How the AI observation system works

The AI observation page is designed to make scientific data collection as simple as possible while preserving the information required for a useful observation. The participant can either take a new photograph with the device camera or select an existing photograph from the phone or computer.

The selected image is previewed directly in the browser. Before the AI analysis can begin, the system also requires the observation to have a geographical position. This ensures that the biological information can later be connected to a real location on the Lichenary map.

The browser automatically creates the observation date and time using the participant's device clock. The geographical coordinates are obtained through the browser's Geolocation API after the participant gives permission.

These values are placed into hidden form fields and displayed to the participant through the interface. Before being accepted by the server, the coordinates are validated by the Python backend to make sure that latitude and longitude fall within their valid geographic ranges.

When the participant selects Analyze Lichen with AI, the browser creates a multipart request containing the photograph and the observation's geographical information.

The Flask application receives the photograph and reads its image data. The server verifies that an image was actually uploaded and that the MIME type represents an image before sending the photograph to the AI system.

Gemini is given a carefully defined instruction: its responsibility is to identify the lichen species visible in the photograph. It is explicitly instructed not to calculate the pollution indicator and not to invent species outside Lichenary's reference list.

Lichenary contains a structured database of lichen species and their associated pollution-indicator values. Each species in this dictionary has a fixed numerical value between 1 and 10.

Once Gemini proposes a species, the backend searches for that exact species in the Lichenary database. If a match is found, the corresponding indicator value is returned. If the species cannot be matched to the reference data, the observation is rejected rather than being assigned an arbitrary value.

The entire process is composed of several connected stages. Each stage has a specific responsibility and passes only the necessary information to the next one.

After the AI successfully identifies a species and the backend finds its corresponding indicator value, the result is returned to the participant's browser. The identified species and pollution indicator are displayed in the observation form.

The participant can then submit the completed observation. At this stage the photograph, date and time, location, coordinates, species and indicator value are sent to the final upload endpoint.

Before the observation is stored, the server performs another species lookup. This second validation is important because information displayed in a browser cannot be considered trustworthy simply because it was originally produced by the application.

Lichenary stores observations in a relational database. Each observation has its own record containing information such as the photograph URL, date and time, location, latitude, longitude, identified species, pollution indicator, approval status and the user who submitted it.

Photographs are uploaded to Cloudinary rather than being stored directly inside the database. The database stores the secure image URL associated with the observation.

Approved observations containing valid geographical data can then be requested through Lichenary's observations API. The map uses this structured information to display observations according to their geographic position and pollution-indicator value.

A photograph does not always contain enough information for reliable species-level identification. Lighting, focus, image quality, the angle of the photograph, overlapping organisms and environmental conditions can all affect visual identification.

For this reason, an AI identification should be understood as an automated suggestion rather than an infallible taxonomic determination. Lichenary's review process adds an additional layer of quality control.

The pollution indicator also has an important limitation. It represents the known relationship between the lichen species and pollution tolerance in the Lichenary reference system. It is not a direct sensor reading and should not be interpreted as an exact concentration of air pollutants.