Analysis - AI Plant detection

This guide explains how to use the AI Plant Detection tool in the CultiWise application to automatically detect crops and weeds using AI detection. 

The resulting detection layer is a required input for generating Selective Herbicide Prescription Maps or performing a Plant Count Analysis.

Starting the AI Plant Detection

To begin, you can start from either the Farm View or the Fields View.

Click 'CREATE NEW', then select 'PRESCRIPTION'.

Alternatively, you can go directly to the field where you want to run detection and click 'CREATE PRESCRIPTION'.

In both cases, after selecting Prescription, choose 'Analysis', and then select 'Plant Detection'.

AI detection

First, you’ll have the option to run fully automated AI detection.
This method uses a generic model trained on common crop and weed types. In many cases, this model works well without any extra setup. 

To test it on your field, click 'GO TO AI DETECTION'

Then draw a Test Area by clicking 'DRAW TEST AREA'.

Once your test area is drawn, the option to run detection will become available. Click 'APPLY AI DETECTION' to proceed.

Note: The Test Area must be within size limits — not too small or too large. You'll be notified during drawing if adjustments are needed. Ensure that all crop and weed types are represented within the test area. You can adjust the area size after it's drawn by clicking on the area and dragging the edges.

Review AI Detection Results

After approximately 10 seconds, you’ll see the detection results displayed on the selected area. Now it's time to review them. 

Here are the criteria on how to check if the detection is accurate:

• No false detections occurred

• All weeds were correctly identified

• Bounding boxes (labels) cover each plant properly

• No duplicate detections on the same plant
  

Available options moving forward

• If you're unsure with the results, test additional areas in different parts of the field — especially where conditions like plant size, type, or soil differ. 

Note: We recommend testing multiple spots before running detection for the entire field. You can do so by clicking 'TEST NEW AREA'.

• If you’re satisfied with the results, click 'DETECT WHOLE FIELD' to apply the model to the entire field. The tool will run AI detection across the whole area. Once complete, a new layer with the results will appear in your field view.
  
  

• If the results are not satisfying, you can start training your own model to better match your specific field conditions. Click 'TRAIN MODEL' to start the process. 

Custom model training

Now we’ll cover how to prepare and train a model that fits your specific field conditions.

1. Training Area

In this step, you will draw Training Areas — zones on the field where model training will take place.

You must define at least two training areas and they should meet these conditions:

• Choose areas with high weed or plant density and diverse field conditions.
• Each crop or weed type on the field should appear in at least two training areas. 
  
  

Training areas must meet both minimum and maximum size requirements. You can adjust the boundaries of a training area after creating it by clicking its edges and dragging them.

• If the area is still empty (contains no labels), you can freely resize it — up to the maximum allowed size.
• If the area already contains labels, resizing is limited to only increasing the size, again within the allowed maximum.

Labeling options will become available once at least two areas are defined.

2. Plant Classes and Labeling

Next, choose one of the following labeling methods:

• Click on 'AI ASSISTANT' (recommended) to use our AI Labeling Assistant. 

• Click on 'CREATE MANUALLY' to define all labels and plant classes yourself. 

AI Labeling Assistant

First, we will cover the recommended method of labeling, which is using the AI Assistant.

Our AI Labeling Assistant automatically pre-labels the plants in your Training areas using a generic model. 

After automatic detection, a guide at the bottom of the screen will assist you through each training area. For every area, you can:

• Discard results – if you're not satisfied delete all labels and prefer to label manually.

• Confirm results – review each area, click "Next" to proceed through each area, then "Confirm" when done.

Creating Labels manually. 

To begin, you must first create at least one plant class by clicking 'ADD PLANT CLASS' and naming it. You can continue adding more classes, or begin labeling immediately by clicking 'Draw Label' and selecting the class you want to use.

You’ll be guided through the Training areas in a similar way as with the AI Assistant. 

• When you’ve finished labeling an area, click Next to proceed.
• After completing all areas, click Finish Editing to submit your labels.

Labeling Recommendations and Requirements

No matter which labeling method you choose, each plant class must contain at least 10 labeled objects. If this requirement isn’t met, you won’t be able to proceed. To resolve it, either label more plants within the existing training areas or create additional ones by clicking 'ADD TRAINING AREA'.

You can monitor labeling progress directly in the class row next to the class name. This section shows the total number of labeled objects for each class, and if the number is below 10, it highlights that the minimum requirement for prediction hasn't been met.

At the top of each class row, you'll also see a colored progress bar. This bar indicates how well the class is prepared for model training:

• If the bar is red, the number of labeled plants is insufficient, and you won’t be able to run detection.
• If the bar is yellow, you will be able to run the prediction but the number of objects is not ideal.
• If the bar is green, the recommended amount of labels has been reached and you have the highest chance of accurate detection.

Editing tools

You can edit labels, whether generated by the AI or created manually, either during the guided review or later by clicking the 'Edit' icon next to the training area name, or by selecting 'Edit Labels' from the More Options menu.

After that, Edit tools become available. 

Click inside a label’s boundary to select it. You can select one or multiple labels.

• If a single label is selected, you can resize it for more accurate labeling.

• If multiple labels are selected, you can use this tool in combination with others (see below).

Inverse selection

Use this with Select labels to quickly select all currently unselected labels.

Draw labels

If a plant was missed, select the correct class and manually draw the label. These labels can be adjusted later just like any other.

Change class

Delete

Test Area

After setting up training areas and labels, you can test your model.

To begin, click 'DRAW TEST AREA' and define one or more test zones across the field.

These areas will be used to evaluate the detection results of your trained model.

To add additional test zones, click 'ADD TEST AREA'. When ready, click 'TRAIN AND TEST DETECTION' to run the model and view the results.

Note: We recommend creating multiple testing areas (up to 4) and including a variety of field conditions and labeled plant types to evaluate how the model performs under different circumstances.

This trains the model using your labels and runs predictions in the test areas. It may take about one minute.

Once the results are ready, review all results carefully to ensure they meet your expectations.

Results review

Now you will be guided through each Test area to closely review the detection results. You can switch back and forth between areas as needed. Once you’ve gone through all of them, options to proceed will become available.

What to do next?

Train the model further

If you're not satisfied with the results, continue refining your model using the same approach as before. 

You can:

• Improve labels by adding, adjusting, or resizing them. 
• Add more training areas either by creating new ones or converting existing Test areas into Training areas via More Options > 'Convert to Training Area'.

Detect Whole Field 

Once you're satisfied with the detection results, click 'DETECT WHOLE FIELD' to apply your trained model across the entire field. This may take a few moments. When complete, a new detection layer will appear in the CultiWise application.

You can then use this layer to generate a Selective Herbicide Prescription Map, or to run a Plant Count Analysis.

Note: After any adjustments, run 'TRAIN AND TEST DETECTION' again. Repeat this process until the results successfully detect all the weeds you want to target.

Labeling rules

1. Create or modify bounding boxes around the entire plant. Each box should fully enclose the plant.

2. Only one bounding box per plant. Avoid duplicates — each plant or weed should be labeled once.

3. Label every crop. No crops should be left unmarked.

4. In dense areas, be as precise as possible, even if plant edges are hard to see.

5. Watch out for double boxes. Sometimes the AI adds two overlapping boxes when unsure of the class — remove the extra one.

6. Delete all false detections. Any box that doesn't correspond to a real plant should be removed.

7. Don't overdo training. After a certain number of training areas, the improvements may be minimal. Minor mistakes can be filtered later when creating a spot-spray map.

8. Include variety. Try to label areas with different plant sizes, row directions, soil types, and weed conditions — both in training and testing.

9. Combine hard-to-distinguish weeds into one class. If two weed types look very similar, it’s okay to label them as one group.

10. Special scenarios – label only weeds in homogeneous crop cover. For cases like overgrown weeds in cereals (e.g., thistles) or grasslands (e.g., rumex), label only the weeds — not the crop.