Forest AI for Wildfire and Deforestation (FAWD)

This milestone focuses on gathering and organizing the necessary datasets for both wildfire detection and deforestation segmentation. For wildfire detection we will search for datasets containing images with labeled bounding boxes for fire and smoke instances. For deforestation detection we will collect satellite imagery datasets with labeled deforested and forested areas. The goal is to ensure balanced and diverse data representation which is essential for the effectiveness of subsequent model training.

Wildfire - Search and collect datasets on wildfire detection using object detection - Document data sources and verify proper licensing for all satellite imagery and wildfire datasets to ensure ethical data usage - Ensure datasets have labeled images with bounding boxes around fire/smoke. - Apply data augmentation (e.g. rotations flipping scaling noise addition) to enhance dataset diversity. - Split the dataset into training (75%) validation (15%) and testing (15%) with balanced representation of images with and without fire. Deforestation - Collect satellite imagery datasets focused on deforestation detection (Landsat Sentinel-2 etc.). - Ensure datasets include labeled masks showing deforestation or forest cover. - Apply data augmentation (e.g. rotations flipping scaling noise addition) to enhance dataset diversity. - Split the dataset into training (75%) validation (15%) and testing (15%).

$15,000 USD

- A ready-to-use dataset for wildfire detection with bounding box labels for fire/smoke instances. - A ready-to-use dataset for deforestation detection with labeled segmentation masks for deforested and forested areas.

This milestone involves preprocessing the collected datasets to prepare them for model training. For wildfire detection image resizing and normalization will ensure consistent input for the object detection model with proper bounding box annotations in the correct format. Similarly for deforestation detection satellite images and segmentation masks will be resized normalized and formatted into binary masks to represent deforested areas.

Wildfire - Resize all input images to a standard size (e.g. 416x416 or 640x640 adjusted depend on model used) maintaining aspect ratio with padding if necessary. - Normalize pixel values to the range [0 1]. - Convert bounding box annotations to the proper format (e.g. YOLO format: x_center y_center width height). - Address any class imbalance using oversampling or class weight adjustment. Deforestation - Resize satellite images and segmentation masks to a consistent size (256x256 or 512x512 pixels). - Normalize pixel values to the range [0 1]. - Convert segmentation masks to binary format (1 for deforested areas 0 for forested areas). - Address any class imbalance using oversampling or class weight adjustment.

$20,000 USD

- All images resized to consistent sizes (e.g., 416x416 or 640x640 for wildfire, 256x256 or 512x512 for deforestation). - Pixel values normalized to the range [0, 1] for both datasets. - Bounding boxes in the proper format for wildfire detection and binary masks for deforestation.

This milestone is focused on training the models for both wildfire detection and deforestation segmentation. For wildfire detection we will use YOLO a popular object detection model and implement a suitable loss function to handle classification localization and objectness. For deforestation we will use a segmentation model like U-Net or DeepLabV3+ and focus on segmentation-specific loss functions such as binary cross-entropy or Dice loss. Hyperparameter tuning will be performed using grid or random search to achieve the best model performance.

Wildfire - Select the YOLO model for object detection. - Implement a loss function combining classification loss localization loss and objectness loss. - Tune hyperparameters (learning rate batch size epochs anchor box sizes) using grid search or random search. - Evaluate the model using Intersection over Union. Deforestation - Choose a model suitable for segmentation tasks (e.g. U-Net DeepLabV3+ FCN). - Implement loss functions like binary cross-entropy or Dice loss. - Tune hyperparameters (learning rate batch size epochs). - Evaluate performance mainly using Intersection over Union.

$10,000 USD

- Wildfire model achieves target performance metrics (IoU). - Deforestation model achieves target performance metrics (IoU). - Comprehensive model performance report for both wildfire and deforestation detection.

This milestone involves implementing core safety and ethical considerations, followed by deploying both wildfire detection and deforestation detection services on the SingularityNET platform. The goal is to ensure the services provide reliable environmental monitoring capabilities while adhering to ethical principles and safety standards.

- Document the potential environmental impact of the service. Ensuring there are no storage of images data and immediate data disposal after processing. - Configure service settings for both wildfire and deforestation detection on SingularityNET. - Deploy the services to the SingularityNET marketplace, ensuring they are accessible to users. - Perform comprehensive testing on the SingularityNET platform to validate the functionality, performance, and reliability of both services. - Resolve any issues identified during testing to ensure smooth operation.

$5,000 USD

- Both services are successfully deployed on SingularityNET and accessible to users and other AI agents. - All identified issues during testing are resolved, and services operate smoothly.