🧠 AI and Ecology: An Unlikely Alliance

Subscribe | AI Toolkit | Meet The Team

Welcome back AI prodigies!

In today’s Sunday Special:

• 📜The Prelude

• 🎞️Watching Without Disturbing

• 🤖Can AI Recognize Wildlife at Scale?

• ⚙️How CNNs Work, in Simple Terms.

• 🔑Key Takeaway

Read Time: 7 minutes

🎓Key Terms

• Serengeti: A vast geographical region in Africa filled with sweeping grasslands and diverse ecosystems that host the world’s most iconic wildlife.

• Convolutional Neural Networks (CNNs): A network for specialized layers that detect complex visual patterns, such as edges, textures, and structures.

🩺 PULSE CHECK

📜THE PRELUDE

Before sunrise, the African Savanna lies hushed. Then, a subtle click disturbs the stillness. It’s not an animal; it’s a camera trap, quietly capturing anything from the soft tread of an elephant’s foot to the glint of a hyena’s eyes.

These camera traps are critical for wildlife research, with each image representing a critical data point for understanding an animal’s behavior.

So, how do ecologists convert these images of animals into actionable insights? How do AI Systems help enhance ecological workflows?

🎞️WATCHING WITHOUT DISTURBING

📸 How Cameras Became Wildlife Tools.

By placing camera traps across the Serengeti, ecologists gain continuous, non-invasive monitoring of wildlife, capturing elusive behaviors like interactions between predator and prey. These captured interactions allow ecologists to precisely track anything from an animal’s population size to current migration patterns. This data helps inform how to allocate anti-poaching patrols and when to intervene to protect at-risk species. But this level of precision wasn’t always possible.

In the late 1980s, U.S. wildlife biologists first experimented with motion-triggered cameras to count common game species. These motion-triggered cameras were housed in metal casings the size of shoeboxes. Each camera was priced at roughly $3,800 in today’s dollars despite only storing up to 36 frames on a standard film cartridge and using batteries that lasted at most two weeks.

This short battery life forced U.S. wildlife biologists to hike dozens of miles to consistently capture images of animals over longer surveying sessions. The payoff was modest, with an entire season of fieldwork hopefully yielding 500 usable images because of poor lighting issues causing under-exposed film or over-exposed film. Despite these technological limitations, early remote monitoring of wildlife through motion-triggered cameras proved to be incredibly valuable.

👀 Counting the Unseen.

By the early 1990s, India’s wild tiger population had plummeted from about 10,000 wild tigers to just 3,000 wild tigers. Nagarahole National Park (NNP), one of the country’s last remaining wild tiger strongholds, became the focus of several critical surveying sessions. In 1991, Conservation Zoologist Dr. K. Ullas Karanth sought to estimate NNP’s remaining population of wild tigers. He placed motion-triggered cameras along commonly used trails for hunting and territorial marking. Within a year, he captured 31 images of multiple wild tigers across nine surveying sessions. So, how did he leverage 31 images to estimate the entire population of remaining wild tigers in India?

He relied on the Capture-Recapture Theory (C-RT). When a wild tiger is photographed once, that’s a Capture. If it’s photographed again, that’s a Recapture. By tracking Recaptures, the C-RT estimates the likelihood that any given wild tiger was missed. For instance, if multiple wild tigers are Recaptured across several surveying sessions, it suggests most of the wild tiger population has been identified.

However, if Captures are common and Recaptures are rare, it suggests that most of the wild tiger population hasn’t been identified yet. In this case, Recaptures were rare for Dr. K. Ullas Karanth. He concluded that the true wild tiger population was larger than the observed count of wild tigers. He identified that about 65 wild tigers lived in NNP, far more than the Indian government’s estimated 40 wild tigers.

🤖CAN AI RECOGNIZE WILDLIFE AT SCALE?

📊 Better Cameras, Bigger Datasets.

By 2005, technological innovation extended to cameras. Complementary Metal Oxide Sensors (CMOS), which convert light into electrical signals to capture digital images, enabled the creation of compact, power-efficient camera traps the size of a notebook. At the same time, memory card capacity increased from 10 megabytes (MB) to 2 gigabytes (GB), a 200x increase in the ability to capture and store high-resolution digital images. Alongside this, Infrared Flash replaced Visible Light, eliminating the white flashes that once spooked wildlife when cameras captured photographs. More importantly, camera prices fell. By 2010, a complete digital camera kit with solar-rechargeable batteries could be purchased for under roughly $1,310 in today’s dollars.

🏷️ Labeling Bigger Datasets.

In 2010, University of Minnesota (UMN) Graduate Alexandra Swanson launched the Snapshot Serengeti Project (SSP). She installed 225 camera traps across the Serengeti. Within two years, these camera traps generated more than 1.2 million images.

In 2012, Swanson published SSP through a public website where anyone could volunteer to label animals in images. Over the next three years, around 30,000 volunteers identified animals across more than 300,000 images. To classify them, Principal Data Scientist Margaret Kosmala developed a simple algorithm that counted the number of votes a specific animal received in a particular image. For instance, if a specific animal received more than 50% of the votes in a particular image, that particular image was labeled as containing that specific animal. This simple algorithm correctly identified 96% of animals in images. However, it wouldn’t have been possible without volunteers who spent over 50,000 hours voting on images.

🚀 Breaking the Labeling Bottleneck.

To address this challenge, Senior Research Scientist Mohammad (i.e., “Arash”) Norouzzadeh of the Bosch Center for AI (BCAI) created smaller wildlife image datasets using MegaDetectorV5. This CNN framework automatically filtered out images that didn’t contain animals.

Next, he applied Transfer Learning to retrain an AI System called ResNet-50 using these smaller wildlife image datasets. MegaDetectorV5 filtered images for animal presence, and ResNet-50 identified which species were in those filtered images, achieving 92% accuracy across 48 species.

⚙️HOW CNNs WORK, IN SIMPLE TERMS.

⦿ 1️⃣ Step 1: Input the Image

The CNN is given an image. This image appears as a grid of numbers (i.e., “pixels”), where each number represents a color or brightness level.

⦿ 2️⃣ Step 2: Find Patterns in the Image

The CNN uses small filters (i.e., think of them as little windows) that slide over the image.

Each small filter looks for specific shapes, like edges, corners, or textures, where colors in the image change sharply.

⦿ 3️⃣ Step 3: Keep the Good Stuff

The CNN keeps the specific shapes that seem important for identifying specific objects. This process helps the CNN focus on useful features within the image.

⦿ 4️⃣ Step 4: Make It Smaller

To make this process faster, the CNN reduces the image size using pooling, which keeps the specific shapes while shrinking other aspects of the image.

⦿ 5️⃣ Step 5: Repeat It Again

The previous steps are repeated multiple times. With each layer, the CNN learns more complex patterns within the image:

• Layer 1: Edges (e.g., “the curve of a lion’s ear”).

• Layer 2: Shapes (e.g., “the loop of a lion’s tail”).

• Layer 3: Objects (e.g., “a lion’s mane”).

⦿ 6️⃣ Step 6: Make a Decision

The CNN takes all the complex patterns identified within the image and sends them to the fully connected layer, which predicts what animal is represented in the image (e.g., “it’s a lion”).

⦿ 7️⃣ Step 7: The Output

The CNN generates a label for the specific shapes identified within the image (e.g., “lion”) along with a confidence score.

🔑KEY TAKEAWAY

AI Systems have transformed how ecologists monitor animals through images, turning a labor-intensive chore into a scalable, data-driven workflow. CNN frameworks now process millions of images of animals with higher accuracy in just hours, enabling ecologists to respond faster to the population decline of a species caused by habitat encroachments or seasonal poaching practices.

📒FINAL NOTE

FEEDBACK

❤️TAIP Review of The Week

REFER & EARN

🎉Your Friends Learn, You Earn!

You currently have 0 referrals, only 1 away from receiving 🎓3 Simple Steps to Turn ChatGPT Into an Instant Expert.

Share your unique referral link: https://theaipulse.beehiiv.com/subscribe?ref=PLACEHOLDER