PlantAI Playbook

Solvi PlantAI™ Playbook

Purpose of This Guide

Tour of the Plant AI Interface
Detection Types in Solvi Plant AI
Preparing the Field Boundary
Creating Effective Examples
Defining the Field Area
Reviewing the Results
Summary: Key Steps for Good Results

Purpose of This Guide

Before following this guide, it is essential to ensure that the drone imagery being used is of sufficient quality. Poor image resolution will undermine the effectiveness of even the most precise annotations and workflows.

As a general rule, the ground sample distance (GSD) of your drone imagery should be no greater than the typical plant size (in cm) divided by 30. For example, for a plant like corn that is approximately 20 cm wide, the recommended GSD would be: 20 cm / 30 = 0.66 cm/pixel

Using imagery that meets or exceeds this standard will significantly improve the accuracy and reliability of the detection results.

Solvi's Plant AI tool is built on a powerful base model that already recognizes a wide variety of common plant types. Therefore, when you create examples, your primary goal is not to teach the system what to include, but rather to guide it on what to avoid—especially problematic areas such as weeds, shadows, or unusual crop conditions. The examples serve as corrective signals to help the model refine its predictions and remove unwanted detections.

This document is a practical playbook for using Solvi's Plant AI tool effectively. The goal is to maximize the probability of getting accurate and useful results by following a structured, step-by-step methodology. This guide outlines the full workflow—from creating high-quality examples to reviewing outputs and refining the model.

1. Tour of the Plant AI Interface

1.1 Main Plant AI Interface and Annotation Tools

This screenshot shows the main Plant AI interface, highlighting the annotation toolbox:

Use the square, polygon, or circle tool to annotate plants.
The Magic Wand tool attempts automatic plant selection but may require manual correction.
The Get Latest Results tool appears after the first detection round, allowing you to load the most recent detection output.
This example also highlights good annotation practice:

Every plant within a bounding box is correctly annotated.
Detection boxes are placed in weedy areas and regions with different plant coloration.

The Show Background Map icon toggles between displaying just the detection polygons and the orthomosaic imagery.

1.2 Review Results Interface

This screenshot displays the Review Results interface after the first detection run:

Shows the detection probability histogram.
Displays previous and current detection iterations in different colors.
Contains tools on the right-hand side for adding and deleting visual notes.
Includes the Show Background Map toggle to switch between polygon view and the orthomosaic.
Note the Highlight Removed Plants feature under the histogram for identifying low-probability detections.

1.3 Adjusting Field Boundaries

This screenshot shows the boundary adjustment interface:

Add vertices by clicking on the boundary lines and dragging them to fine-tune the field area. Remove vertices by using “Alt” key, then click on vertex.
Observe the creation of negative areas to exclude unwanted parts of the field from processing.
Tight, clean boundaries that exclude irrelevant regions and improve processing efficiency.

2. Detection Types in Solvi Plant AI

Solvi Plant AI provides three main detection modes, each suited to different use cases:

Detection Type	Purpose	Annotation Tools
Basic Plant Detection	General plant presence identification	Square
Plant-Level Data	Detailed analysis of individual plant shapes & sizes	Circle, Polygon
Weed Detection	Identifying and distinguishing weeds from crops	Circle, Polygon

Aside from the shape tools provided, the overall workflow described in this guide applies equally across all three detection types.

3. Preparing the Field Boundary

Before creating any examples, and once image quality has been confirmed, the first action should be to set the field boundaries within Solvi.

You can create additional vertices to define complex boundaries. This allows you to cut out unnecessary regions and avoid including anything that may interfere with the plant detection algorithm. This is done from the Imagery tab.

While the Plant AI detection boundaries can be set or refined later within the PlantAI Tool, establishing the field boundary early ensures that all subsequent operations—such as example creation and detection—are correctly scoped to the field of interest.

The field boundary applies across the entire project workflow, helping to exclude irrelevant surroundings and maintain a clean focus on the target area.

4. Creating Effective Examples

Creating accurate examples is critical to the success of the AI's plant detection process. This section covers how to create good examples and what pitfalls to avoid.

4.1 Focus on the Bad Areas First

Spend some time scanning across the field before creating examples, looking for problematic areas such as:

Weeds along the edges
Bushes or shrubs that are not part of the intended target crop
Any zones containing plants or objects that you do not want to include in the detections

For the very first detection round, aim to create a minimum of three examples to provide enough data for the algorithm to generate initial results.
Prioritize creating examples in problematic zones (e.g. weed-heavy areas, overlapping crops, shadows).

4.2 Characteristics of a Good Example

It is absolutely critical that within a bounding box, all plants of interest are correctly annotated. Any plants that are left unannotated will train the model to avoid detecting them, which will lead to poor results from the algorithm.
When running a weed detection task, keep example areas small—even if weeds are large or spread out. Training on small, clearly defined sections helps the algorithm learn more effectively.
The Magic Wand tool can be helpful when selecting plant boundaries quickly, but be mindful that the results it generates may require manual refinement or complete removal, depending on accuracy. You will need to delete any detections created with the Magic Wand that look strange or do not fully encompass a plant.
Ensure the full boundary of each plant is completely included in the annotation.
Clear plant boundaries, with well-separated individuals.
As always, areas where weeds or other irrelevant vegetation should be excluded.
You can have annotations that overlap or extend beyond the boundary box edges

4.3 Geographical Placement

Distribute examples throughout the entire field, covering different plant densities, soil conditions, and light exposures.
If your imagery shows different lighting conditions or variations in plant coloration, ensure your examples cover these changes to help the model generalize across the entire field.
Pay special attention to:

Shadowed zones
Weedy patches
Boundary zones

5. Defining the Field Area

5.1 Setting the Map Boundary

By reducing the boundary area as much as possible, you also speed up processing time, making the detection process more efficient.
After creating examples, it is important to define the map boundary as tightly as possible to include only the field of interest.
Cropping out irrelevant peripheral areas (such as surrounding vegetation, equipment, or non-target regions) reduces noise and improves detection accuracy.
A well-defined boundary prevents the AI from evaluating regions that are outside the actual crop zone.

5.2 Using Negative Areas

Negative areas can be created using the scissors icon in the toolbar.
Solvi's new feature allows users to define "negative areas" within the boundary.
These zones are explicitly marked to be ignored by the AI during detection.
Negative areas can be used both within the field and along its edges, helping to exclude persistent problem zones (e.g. water tanks, fence lines, or weed patches).

6. Reviewing the Results

Once examples are created, it's essential to critically evaluate the output and refine the model as needed.

6.1 Visual Inspection

On your first detection round, focus primarily on scanning for missing plants—these are the important corrections to make early on.
After addressing missing detections, look for false positives—detections that do not correspond to actual plants.
Use the icon in the toolbar to toggle the orthographic map view on and off. This removes the background imagery so only the detection polygons are visible.
Alternatively, press the 'B' key on your keyboard to quickly toggle this view.
This is especially helpful for inspecting the geometry of polygons at a glance and identifying anomalies, such as detections occurring between crop rows or unusually shaped detections.
Use the ortho toggle to compare the AI overlay with the base imagery.
Zoom into areas with sparse or dense detections.
Look for both false positives (extra detections) and false negatives (missed plants).

6.2 Threshold Adjustment

Once the first results are generated, you will be presented with a histogram showing the detection probabilities across all the samples.
While this histogram can be used to filter the examples, we generally recommend using it as a tool to identify areas with low detection probabilities that may require further refinement rather than filtering the examples outright.
A new feature, "Highlight Removed Plants," allows you to filter detections by confidence level.
Plants below the current detection probability threshold are highlighted in purple.
This visual cue enables quick identification of problematic zones from a zoomed-out view.

6.3 Annotating Issues

The purpose of the "Add Note" feature is to mark areas on the map that require improvement.
This makes it easier to return to these specific spots in subsequent iterations and observe how the algorithm performs after adjustments.
Use the 'A' key on your keyboard to quickly activate the "Add Note" feature and place geographic markers.

6.4 Iteration & Comparison

After creating notes and adding additional examples, run a second, third, or fourth iteration. Once the iteration is complete, you can compare results on the same map: the previous iteration will be colored blue and the current iteration will be colored red. This visual comparison helps assess improvements in detection accuracy.
If successive iterations yield minimal change in the total number of detected plants, this can be a good indicator that the algorithm is performing well and is unlikely to benefit significantly from further iterations.

7. Summary: Key Steps for Good Results

Ensure drone imagery is of high quality, with an appropriate GSD.
Set the field boundary first to limit the working area.
Start with a minimum of three well-placed examples, focusing on problem areas.
Carefully annotate all plants of interest in your examples.
Keep weed detection examples small and clearly defined.
Use the Magic Wand carefully and delete any inaccurate detections.
Cover different lighting and plant coloration conditions in your examples.
Use the map boundary and negative areas to cut out irrelevant regions.
Review results by checking for missed plants and false positives.
Use threshold adjustment and the Highlight Removed Plants feature to spot weak areas.
Annotate problematic zones with notes to revisit in future iterations.
Use the Download Last Results tool to preload detections and fix issues efficiently.
Compare iterations visually using the color-coded results (blue vs red).
Run multiple iterations, improving examples each time, until there is a minimal change in number of plants detection between iterations, or that the final results are acceptable