🌳 Forest Guardian AI

Real-time Forest Sound & Image Threat Detection

An AI-powered forest monitoring system that detects illegal activities — chainsaw sounds, gunshots, and heavy machinery — in real time using deep learning audio classification and Gemini-powered visual analysis.

---

📋 Table of Contents

1. Problem Statement
2. System Architecture
3. Dataset
4. Classes Predicted
5. Features Used
6. Algorithms & Model Evolution
7. CNN Architecture (Final)
8. Training Strategy
9. Results & Accuracy
10. Project Structure
11. How to Run
12. Streamlit Web App
13. Deployment
14. Tech Stack

---

🚨 Problem Statement

Illegal deforestation, poaching, and unlawful forest activity cause irreversible environmental damage. Manual forest patrols are:

• Expensive — require large teams across vast areas
• Reactive — threats are detected too late
• Impractical at scale — forests can span thousands of kilometres

Forest Guardian AI solves this by providing an automated, AI-powered monitoring system that:

• 🎙️ Listens to forest audio and flags chainsaw sounds, gunshots, and heavy machinery
• 🖼️ Analyzes camera trap images for visual signs of illegal activity
• ⚡ Works in real-time with a confidence threshold to reduce false alarms

---

🏗 System Architecture

┌─────────────────────────────────────────────────────────┐
│                  Forest Guardian AI                      │
│                    (Streamlit App)                       │
└──────────────────────┬──────────────────────────────────┘
                       │
          ┌────────────┴────────────┐
          │                         │
          ▼                         ▼
┌─────────────────┐       ┌──────────────────────┐
│  AUDIO MODULE   │       │    IMAGE MODULE       │
│ audio_model/    │       │   image_model/        │
│                 │       │                       │
│ 1. Upload .wav  │       │ 1. Upload image       │
│ 2. Extract Mel  │       │ 2. Send to Gemini API │
│    Spectrogram  │       │ 3. Parse JSON threat  │
│ 3. Normalize    │       │    assessment         │
│ 4. CNN Predict  │       │ 4. Display result     │
│ 5. Show result  │       └──────────────────────┘
└────────┬────────┘
         │
         ▼
┌─────────────────┐
│  CNN MODEL      │    Input: (64 × 130 × 1) Mel Spectrogram
│  4 Conv Blocks  │    Output: 4-class softmax
│  ~4.5M params   │    Threshold: ≥ 65% confidence
└─────────────────┘

Data Flow (Audio):

WAV File → Librosa Load → Mel Spectrogram (64×130) → Normalize → CNN → Softmax Probabilities → Prediction

Data Flow (Image):

Image Upload → PIL → Gemini 2.0 Flash API → JSON Response → Threat / Safe UI

---

📦 Dataset

Property	Detail
Format	WAV, mono, 22,050 Hz, 3 seconds
Total classes	4 core + 4 extended (future)
Balancing	Augmentation to equalise minority classes

Sources explored:

• 🔊 ESC-50 Environmental Sound Dataset
• 🔊 UrbanSound8K
• 🔊 Custom-curated & recorded .wav files
• 🔊 Augmented copies of minority classes

Data Cleaning Rules:

• Format: WAV only (MP3 converted via ffmpeg / librosa)
• Duration: exactly 3 seconds (padded or trimmed)
• Sample rate: 22,050 Hz
• Channels: Mono (stereo mixed down)

---

🎯 Classes Predicted

Core Classes (trained)

Class	Description	Threat Level
🪚 chainsaw	Electric / petrol chainsaw sounds	🔴 HIGH
🔫 gunshot	Single or burst gunfire	🔴 HIGH
🚜 heavy_machine	Bulldozers, excavators, trucks	🔴 HIGH
🍃 normal	Ambient forest — birds, wind, rain	🟢 SAFE

Predictions below 65% confidence are flagged as OTHER / AMBIENT to prevent false alarms.

Extended Classes (dataset collected, future training)

Class	Description
dangerous_animals	Predator / dangerous wildlife calls
wildlife_large	Elephant, rhino, large animal movement
mistake_mimicry	Sounds that mimic threats but are not
human_activity	Footsteps, voices (non-illegal)

---

🔬 Features Used

Primary: Mel Spectrogram

Raw audio waveforms are converted into Mel Spectrograms — 2D image representations of sound that mimic how the human ear perceives frequency (logarithmic scale). This allows the CNN to treat audio classification as a visual pattern recognition problem.

SAMPLE_RATE = 22050   # Hz
DURATION    = 3       # seconds
N_MELS      = 64      # Mel filter banks  → image height
MAX_LEN     = 130     # time-axis columns → image width

Each audio clip becomes a (64 × 130) grayscale image fed into the CNN.

Why Mel Spectrograms over raw MFCCs?

Feature	MFCC	Mel Spectrogram
Information retained	Compressed	Full spectral detail
CNN suitability	Moderate	✅ Excellent (image-like)
Noise robustness	Good	Very good
Used in final model	Baseline only	✅ Yes

Data Augmentation (applied during training)

Technique	Details
Time stretching	±10% speed change
Pitch shifting	±2 semitones
Gaussian noise	Small random noise injection
Time shift	Circular roll of waveform ±20%
SpecAugment	Random frequency + time band masking

---

🤖 Algorithms & Model Evolution

The model went through 4 major iterations:

Version	Algorithm	Feature	Accuracy
Baseline	Logistic Regression	Raw MFCC	~62%
v1	Decision Tree	MFCC	~72%
v1	Random Forest	MFCC	~78%
v2 (CNN v1)	3-block CNN + Flatten	Mel Spectrogram	~84.9%
v2 (Final)	4-block CNN + GAP + AdamW + SpecAugment	Mel Spectrogram	94.4%

Key Improvements in Final CNN

Change	Impact
Flatten → GlobalAveragePooling2D	Reduced overfitting significantly
Added 4th Conv block (256 filters)	Deeper feature learning
SpecAugment during training	Better generalisation to real-world noise
AdamW optimizer (weight decay)	Prevented weight explosion
Label smoothing (0.1)	Prevented overconfident predictions

---

🧠 CNN Architecture (Final)

Input: (64, 130, 1)  ←  Mel Spectrogram as grayscale image
│
├── Conv2D(32, 3×3, ReLU) → BatchNorm → MaxPool(2×2) → Dropout(0.25)
│
├── Conv2D(64, 3×3, ReLU) → BatchNorm → MaxPool(2×2) → Dropout(0.25)
│
├── Conv2D(128, 3×3, ReLU) → BatchNorm → MaxPool(2×2) → Dropout(0.25)
│
├── Conv2D(256, 3×3, ReLU) → BatchNorm → GlobalAveragePooling2D
│
├── Dense(256, ReLU) → Dropout(0.5)
├── Dense(128, ReLU) → Dropout(0.3)
│
└── Dense(4, Softmax)  ←  Output: probability per class

Model stats:

• Total parameters: ~4.5 million
• Model size: ~52 MB (.h5 format)
• Input shape: (64, 130, 1)
• Output: 4-class softmax

---

⚙️ Training Strategy

optimizer  = AdamW(learning_rate=1e-3, weight_decay=1e-4)
loss       = CategoricalCrossentropy(label_smoothing=0.1)
epochs     = 50          # with early stopping
batch_size = 32
val_split  = 0.20        # 80% train / 20% validation

Callbacks:

Callback	Config
EarlyStopping	patience=10, restore best weights
ReduceLROnPlateau	factor=0.5, patience=5, min_lr=1e-6
ModelCheckpoint	saves best val_accuracy model

---

📊 Results & Accuracy

Evaluated on a held-out test set of original (non-augmented) forest audio files:

Metric	Value
Overall Accuracy	94.4%
Weighted F1-Score	0.944
Chainsaw Recall	95%
Gunshot Recall	95%
Heavy Machine Recall	93%
Normal Forest Recall	95%

Confidence Threshold

Predictions below 65% confidence are shown as OTHER / AMBIENT — this prevents the model from making overconfident wrong calls on unknown sounds.

Confidence	Label shown	Colour
≥ 65% + threat class	Class name (e.g. CHAINSAW)	🔴 Red
≥ 65% + normal	NORMAL	🟢 Green
< 65%	OTHER / AMBIENT	⚫ Grey

---

📁 Project Structure

Audio_Classification_Model/
│
├── audio_model/                  # CNN audio classifier
│   ├── __init__.py
│   ├── classifier.py             # extract_mel(), predict(), constants
│   ├── forest_sound_model_v2.h5  # trained model weights (~52 MB, Git LFS)
│   └── norm_stats.npy            # mean & std from training data
│
├── image_model/                  # Gemini visual analyzer
│   ├── __init__.py
│   └── analyzer.py               # analyze_image() via Gemini API
│
├── app.py                        # Streamlit entry-point (UI only)
├── requirements.txt              # Python dependencies
├── packages.txt                  # Linux system libs (Streamlit Cloud)
├── .gitignore                    # excludes .env, __pycache__, etc.
├── .gitattributes                # Git LFS config for model files
├── model_summary.md              # Full technical walkthrough
└── README.md

---

🚀 How to Run

Prerequisites

• Python 3.10+
• A Google Gemini API key → get one free at aistudio.google.com

1. Clone the repository

git clone https://github.com/ishwarc04/Audio_Classification_Model.git
cd Audio_Classification_Model

2. Install dependencies

pip install -r requirements.txt

3. Set your API key

Create a .env file in the project root:

GOOGLE_API_KEY=your_gemini_api_key_here

4. Run the app

streamlit run app.py

Open http://localhost:8501 in your browser.

---

🖥️ Streamlit Web App

The app has two tabs:

🎙️ Audio Monitoring Tab

1. Upload a .wav file
2. App extracts the Mel Spectrogram
3. Spectrogram is normalized using saved norm_stats.npy
4. CNN predicts class probabilities
5. Results shown with colour-coded alert + full probability bars

🖼️ Image Monitoring Tab

1. Upload a forest image (.jpg, .png)
2. Image is sent to Gemini 2.0 Flash
3. AI returns a structured JSON threat assessment
4. Results shown with confidence score + threat/safe status
5. Adjustable confidence threshold slider

---

☁️ Deployment

Deployed on Streamlit Community Cloud (free tier).

Steps to deploy your own

1. Push repo to a public GitHub repository
2. Go to share.streamlit.io and log in with GitHub
3. Select app.py as the entry point
4. Under App Settings → Secrets, add:

   GOOGLE_API_KEY = "your_key_here"

5. App is live at a public URL — no server management needed

Key deployment files:

File	Purpose
app.py	Streamlit entry point
audio_model/forest_sound_model_v2.h5	Trained model (Git LFS)
audio_model/norm_stats.npy	Normalization stats (Git LFS)
requirements.txt	Python packages
packages.txt	Linux libs (libsndfile1 for audio)

---

🛠 Tech Stack

Tool	Purpose
Python 3.10+	Core language
TensorFlow / Keras	CNN training & inference
Librosa	Audio loading & Mel Spectrogram extraction
NumPy / Scikit-learn	Data handling, metrics, class weights
Matplotlib	Spectrogram visualization
Streamlit	Web dashboard
Google Gemini 2.0 Flash	Image threat analysis (Vision AI)
Pillow	Image handling
python-dotenv	Local environment variable management
Git LFS	Large file storage for model weights

---

Built for forest conservation and environmental protection. 🌿

If this project helped you, please ⭐ star the repository!