In the Spotlight

The history of agricultural science is built on the patient accumulation of field observations, trial data, and expert interpretation. An experienced agronomist who has spent two decades walking Indonesian potato fields carries a knowledge base that no dataset can fully encode — the subtle interplay of soil colour and compaction at depth that signals drainage problems, the particular pattern of yellowing that distinguishes magnesium deficiency from iron toxicity, the timing window when a crop that looks fine today will show stress symptoms by next week if rain doesn't come.

Data science, at its best, is not a replacement for this expertise. It is an amplifier — a system that can take patterns from thousands of fields and millions of data points, identify relationships that no individual expert could hold in mind simultaneously, and surface insights that sharpen the agronomist's recommendations rather than replacing their judgement.

The Data Foundation

Building useful agricultural data science for Indonesia requires several years of structured data collection before reliable predictive models become possible. At DayaTani, the data architecture is designed around this reality. Every field observation logged by an agronomist in the farm management application — every pest severity rating, every growth stage assessment, every soil observation — is structured data that accumulates in a format that machine learning models can use.

The sensor network contributes a continuous stream of weather and soil condition data, timestamped to specific plot locations. Satellite imagery, processed at weekly intervals for enrolled plot areas, provides vegetation indices — particularly NDVI and EVI — that track canopy health at a spatial resolution of 10 metres.

Farm outcomes — yield, quality grade, marketable fraction, input costs — complete the training data picture. The model needs to know not just what conditions were observed, but what happened to the crop when those conditions prevailed.

What Indonesian Conditions Demand

Agricultural data science models built on temperate-climate data — which represents the majority of globally available agricultural research literature — translate poorly to Indonesian conditions. The equatorial climate produces two or more growing seasons per year with no winter break in the pest and disease cycle. Soil types, particularly the andosols of highland Java and Sumatra derived from volcanic parent material, have distinct chemical and physical properties that affect nutrient availability in ways that models trained on European or North American soils will mis-specify.

This means that the value of Indonesia-specific training data is disproportionately high. Every additional season of field data collected in DayaTani's operational areas is not merely incremental — it is building a foundation that does not exist elsewhere and cannot be sourced from global agricultural databases.

Current Model Applications

DayaTani's current crop modelling work focuses on three practical applications where the data quality is sufficient to support reliable predictions:

Harvest yield forecasting: Using accumulated growing degree units, satellite NDVI trajectories, rainfall totals, and fertiliser application records, models trained on DayaTani's potato data can predict plot-level yield outcomes 3–4 weeks before harvest with a mean absolute error of approximately 12%. This advance warning allows logistics and marketing teams to plan procurement and cold storage allocation rather than reacting to harvest outcomes after the fact.

Disease risk scoring: Daily disease risk scores for key pathogens — late blight in potato, bacterial wilt in chilli, blast in rice — are generated for each monitored plot using logistic regression models incorporating weather, crop age, and local disease history. Scores above the advisory threshold trigger preventive alert messages to farm supervisors before visual symptoms appear.

Fertiliser response modelling: Analysing the relationship between soil nutrient baseline measurements, fertiliser application records, and yield outcomes across multiple seasons allows the nutrient recommendation engine to become progressively more calibrated to local soil conditions over time — moving from generic crop requirement tables toward plot-specific response curves.

The Human-AI Partnership

In practice, the most effective use of these models is not autonomous decision-making but agronomist augmentation. A field agronomist who can see a model-generated disease risk score alongside their own field observations makes better decisions than either the model or the agronomist operating alone. The model catches patterns across hundreds of plots simultaneously that no individual could monitor. The agronomist catches the local context — the specific plot's drainage problem, the unusual weather event from three days ago — that the model doesn't yet encode.

As the data accumulates and the models improve, the balance shifts. But the goal is not to replace the agronomist. It is to make each agronomist's expertise reach further — more plots, more farmers, more decisions supported by the same underlying expert knowledge.