Roadmap

In this version.

• Forecast totals over 4 and 8 weeksShipped

  Each forecast, the Overview and the Accuracy page show the total over the next 4 and the next 8 weeks with a range. Totals are more accurate than single weeks, and are checked against the harvest with their own trust level. How the forecast works

• Walk correctionsShipped

  On a walk's results or in the ledger, change a fruit's ripeness, mark a box as not a fruit, draw a missed fruit, or type the true counts, and undo any of it. The scored walk stays as it is. The week summary, forecast, backtest, accuracy, and CSV export use the corrected counts and show the model's next to them. Correct a walk

• Climate computer importsShipped

  Climate exports are read without a column mapping. Columns, units, and the layout are detected, with headers in English, Dutch, Spanish, or French. Files with several compartments and day-first or month-first dates are handled, and a preview shows what was found before anything is saved. Climate CSV import

• Self-service sign-upShipped

  A new organization requests access with a short form and keeps a one-time claim code. Once DeepLeaf approves the request, the grower collects the first admin key with that code, once. Get your key and sign in

• Offline mobile captureShipped

  Add the app to the phone's home screen, film or pick walks without a connection, and they upload on their own when the connection returns, once each. No signal in the greenhouse

• Overview of all greenhousesShipped

  Every greenhouse on one page after sign-in: this week's and next week's forecast with range, trust level, last walk and harvest, water productivity, alerts, and a weekly chart across the organization. Overview of all greenhouses

• The new appShipped

  A redesigned app at /app with Walk, Records, Forecast, Accuracy, and Setup, a week calendar picker, edit sheets, and delete confirmations that list what goes with it.

• Guided capture while filmingShipped

  Film in the app with a framing guide and a live verdict about once a second: move closer, step back, hold steady, or good. The check doesn't count as a walk. Filming guide

• Batch import of multi-clip walksShipped

  Up to 12 clips of one aisle, in order, scored and counted as one walk, with a capture check per clip.

• Per-house accuracy reportsShipped

  How far the forecast was off in each greenhouse, per horizon: error, bias, range coverage, models used, the calibration scale, and a plain trust level. How far to trust the forecast

• Key management for adminsShipped

  Admins create, name, and revoke their organization's keys on the Access keys page. Member keys do everything except keys and payment.

• Bank payment matchingShipped

  A bank wire or bank transfer that carries the invoice number is matched to its invoice automatically, and the plan takes effect.

• Plans and billingShipped

  A plan per organization, paid by card payment, bank transfer, or bank wire invoice, monthly or yearly, with walks counted against the monthly allowance.

• Walk scoring and capture checkShipped

  Fruit by ripeness, plants from stems, leaves, trusses, and flowers, from a photo or a video, with warnings when a walk was filmed too far, too close, or blurry.

• Measured fruit weightShipped

  Diameter and mass per fruit against a calibration board. Merged or near-camera boxes are left out of the average.

• Eight-week forecastShipped

  Ripening model that follows green fruit through turning to red at the expected temperature and learns each crop cycle's pace from its walks, with calibration and ranges, per-cycle choice against harvest history, and backtests. Weeks 4 to 8 add the open flowers on the walk, carry a widening range, and are marked rough until they have 8 checked weeks. How the forecast works

• Records, weather, and exportShipped

  Harvests, irrigation, and climate files; outdoor weather and a 16-day forecast per greenhouse; water productivity and a per-week CSV.

What we plan to work on next, based on what the measurements showed.

• Reduce close-range overcountPlanned

  Close 4K phone frames first gave about twice the labelled count (93% high). Dropping the small boxes that split one large tomato brought that to about 53% high, for about 3 points of fruit found, and the capture check flags footage filmed too close. Frames that stay under the too-close limit still count about twice the labels. The goal is still to count one tomato once, even up close.

• Validate the ripening forecast on a full seasonPlanned

  The new ripening model shipped after checks on public data, where it lowered the error for next week and the week after. Those data had fruit counts without ripeness, from one site and one season. Weekly walks and harvests of the same rows over a real season are still needed to measure the model itself. Accuracy and limits

• Learned forecast correctionPlanned

  A small learned model that corrects what the forecast tends to miss on a cycle. It is built in and off: per week ahead it turns on only after 12 checked weeks, and only where replaying those weeks in time order showed a lower error. Real seasons will show where it helps. How the forecast works

Ideas we're considering. None is committed yet.

• Fine-tuned model from correctionsnot decided

  Train a model version on the corrections growers make. Corrections are not used for training today. Optional, and pending a decision on how corrected data is handled.