CNN-based terrain classification from satellite and ground-level imagery, achieving 92.40% test accuracy with EfficientNet-B0 transfer learning across five terrain classes.
Developed for SIH (Smart India Hackathon) under the problem statement "Deep Learning for Terrain Recognition," this project iterates through seven custom CNN architectures and multiple transfer learning approaches. The model was progressively optimized from 150 MB down to 15.6 MB while improving accuracy, making it deployable on mobile and edge devices.
graph LR
A[Input Image] --> B[Data Pipeline]
B --> B1[Web Scraping]
B --> B2[10K Train / 500 Test]
B2 --> C{Architecture}
C --> D[Custom CNN V1-V7]
C --> E[Transfer Learning]
D --> D1[Xception: 92.8%]
E --> E1[EfficientNet-B0: 92.40%]
E --> E2[ConvNeXT Tiny: 92.11%]
E --> E3[MobileNetV2: 91.33%]
D1 & E1 --> F[Prediction]
F --> G1[Coast]
F --> G2[Desert]
F --> G3[Forest]
F --> G4[Glacier]
F --> G5[Mountain]
The custom CNN architecture used across V1-V7 experiments:
graph TD
A["Input: 64x64 RGB"] --> B["Conv2D(32, 3x3) + ReLU"]
B --> B1["BatchNorm + MaxPool(2x2)"]
B1 --> C["Conv2D(64, 3x3) + ReLU"]
C --> C1["BatchNorm + MaxPool(2x2)"]
C1 --> D["Conv2D(128, 3x3) + ReLU"]
D --> D1["BatchNorm + MaxPool(2x2)"]
D1 --> E["Flatten"]
E --> F["Dense(1024) + ReLU + BatchNorm"]
F --> F1["Dropout(0.2)"]
F1 --> G["Dense(1024) + ReLU + BatchNorm"]
G --> G1["Dropout(0.2)"]
G1 --> H["Dense(5) + Softmax"]
H --> I["Output: 5 Classes"]
| Model | Architecture | Test Accuracy | Model Size | Framework |
|---|---|---|---|---|
| EfficientNet-B0 | Transfer Learning | 92.40% | 15.6 MB | PyTorch |
| Xception | Transfer Learning | 92.8% | 23 MB | Keras |
| ConvNeXT Tiny | Transfer Learning | 92.11% | 106 MB | PyTorch |
| MobileNetV2 | Transfer Learning | 91.33% | 9.71 MB | PyTorch |
| Custom CNN V6 | RMSprop | 88.46% | ~150 MB | Keras |
| Custom CNN V4 | Adam (4-layer) | 82.69% | ~150 MB | Keras |
| Custom CNN V5 | SGD + Momentum | 80.77% | ~150 MB | Keras |
| Custom CNN V7 | Adagrad | 76.92% | ~150 MB | Keras |
| Optimizer | Version | Learning Rate | Params | Train Acc | Val Acc |
|---|---|---|---|---|---|
| RMSprop | V6 | 0.001 | default | 77.67% | 88.46% |
| Adam | V4 | default | default | 76.43% | 82.69% |
| SGD + Momentum | V5 | 0.01 | momentum=0.9 | 75.83% | 80.77% |
| Adagrad | V7 | 0.01 | default | 74.39% | 76.92% |
- RMSprop achieved the best validation accuracy among classical optimizers due to its adaptive learning rate, particularly effective for high-variance image datasets.
- SGD with Momentum showed steady but slower improvement, plateauing after a certain number of epochs without learning rate scheduling.
- Adagrad exhibited early convergence due to accumulated gradient squares, leading to diminishing learning rates.
- Transfer learning models universally outperformed custom CNNs, with Xception achieving the highest single-model accuracy at 92.8%.
- MobileNetV2 offers the best accuracy-to-size ratio at 91.33% with only 9.71 MB, ideal for edge deployment.
| Metric | Value |
|---|---|
| Parameters | ~4 million |
| Model Size | ~15.6 MB |
| Memory Usage | ~34.7 MB |
| CPU Inference | ~30-60ms/image |
| GPU Inference | ~10-15ms/image |
| Validation Accuracy | 91.93% |
| Test Accuracy | 92.40% |
Training config: Adam optimizer, LR=1e-4, batch size=32, 25 epochs with early stopping.
- Parameters: 27.8 million | Size: 106 MB
- Accuracy: 92.11% | Loss: 0.2168
- Pre-trained on ImageNet, fine-tuned with standard 224x224 transforms
- Parameters: 2.2 million | Size: 9.71 MB
- Accuracy: 91.33% | Loss: 0.2798
- Lightweight architecture designed for mobile deployment
- Base: Xception (ImageNet pretrained)
- Head: GlobalAveragePooling → Dense(1024, ReLU) → Dense(5, Softmax)
- Accuracy: 92.8% on test set
- Training on 150x150 images with EarlyStopping and ReduceLROnPlateau
- Source: Custom-scraped via BeautifulSoup + Selenium
- Classes: Coast, Desert, Forest, Glacier, Mountain
- Split: 10,000 training / 500 test images
Terrain-Recognition/
├── README.md
├── LICENSE
├── CONTRIBUTING.md
├── CODE_OF_CONDUCT.md
├── requirements.txt
├── .gitignore
├── .github/
│ └── workflows/
│ └── greetings.yml
├── assets/ # Diagrams and charts
│ ├── analysis_chart.png
│ ├── data_sample.png
│ ├── model_v2_diagram.png
│ ├── optimization.png
│ ├── training_history.png
│ └── v2_architecture.png
├── data/
│ ├── train/ # 10,000 training images
│ │ ├── coast/
│ │ ├── desert/
│ │ ├── forest/
│ │ ├── glacier/
│ │ └── mountain/
│ └── test/ # 500 test images
│ ├── coast/
│ ├── desert/
│ ├── forest/
│ ├── glacier/
│ └── mountain/
├── notebooks/
│ ├── experiments/ # Custom CNN iterations
│ │ ├── v1_baseline.ipynb
│ │ ├── v2_improved.ipynb
│ │ ├── v3_deeper_cnn.ipynb
│ │ ├── v4_4layer_cnn.ipynb
│ │ ├── v5_sgd_momentum.ipynb
│ │ ├── v6_rmsprop.ipynb
│ │ └── v7_adagrad.ipynb
│ ├── transfer_learning/
│ │ ├── efficientnet_b0.ipynb # Best model (92.40%)
│ │ ├── convnext_mobilenet.ipynb # ConvNeXT + MobileNet
│ │ └── past/
│ │ ├── transfer_learning_v1.ipynb
│ │ ├── cal_inference.ipynb
│ │ └── load_model.ipynb
│ ├── xception/
│ │ └── xception_v1.ipynb
│ └── inference.ipynb # Model inference demo
├── models/
│ ├── xception/
│ │ └── model_xception.h5
│ └── transfer_learning/
│ ├── best_model.pth
│ ├── best_model_sg.pth
│ └── efficientnet_terrain_classifier.pth
└── results/
├── xception_confusion_matrix.png
└── transfer_learning/
├── confconnet.jpg
├── confnet.jpg
├── convnextaccu.jpg
├── convnextarch.jpg
├── mobilenetacc.jpg
└── mobilenetarch.png
git clone https://github.com/Akasxh/Terrain-Recognition.git
cd Terrain-Recognition
pip install -r requirements.txt
# Best model (EfficientNet-B0)
jupyter notebook notebooks/transfer_learning/efficientnet_b0.ipynb
# Xception model
jupyter notebook notebooks/xception/xception_v1.ipynb
# Inference
jupyter notebook notebooks/inference.ipynb- Frameworks: TensorFlow/Keras, PyTorch
- Architectures: Custom CNN, Xception, EfficientNet-B0, ConvNeXT Tiny, MobileNetV2
- Data Collection: BeautifulSoup, Selenium
- Visualization: matplotlib, seaborn
- Training: Google Colab, local GPU
Lead: S Akash
Contributors
- Vihaan Agrawal
- Ruchi Chand Thakur
- Manas Gupta
- Tanmay Singh
- Harshith Patnaik
Licensed under GPL-3.0. See LICENSE for details.