rainfall index indicates how typical the month’s precipitation was compared to the 20-year record for that location and month. For the rainfall index, low levels of precipitation at a typically dry location and (or) typically dry month may be at or near median levels and thus be categorized as typical. In contrast, precipitation less than 8 or 4 in. would be categorized as “concerning” using the PEAC criteria, regardless of whether those precipitation amounts would be typical or unusual at that location or month. Therefore, it is likely that for some locations, months, or both, the rainfall index may indicate typical conditions whereas the PEAC index may indicate low levels of precipitation.

Developing the Layout for the Report Template

At the core of the monthly report are two estimated precipitation values, two colorized rainfall indicators, and a bar graph with recent months’ precipitation history. This information is delivered in multiple ways that provide different quick-look summaries of the precipitation status. Internet connectivity in the Marshall Islands at times is unstable or limited by bandwidth (Okuda and others, 2018); thus, we sought to maximize ability to communicate by minimizing file size of the report by developing relatively dense graphics. The layout evolved during the 1-year co-production partnership through cycles of production, evaluation, and refinement. Here, we describe the content and layout of the monthly report that is available (Senay and others, 2023; https://doi.org/10.5066/P90J1426). Information graphics in the report are numbered to facilitate the co-production process and are referenced here: