Random-forest crop prediction from the per-field statistical time series.
Models answers a different question from Checks. Checks asks whether a field behaves like its declared crop; Models asks what crop the field most looks like, given every crop the classifier knows.
A random forest is only as good as the training set. Most of the work in this app is in Training Samples — building a set that is reliable, balanced across crops, and drawn from parcels whose declaration can be trusted. A model trained on a set inheriting the declaration errors it is meant to detect will agree with them.
Overall accuracy hides the crops that matter. A classifier can score well overall while being useless on a minority crop that happens to be the one under control. The Accuracy panel reports per-crop performance for that reason.
Per-field predictions with per-crop accuracy figures, written as CSV into the selected period's results/ folder, from where Results validates them and Database packages them for the client.
This app opens on a year · project · period selector in the top bar. A run reads only the statistical tables inside the selected period, and writes only into that period's results/ folder. Training on statistics silently drawn from two periods produces a model whose accuracy figures cannot be reproduced.
Use the menu on the left; each panel has its own help topic.
Status: not written yet. This topic is a stub created with the project template. The panel's owner (see
.github/CODEOWNERS) fills it in as the panel is built. Delete this quote block when you do.
TODO — one or two sentences: what question does this panel answer, and at what point in the CAP-control workflow does an operator open it?
The manifest describes it as: Build reliable training samples from the selected period's statistics.
TODO — what must already be true. This app needs a year, project and sampling period selected in the top bar.
TODO — name the outputs exactly: which table, which CSV, which folder under the project tree. An operator reading this later needs to find the artefact without reading the code.
TODO — the failure you actually hit while building this, and the check that distinguishes it from the next-most-likely cause. This section is the reason the help system exists; a list of buttons is not.
Status: not written yet. This topic is a stub created with the project template. The panel's owner (see
.github/CODEOWNERS) fills it in as the panel is built. Delete this quote block when you do.
TODO — one or two sentences: what question does this panel answer, and at what point in the CAP-control workflow does an operator open it?
The manifest describes it as: Train random-forest models and predict the crop on every field.
TODO — what must already be true. This app needs a year, project and sampling period selected in the top bar.
TODO — name the outputs exactly: which table, which CSV, which folder under the project tree. An operator reading this later needs to find the artefact without reading the code.
TODO — the failure you actually hit while building this, and the check that distinguishes it from the next-most-likely cause. This section is the reason the help system exists; a list of buttons is not.
Status: not written yet. This topic is a stub created with the project template. The panel's owner (see
.github/CODEOWNERS) fills it in as the panel is built. Delete this quote block when you do.
TODO — one or two sentences: what question does this panel answer, and at what point in the CAP-control workflow does an operator open it?
The manifest describes it as: Post-process predictions; assess per-crop accuracy; recommend improvements.
TODO — what must already be true. This app needs a year, project and sampling period selected in the top bar.
TODO — name the outputs exactly: which table, which CSV, which folder under the project tree. An operator reading this later needs to find the artefact without reading the code.
TODO — the failure you actually hit while building this, and the check that distinguishes it from the next-most-likely cause. This section is the reason the help system exists; a list of buttons is not.
Status: not written yet. This topic is a stub created with the project template. The panel's owner (see
.github/CODEOWNERS) fills it in as the panel is built. Delete this quote block when you do.
TODO — one or two sentences: what question does this panel answer, and at what point in the CAP-control workflow does an operator open it?
The manifest describes it as: Export period results as CSV into the period's results folder.
TODO — what must already be true. This app needs a year, project and sampling period selected in the top bar.
TODO — name the outputs exactly: which table, which CSV, which folder under the project tree. An operator reading this later needs to find the artefact without reading the code.
TODO — the failure you actually hit while building this, and the check that distinguishes it from the next-most-likely cause. This section is the reason the help system exists; a list of buttons is not.