Species Recognition Models

The accuracy of a species recognition model largely depends on the quality of the dataset it was trained on. In computer vision, there’s a well-known saying: “garbage in equals garbage out.” This means that a low-quality dataset will result in a low-quality model. Therefore, investing time and effort into creating a high-quality dataset is essential for achieving the best results.

Image variability
Species recognition models are trained on labelled images from the project area in which they will be deployed. By repeatedly analysing these examples, the model learns to identify species effectively. A good rule of thumb is to provide at least 10,000 images per class. While more images per class are always beneficial, the returns diminish as the quantity increases.

What matters most is the variability of the images. For example, 5,000 images from 100 locations are far more valuable than 10,000 images from a single location. To develop a robust model, the dataset should be as diverse and heterogeneous as possible, incorporating different locations, backgrounds, camera types, angles, weather conditions, habitats, etc.

However, creating a perfect dataset may not always be feasible in real-world scenarios. For classes with limited data, don’t worry—Addax will work with whatever you have, even if a class contains only one or a few images. When necessary, the dataset might be supplemented with images from other ecological studies. While these external images are less effective than project-specific data, they still contribute to improving the model’s accuracy.

Location IDs
When preparing a dataset for training, it is crucial to know which images come from the same location. However, we do not need the actual latitude and longitude—just a unique identifier such as "Location 1," "Mossy ridge," or "Fernwood grove." The specific place doesn’t matter; what matters is grouping images that were taken in the same area.

Why is this important? Images from the same location are statistically dependent—they often share similar backgrounds, lighting conditions, and environmental factors. If these images are randomly spread across training, validation, and test sets, the model may learn shortcuts rather than generalizing to new locations. By using location-based splitting, we ensure that images from the same place do not appear in multiple subsets, leading to a more robust and reliable model.

If providing location IDs is difficult, we can estimate dependence based on image path similarity (e.g., folder structure or file naming patterns). However, this should be considered Plan B, as it is less reliable than using actual location IDs. Whenever possible, providing location identifiers will lead to better dataset integrity and stronger model performance.

Image tagging
As the expert on your region’s wildlife, you are best suited to annotate the data. Addax essentially requires information about which animal appears in each image or video, and preferably also its location ID. This information can be provided in various ways. In many cases, organisations have been labelling their images for years and already have a preferred method. That’s great! If you still need to label or organise your data, below are some points to consider. Applications like ZIP-classifier are excellent for tagging images efficiently. They allow you to organise and label large datasets effectively and prepare them for model development. Below are two common methods of labelling your data.

• Tagging with spreadsheet metadata - The most efficient and preferred way to provide image annotations is through a metadata file (e.g., XLSX, CSV, TXT, or JSON) containing the file paths along with their species labels and, if possible, location IDs. This method allows Addax to handle dataset organisation programmatically, ensuring consistency and flexibility.

• Tagging by folder structure - If you prefer to organise your data by folders, the images can be grouped according to species. Each species should have its own dedicated subfolder. Below is an example of how to structure the folders. Group the data starting from larger groupings (e.g., "Project A", "Project B") to smaller groupings (e.g., "Site 1", "Site 2"). This hierarchy helps ensure that images taken from the same location or deployment are grouped together, making it easier to track the statistical dependence of the images. It is important to maintain a consistent folder structure. The better the folder structure, the more smoothly the dataset can be processed programmatically, ensuring accurate model training. This approach allows the model to better understand and handle the relationships between images based on their metadata, and ensures smooth and efficient processing.

A metadata file is strongly preferred. However, if a metadata file cannot be provided, folder structure can be used instead— but please note that inconsistencies in folder naming or hierarchy may make it difficult or unreliable to extract location IDs automatically.

Don't worry if all this seems like a lot. We can work together to prepare the dataset, and even if the dataset isn't perfectly structured, we can still develop a model from it. However, keep in mind that the model's accuracy will be better when the dataset follows the structure outlined in the two methods above. With a bit of effort on organizing the data, we can ensure the model performs at its best.

Image variability
Choosing the right granularity for your labels is crucial. Generally, the more specific your labels, the more fine-grained your model's predictions will be. However, we understand that labeling every animal down to the species level may not always be feasible (or necessary for your project). Keep in mind that the model can only predict based on the labels you provide. For example, if you label all birds simply as "bird," the model won't be able to distinguish between different bird species.

The level of specificity you choose depends on your expectations for the model, the time available for labeling, and the labels already present in your project database. You can annotate using various taxonomic levels (e.g., "bird" (order), "rodent" (family), "hyena" (genus), "lion" (species), etc.), even if you only have a few images available. Let Addax handle the class organisation—just provide the most specific labels you can.

Multi-species occurrences
Images containing multiple species can present challenges during data processing. Addax’s annotation method will label all animals in an image with the tag provided by the client. For instance, the image below contains both ostriches and gemsboks in the background. If this image was labelled “ostrich”, all individual animals, including the gemsboks, will receive the “ostrich” label.

In most cases, multi-species occurrences are rare and not a major concern. However, they can create issues in specific contexts, such as studies conducted in pastures with domesticated animals or cameras monitoring watering holes. A few misclassifications are manageable and won’t significantly affect the model’s performance. However, frequent errors will impact accuracy. If you are aware of such images in your dataset, please provide a list of the affected images, deployments, or locations to help mitigate potential issues.

Please note that the resulting model will be capable of detecting multiple species in a single image. Only during the training phase, multi-species occurrences can pose a challenge and are best excluded from the dataset or kept separate.

Duplicate folders
Duplicates in the dataset can negatively impact model accuracy. If duplicate folders or images are present, they might inadvertently end up in different data splits (train, validation, or test). This increases the risk of the model becoming overconfident, which reduces its ability to generalise effectively. While a few duplicates won’t pose a significant problem, many of them can hinder the model’s learning process. Ensuring the dataset is free of duplicates is an important step to maintain its quality and improve the overall performance of the model.

Multiple camera setups
Some projects involve multiple camera setups with different angles or backgrounds, such as downward-facing cameras, enclosures, or arboreal camera traps. While this provides valuable diversity in data, an AI model may struggle to generalise when multiple setups are included in the same dataset. For example, the two images below show the same species, the striped skunk (Mephitis mephitis), but taken with different camera setups. As you can imagine, the model might have difficulty recognising the same species under these different conditions.

.

If you are working with multiple camera setups, it is best not to combine them into a single dataset. For the most accurate results, develop separate models for each camera setup. This allows the model to specialise in the particular conditions of each setup, improving its accuracy and reducing confusion during predictions.

More questions?
If you have further questions or need additional help with labelling, feel free to reach out.