Production release: insect.id & mushroom.id

It's been over a year since the beta release of mushroom.id and insect.id. During this year we have been testing our API's, collecting feedback, and improving, improving, and improving. Now it is time for the release of the production version, which has been on schedule for 17.09.2024.

With the production release, we have prepared exciting improvements, new features, and content:

1. Even more species
2. Accuracy better than ever
3. New content: danger of insects and more
4. Binary classification for insect/mushroom presence
5. Filter your results

1. Even more species

We have increased the number of species classes (or other taxa):

• twofold in insect.id, from 6,387 to 14,125 classes
• by about 50% in mushroom.id from 3,100 to 4,592 classes.

With insect.id, we have also decided to add slugs and snails (Gastropoda), as well as earthworms and leeches (Clitellata). Here is an overview of the classes in the new insect.id model.

2. Accuracy better than ever

The new models improved accuracy, particularly in identifying insects down to the species level, not just the genus. In tests on our internal validation dataset, which closely reflects the photos we typically receive, the new model outperformed previous versions. It achieved 92% accuracy for insect.id and 88% for mushroom.id for the top 3 results on our validation dataset.

To ensure that the accuracy is still high even though the number of classes increased, we have made the following changes to both APIs.

• Adding more photos to our training dataset.
• Weighting based on taxon popularity: we gave more weight to taxa that are more common and more likely to encounter.  
• Including GPS coordinates: allowing identification to be specific to the location where the insect/mushroom was seen. 
• Better coverage of photos throughout the year: to ensure that all stages of insect development are covered. 
• Focusing the model on regions of interest.

3. New content: danger of insects and more

Biting, transmitting disease, or causing allergies. These are some of the many ways that insects can harm people or animals. We have now added this information to the details in the insect.id API response. Species are tagged with the type of hazard they pose, and brief descriptions of how dangerous they are. We have also included the insect's role in the environment or to humans (e.g. pest, pollinator, or wood destroyer), its conservation status and risk of extinction from the IUCN Red List of Threatened Species.

See our documentation for more information and incorporate these new details into your solutions.

For mushroom.id, we have increased the content coverage of look-alike mushrooms.

4. Binary classification

One of the most requested features in the feedback we collected during the year of beta testing was a binary classification of whether or not there was an insect or fungus species in the image. We listened to you and added this functionality: there is a new part is_insect and is_mushroom in the identification result. 

5. Filter your results

A little jealous of Plant.id's suggestion filters, insect.id has earned its own taxonomy-based suggestion filters. This new feature allows you to filter insect.id results based on the taxonomic class of your choice (e.g. only butterflies and dragonflies). See the FAQ for details, and API implementation documentation. You can find all the filters here.

Let's look at an example of how it works. Here's an image of 2 insects - let's say I'm only interested in bee species, but not the butterfly on the right. 

Here are the top 5 suggestions from the full model:

1. Macroglossum stellatarum..... 81%
2. Macroglossum trochilus..... 5%
3. Megachile perihirta..... 2%
4. Hemaris thysbe..... 2%
5. Hemaris fuciformis..... 1%

The model mainly suggests butterfly and moth species, referring to the insect on the right of the image, which is larger and easier for the model to recognize. 

However, after applying a “hymenoptera” suggestion filter, the result changed:

1. Megachile perihirta..... 55%
2. Megachile latimanus..... 10%
3. Megachile ericetorum..... 4%
4. Xylocopa caffra..... 4%
5. Anthidium manicatum..... 3%

The correct bee species (Megachile perihirta) is now in first place with greater certainty.