The consultation gap in West Java smallholder farming
Commercial farms with 50 hectares or more have regular access to agronomist consultants. Farms affiliated with large off-taker companies often have mandatory scouting schedules. Smallholder farmers in West Java with 0.3 to 0.8 hectares typically have none of that. When they see something wrong on a leaf, they call whoever picks up the phone: the kios dealer, a neighbour who once took an extension course, or a relative with farming experience from a different soil zone.
This is not a criticism of those informal networks. They are what exists. But they have limitations. A kios dealer has a business interest in selling product. A well-meaning neighbour may have seen something similar but in a different crop variety or at a different growth stage. The farmer describing "brown patches on the leaves" over a phone call has almost no way to distinguish rice blast (Pyricularia oryzae) from brown spot (Bipolaris oryzicola) from bacterial leaf blight (Xanthomonas oryzae pv. oryzae) from simple manganese deficiency, all of which present with some variation of brown discolouration depending on the stage and severity.
The wrong diagnosis leads to the wrong treatment. A fungicide spray for what turns out to be a nutrient deficiency wastes money and delays the correct intervention. A nutrient application for what is actually an aggressive blast infection loses critical days.
What image-based detection can do at field conditions
Image-based crop disease detection has been an active area of computer vision research for several years, and a significant fraction of the published work uses standard benchmark datasets that were collected in controlled conditions: isolated leaves on white backgrounds, consistent lighting, high-resolution cameras. Field conditions in West Java are none of those things. Farmers take photos with low-end Android devices under variable light, with backgrounds of soil and other plants, often of a leaf that is partially obscured and in an awkward position to photograph.
What we have been building at Elevarm is calibrated explicitly for this failure mode. Our detection tool was trained on field images collected with the same device quality range that our farmers actually use. We prioritized getting the model to work reliably with imperfect inputs rather than achieving high accuracy on clean images. A model that reaches 92 percent accuracy on clean test images but drops to 61 percent accuracy on actual farmer submissions is not a useful field tool. We are more interested in consistent performance at 78 percent on realistic inputs than impressive performance on a benchmark that does not reflect deployment reality.
The four most common diseases in our calibration scope
We started with the four diseases that account for the large majority of reportable crop loss events in West Java rice cultivation based on BPTPH regional plant protection records and Dinas Pertanian district reports: blast, brown spot, bacterial leaf blight, and sheath blight (Rhizoctonia solani). These four represent a large fraction of identifiable disease-related yield loss events reported in the West Java rice belt.
The detection model is not meant to handle the full taxonomy of rice diseases. It is meant to handle the diseases that our farmers are most likely to encounter, with sufficient confidence to give a useful first-response recommendation. For cases where the confidence score falls below our threshold or where the image quality is too poor to classify, the output is "unclassifiable, recommend physical inspection" rather than a forced incorrect diagnosis. We built an explicit "I don't know" output because a wrong answer in this context causes real harm.
From detection to action
Identifying the disease is the first step. The second step, which matters more for farmer outcomes, is what they are supposed to do about it. We link each detection output to a recommended intervention that specifies the treatment category, the timing relative to growth stage, and the application rate range. We are specific about treatment category (fungicide, bactericide, or cultural intervention) but deliberately general about product brand, because product availability varies by kecamatan and we do not have a real-time view of what each kios dealer currently stocks.
We are also working on integration with the Mitra dealer network so that when a farmer reports a disease event and the detection tool outputs a recommendation, the nearby Mitra dealer gets a notification of potential demand for that treatment category. This closes a loop that currently has a large gap: the farmer knows what they need but does not know who has it, and the dealer does not know the demand is coming until the farmer physically shows up.
Honest limits of what this replaces
A mobile detection tool is a first-response triage aid, not an agronomist. For severe or unusual outbreaks, physical inspection by a trained person is still necessary and we say so explicitly in the tool output. The BPTPH (provincial plant protection board) and Dinas Pertanian district offices maintain field extension teams for this purpose, and our recommendation for cases that exceed the tool's confidence or severity threshold is always to contact those services, not to treat based on the app output alone.
We are building a tool for the 95 percent of cases that are straightforward enough to handle with good first-response information, and directing the remaining 5 percent to the experts who should handle them. The goal is to shift the burden of routine disease identification away from the informal phone call network and toward a more consistent, documented, actionable process. It will not be a complete solution until the formal extension services are better resourced. But it is a real improvement over the current situation.