Pola Hujan

The problem

Indonesians learn that the country has two seasons, rainy and dry. That's true for Java and wrong in much of the rest of the country. Ambon's wet season falls in Java's dry months, and parts of the equatorial belt get two wet peaks a year. BMKG, the national weather agency, knows this. It maps 699 season zones in three families: Monsunal, Ekuatorial and Lokal. But that knowledge lives in PDF bulletins and expert judgement, not in something a curious person can open and explore.

I built Pola Hujan for general Indonesian readers who want to know why "rainy season" means different things in different places. It describes long-term climate. It isn't a forecast or a planting calendar, and the site says so on its first screen.

What I built

I computed the map from raw data

The easy route was to digitise BMKG's zone map and colour it in. I computed my own classification from satellite rainfall instead. Each place's twelve-month rainfall cycle is treated as a wave and split into a yearly part and a twice-a-year part, the same maths that describes tides. Which part is stronger, and when it peaks, decides the family. Anyone can check the result.

Disagreements with BMKG stay on the map

The thresholds live in one file, each cited to Aldrian and Susanto (2003), the paper behind the three-family system. I set a rule never to adjust them until the map matches BMKG. Four places where my result differs, Medan, Palu, Tobelo and Labuha, stay on the map as reported findings, not quietly resolved. A separate layer marks every place where the two maps disagree, with sources, and the agreement rate has moved as I've added more cited comparisons, never from touching a threshold.

I tested the classifier on made-up data first

A generator builds rainfall cycles from known wave settings, and the classifier has to recover those exact settings. One permanent test covers a cycle that peaks six months after Java's, which must come out as the inverted family. Only after those tests passed did I run it on satellite data.

A static site with no heavy libraries

The pipeline downloads CHIRPS satellite archives and reads the raster files with decoders I wrote by hand. There is no mapping, charting, statistics or machine-learning library. Everything runs once at build time, so the site is static, makes no network calls and costs nothing to host. Download cost is dominated by the fixed set of monthly rasters, not by how many places get sampled from them. Coverage grew from an initial 15 cities to 34, then to 86, at almost no extra cost each time.

I turned down the obvious build

The easy pitch was an animated wind map: a gridded field, a month dial, particles drifting across it. I would have built a third near-identical animated-field project. I chose a categorical atlas instead, small multiples and time series rather than a continuous vector field, because the map's job here is to show which family a place belongs to, not to look impressive in motion.

Result

Pola Hujan is live, in Indonesian. A story-driven home page introduces the three families, then the atlas, individual city pages, a two-place comparison, and a merged method-and-explainer page carry the detail. It covers 86 locations. I've compared 52 of them against BMKG's published families, and my map agrees on 71%. Of those 52 comparisons, 32 are independently verified against the agency's own bulletin, up from zero; the other 20 are marked as estimates. I built the first version in three days in mid-August 2026, then kept returning over the following weeks to redesign the palette and type system, merge the method pages into one, and more than double the city count.