The sensor-free precision problem
Precision fertilizer application in commercial agriculture typically implies in-situ soil sensors, variable-rate application equipment, and GPS-guided spreaders. None of that is available to the median smallholder rice farmer in West Java working a plot of 0.3 to 0.8 hectares. The closest thing most of them have to a soil instrument is memory of how last season went and advice from whoever sold them the fertilizer.
The challenge we set ourselves was this: can you meaningfully improve fertilizer dosing accuracy without requiring any of that hardware? The answer we have found, with caveats, is yes. But the path there is different from what precision agriculture literature describes, because the information sources are different.
Starting with soil-class mapping, not sensors
West Java soil classification data exists. BBSDLP (the government land resources research body) has published soil mapping data covering the major rice-growing regencies, and many cooperative records include at least a basic soil type descriptor from historical land registration paperwork. The resolution is lower than a field sensor, but the data is real, it is collected over decades, and it covers the key agronomically relevant distinctions: alluvial lowland, volcanic upland, and the various andosol and latosol subtypes across Priangan.
When we map a farmer's registered plot to its soil class, we get an approximate cation exchange capacity (KTK) range, organic matter fraction estimate, and pH tendency for that area. These three parameters drive most of the relevant adjustments to the standard urea and NPK recommendations. A plot on high-organic-matter andosol in Garut behaves very differently from an alluvial plot in Karawang, even if both farmers are growing the same Ciherang variety at the same yield target. Generic national dosing tables treat them identically. A soil-class-aware model does not.
How rainfall data adjusts the dosing schedule
Timing matters almost as much as quantity. Urea applied before a heavy rainfall event leaches nitrogen before the plant can take it up. A 200 kg/ha urea application split into three top-dressings timed around the rainfall pattern is more efficient than the same quantity applied as two top-dressings on calendar dates.
We use the BMKG regional rainfall index, which is publicly available at the kecamatan level, to adjust the recommended split timing. When the 10-day rainfall forecast shows above-average precipitation for the region, the recommendation engine pushes the next urea application window by 5 to 7 days to avoid the leaching window. This is not a new agronomic principle. Extension workers in West Java have been teaching this for years. The issue is that individual farmers rarely have a quantified rainfall forecast in front of them when they are making the day-of decision. The model makes the adjustment automatically and surfaces it as a specific date range rather than general advice to "wait for drier weather."
What we cannot do without direct measurement
We are not claiming to replace soil testing. If a farmer has access to a cooperative that runs periodic soil tests, that data should override the class-based estimates in our model. The model is a good enough proxy when no test data exists. It is not the best possible information when test data does exist.
The other honest limitation is micro-variation within a plot. A farmer with a 0.5-hectare plot may have a soggy corner that receives subsurface water from an upstream plot. That corner will leach nitrogen faster regardless of what the plot-level soil class says. Our recommendation is for the plot average. A farmer who knows their field well knows which parts of it to manage differently, and that local knowledge is something a plot-level model cannot fully replace.
The waste reduction framing
In West Java, the national fertilizer subsidy system provides urea and phonska at subsidized prices up to a quota limit per farmer. The subsidy is real and meaningful. But the consequence is that fertilizer over-application has historically been subsidized too, which masks the true economic cost.
As subsidy quotas tighten and the cost-sharing structure changes under recent Kementan policy revisions, over-application is starting to show up more directly in farmer input costs. Our dosing recommendations generally come in 15 to 25 percent below the village-average application rate for the same yield target, based on the soil-class and rainfall adjustments. We model this as a cost reduction per hectare rather than a "precision agriculture" benefit, because that is the framing that connects directly to what farmers are actually trying to manage: their net income from a fixed area of land.
Building the feedback loop
The recommendation engine improves when farmers report their actual yields and input quantities at season close. We have been asking enrolled farmers to enter their harvest weight and the quantities they actually applied, which is a different number from what we recommended. The gap between recommended and actual tells us something about where the model is generating recommendations that farmers feel they cannot follow, either because the input is unavailable locally or because the timing does not fit their labour schedule. That feedback loop is what will, over several seasons, make the model calibrated to West Java field realities rather than calibrated to what the textbook says should work.