Library Master| Scanning for Lifeforms
Scanning for Lifeforms
Biodiversity Exploration for Rare Species
Summary
Recently, machine learning methods are powerful to survey and analyze biological data. Even though these techniques save a large amount of time and labor costs for big data processing, they often perform worse when training data is limited.In recent years, some researchers study to solve this problem. The solution to this problem is called few-shot learning. In this work, we aim to effectively explore biodiversity for rare species.First, adopting a few-shot learning algorithm to train the models. Due to collecting rare species data is hard, applying conventional methods will cause accuracy degradation. Second, our models deploy to AIoT devices to only store valuable data on local storage
How We Addressed This Challenge
1.The importance of images for biodiversity
Images can record animals’ present population size or habitat preferences. They can help researchers to build up knowledge of the identity, geographic distribution, and evolution of living species. Hence, our work focuses on using images to trace biodiversity.
2.Communication problem and hardware cost
The Existing edge solution tracks biodiversity of rare species by using motion cameras to capture images and save in local storage. Until edge devices running out of battery or storage, researchers go to collect data. However, this solution causes pressure to storage and researchers require extra efforts to extract interesting rare species images. Hence, to solve these problems, we aim to develop novel rare species data collection processes on edge devices.
3. Limited supervised data for categories of rare species
In order to decrease pressurement of storage and researchers, we use classification models to help us. However, rare species data is always hard to collect. Using naive supervised learning methods causes poor classification accuracy. Hence, we introduce a few-shot learning method to solve this problem.
performance.
How We Developed This Project
In order to solve the problems of existing methods. We propose a novel biodiversity exploration method on edge devices. More specifically, our method analyzes input from a motion camera by neural network. This neural network uses a few-shot learning approach to build up. After analysis by our neural network, our system will discard non-rare data and store rare data. This mechanism can diminish pressurement of storage.
Moreover, our system has the ability to early notify researchers. If the input image belongs to rare species, our system will remind researchers to increase interactivity.
Power consumption is a critical issue on edge devices. In our work, we introduce NeuroPilot to help us. Our model is quantized and deployed by NeuroPilot.
How We Used Space Agency Data in This Project
We use iNaturalist dataset, which is from NASA resources link, to explore rare species biodiversity.
Project Demo
When we lack training data, Few-shot learning performs more efficiently and robustly than traditional classification. Moreover, the quantized model is faster and has less power cost when doing inference.
- The figure(day view) on the left shows the superior of model trained on Few-shot learning, the right one(night) shows the further impact compare to naïve classification model
- The table shows that our model compression work significantly improve model inference time and power consumption by model quantization
Data & Resources
What is remote sensing?: https://earthdata.nasa.gov/learn/remote-sensing#data-pathfinders
GIBS API for Developers: https://wiki.earthdata.nasa.gov/display/GIBS/GIBS+API+for+Developers#GIBSAPIforDevelopers-ImageryAPI/Services
Taxonomy: https://zh.wikipedia.org/wiki/%E7%94%9F%E7%89%A9%E5%88%86%E9%A1%9E%E6%B3%95
Deep Learning for Large Scale Biodiversity Monitoring: https://conservationmetrics.com/wp-content/uploads/Klein_2015_bloomberg_data4good-2015.pdf
A deep active learning system for species identification and counting in camera trap images: https://www.researchgate.net/publication/336735839_A_deep_active_learning_system_for_species_identification_and_counting_in_camera_trap_images
Scene‐specific convolutional neural networks for video‐based biodiversity detection: https://besjournals.onlinelibrary.wiley.com/doi/full/10.1111/2041-210X.13011
(From NASA resources link) iNaturalist, https://www.inaturalist.org/