We made Chirrup 5X faster than competitor bird recognition models
Chirrup wanted to create a bird sound recognition system that supports sustainable farming
Biodiversity loss is one of the thorniest problems farmers face. It affects soil fertility, crop yields, and ecosystem services. Chirrup has the solution to this problem and it lies in bird sounds.
Namely, birds play a critical role in maintaining the health of farmland ecosystems. However, identifying and monitoring bird species can be highly demanding and time-consuming for farming professionals. To address this problem, the client developed a cloud-hosted bird sound recognition system. Via this easy-to-use platform, users can upload and manage recordings, as well as generate reports.
Facing multiple challenges while bringing the platform to life
Time-consuming improvements of poor-quality audio files
The client’s primary goal was to create a highly intuitive application that enables fast and accurate bird recognition from audio clips. However, the process was more complex than they expected. Namely, they faced numerous difficulties, such as poor-quality audio files that required lots of preprocessing. They were wasting lots of time clearing and improving the audio data so that it could be further analyzed.
Collecting and validating data
To the precision of bird sound predictions, we focused on collecting as much data as possible. We did that by scraping relevant information from multiple sources. During the process, data validation was our top priority. We made sure all data was properly licensed and met ethical standards.
Ensuring platform accuracy
One of the client’s requirements was improving the accuracy of predictions. We did that by developing an occurrence mask for the output of the model. The purpose of the mask was to check whether some bird species could appear in certain locations during specific times of the year. That helped the client reduce false positives in platform predictions. While creating the occurrence mask was a highly time-consuming process, it delivered immense value to the client and boosted its accuracy.
The team: the data science and software development firepower
The platform development required a team highly skilled in data science and software development. We deployed a team that consisted of:
- Two data scientists, responsible for processing and analyzing audio data. Their task was also to implement the machine-learning algorithms used for bird sound recognition.
- A back-end developer, whose task was to build and maintain the server-side platform infrastructure. It included data storage, data aggregation, and app deployment.
- A front-end developer, who designed and implemented the user-facing web application. This included the UI/UX design.
- A product owner/project manager (PO/PM), whose task was to oversee the platform deployment and delivery. They defined and prioritized product features, managed the product backlog, and ensured the team met the client’s expectations.
While working on the project, we closely collaborated with ecology consultants. That way, we ensured the platform provided invaluable insights into the biodiversity and health of the land.
The results: a fast-growing platform, 5X faster than similar models
The model we developed exceeded the capabilities of its existing counterparts. It is 5 times faster than similar bird sound recognition solutions. Speed improvements let the Chirrup model perform tasks faster and more efficiently. That is exactly what gives it a major competitive advantage in this growing market.
Chirrup proved to be super-useful when it comes to detecting rare species and generating a large number of true positive predictions. Currently, 21 farms use the platform. Audio samples were collected from two sites on each farm. It is expected that the figures will double by the end of May once production starts.
How we achieved those results:
A deep learning network
We used a deep learning network to analyze the chirps of birds and identify the species in different areas.
Building a scalable and user-friendly system
We focused on boosting its accuracy while maintaining its user-friendliness and ease of use. The platform’s machine-learning algorithms were tested and optimized to ensure the highest level of accuracy.
The back-end infrastructure was designed with scalability in mind. The platform is built to handle large volumes of recordings and data. On the other hand, the front-end interface was designed to be highly intuitive and user-friendly. It made it easy for farmers and other users to upload recordings, manage their farms, and easily generate reports.
The full tech stack
- React, Material UI, and React-Route for the front end of the platform
- NestJS for the back end of the platform
- MongoDB as the database due to its scalability and ease of implementing data aggregation
- AWS infrastructure for deploying the back-end and artificial intelligence components
- CloudFront for the front-end deployment
- AWS services such as Amazon Simple Queue Service (SQS), Simple Email Service (SES), and Simple Storage Service (S3)
- SQS for the back-end and artificial intelligence components
- A custom-built Convolutional Neural Network (CNN) for the artificial intelligence components
- Classic libraries such as Librosa, Skimage, and OpenCV for spectrograms and extracting relevant features. Along with the powerful TensorFlow Keras API, they are also leveraged in training the CNN model using advanced techniques like transfer learning and data augmentation.