By Mushahid Nazir Bhat, PhD scholar, Department of Computer Science and Engineering, IIT Delhi
A decision made once, lived with for thirty years
I am from Kashmir, and I have always known apple as more than a crop. An orchard is planted once and stays for around thirty years. So the first choice a grower makes, which variety and which rootstock to put on a given piece of land, shapes everything that follows.
That choice is usually made at a nursery counter, on the word of the person selling the trees. The grower cannot tell one rootstock from another by looking at it, and cannot check what the graft really is. Whether the advice was good only becomes clear four or five years later, when the trees begin to bear, and by then it is too late to change anything.
This is not because growers know less. The fact that matters, how these trees will do on this particular plot, sits with the seller and cannot be checked by the buyer. I wanted to see if data could help fix that.
What I tried first, and why it did not work
My first idea was simple: take soil, climate and terrain for a plot, and learn from existing data which cultivar and rootstock do best there. Before building anything, I checked whether the data could support this at all. It could not.
- The only Kashmir dataset that links zone, cultivar, rootstock and yield has every rootstock planted at one spacing only. A model could not tell whether a result came from the tree or from the spacing.
- Nothing was measured at the plots themselves: no soil pH, no slope, no aspect, no chilling hours.
- The paper prints only averages, so just fourteen usable numbers are left.
Anything trained on this would mostly learn spacing and call it rootstock. So I stopped and changed the plan.
What I built instead
Rather than train on data that does not exist, I decided to pull the facts out of the research papers and guides we already have, and keep a clear record of where each fact came from.
- A land score. It uses seven factors, height, temperature, slope, land cover, the direction the land faces, rainfall and soil texture, with weights taken from a published study. All the map layers are free and open, at about 230 m resolution.
- Winter chill. Apple needs roughly 1,000 to 1,500 hours between 0 and 7.2 °C to break dormancy. No Kashmir study reports this, so I worked it out from NASA daily temperature data.
- Hard limits. Below about 1,000 chilling hours, or soil pH below 5.5, no good score elsewhere can save the site. These are shown separately, never averaged away.
- A plant list. Four JSON files covering 22 varieties, 13 rootstocks, 18 diseases and 33 official rules. Every value carries the document, the page and the sentence it came from.
A language model on my own machine writes a first draft of each entry, but it is only allowed to copy a fact, never to guess one. Then I read the original sentence on the original page myself, and the program refuses to load any field that has no source behind it.
What it looks like on a real plot
I tried it on a plot in Büirhālan, in Shupiyan district. The land scored 2.45 out of 4, which is suitable, and both hard limits passed.
Then the tool looked at the plants. Gala came back as plantable, with the condition that another variety must flower nearby for pollination. M-9 was refused, and the reason is worth noticing: the land is fine, the plant is the problem. It is recorded as prone to white root rot, woolly apple aphid and fire blight. The grower is told to change the tree, not the land.
What comes next
This is an early tool and not a trained model, and I want to be honest about that. The next steps are to take it to nurseries and growers, see where it agrees and disagrees with their experience, and add a way for growers to report whether a recommendation worked. Over seasons that feedback can become the dataset that is missing today.
I also want to bring this into VARTA-Seeds as an open, Digital Public Good, instead of making one more app that growers have to find and trust. Himachal Pradesh and Uttarakhand are the next regions on the list.
In the coming posts I will write about the method, the JSON files, and how I worked out winter chill.