|
|
--- |
|
|
tags: |
|
|
- computer-vision |
|
|
- image-classification |
|
|
- remote-sensing |
|
|
- sar-imagery |
|
|
- land-cover-mapping |
|
|
datasets: |
|
|
- Sentinel1_LandCover_SAR_Patches |
|
|
license: cc-by-sa-4.0 |
|
|
model-index: |
|
|
- name: SatelliteSAR_LandCover_Classifier |
|
|
results: |
|
|
- task: |
|
|
name: Image Classification (SAR) |
|
|
type: image-classification |
|
|
metrics: |
|
|
- type: accuracy |
|
|
value: 0.887 |
|
|
name: Classification Accuracy |
|
|
- type: kappa_score |
|
|
value: 0.854 |
|
|
name: Cohen's Kappa |
|
|
--- |
|
|
|
|
|
# SatelliteSAR_LandCover_Classifier |
|
|
|
|
|
## 📡 Overview |
|
|
|
|
|
The **SatelliteSAR_LandCover_Classifier** is a deep Convolutional Neural Network (CNN), based on a **ResNet** architecture, fine-tuned for classifying land cover types using **Synthetic Aperture Radar (SAR)** imagery. Unlike optical imagery, SAR penetrates clouds and operates day or night, making it ideal for continuous monitoring. This model classifies 128x128 pixel patches into one of five key land cover types. |
|
|
|
|
|
## 🧠 Model Architecture |
|
|
|
|
|
The model utilizes a modified ResNet architecture, adapted to handle the unique characteristics of SAR data, which typically involves two channels (VV and VH polarization) instead of the three RGB channels of optical images. |
|
|
|
|
|
* **Base Model:** A version of **ResNet-50** with the initial convolutional layer modified to accept 2 input channels. |
|
|
* **Input:** SAR image patches (128x128) with two polarization channels (VV/VH). |
|
|
* **Feature Extraction:** The deep ResNet residual blocks are crucial for extracting complex textural and structural patterns specific to SAR (e.g., distinguishing between rough forest canopies and smooth water surfaces). |
|
|
* **Classification Head:** A standard fully connected layer predicts the probability distribution over the five land cover classes. |
|
|
* **Target Classes:** Water\_Bodies, Urban\_Areas, Forests, Agriculture, and Barren\_Land. |
|
|
|
|
|
## 🎯 Intended Use |
|
|
|
|
|
This model is critical for operational geospatial intelligence and environmental monitoring: |
|
|
|
|
|
1. **Disaster Monitoring:** Mapping flood extent (identifying Water\_Bodies) under cloud cover, where optical sensors fail. |
|
|
2. **Deforestation Detection:** Monitoring changes in the Forests class regardless of weather or time of day. |
|
|
3. **Urban Sprawl Tracking:** Regularly updating maps of the Urban\_Areas class. |
|
|
4. **Agricultural Management:** Classifying Agriculture areas to track seasonal changes and crop type. |
|
|
|
|
|
## ⚠️ Limitations |
|
|
|
|
|
1. **Speckle Noise:** SAR data inherently contains "speckle" noise, which can reduce classification accuracy, though the CNN is robustly trained against it. |
|
|
2. **Confusion Matrix:** It may occasionally confuse low-density urban areas with complex Agriculture/Forest boundaries due to similar scattering patterns. |
|
|
3. **Feature Dependence:** It relies solely on backscatter intensity and ignores spectral (color) information, which can sometimes be helpful for classification. |
|
|
|
|
|
--- |
|
|
|
|
|
### MODEL 5: **QuantumCircuit_Optimization_RL** |
|
|
|
|
|
This model is a Proximal Policy Optimization (PPO) reinforcement learning agent for optimizing the depth of quantum circuits. |
|
|
|
|
|
#### config.json |
|
|
|
|
|
```json |
|
|
{ |
|
|
"_name_or_path": "custom-ppo-quantum-optimizer", |
|
|
"architectures": [ |
|
|
"PPOAgentForQuantumCircuitOptimization" |
|
|
], |
|
|
"model_type": "reinforcement_learning", |
|
|
"environment": "QuantumCircuitEnv-v1", |
|
|
"state_space_size": 256, |
|
|
"action_space_size": 10, |
|
|
"policy_network": "MLP", |
|
|
"hidden_layers": [128, 128], |
|
|
"gamma": 0.99, |
|
|
"lambda_gae": 0.95, |
|
|
"learning_rate": 3e-4, |
|
|
"optimization_goal": "Minimize Circuit Depth", |
|
|
"reward_function": "CircuitDepthReductionDelta", |
|
|
"transformers_version": "4.36.0" |
|
|
} |
