Model Documentation: DeepLabV3+ for Semantic Segmentation

Version: v6.0 | Date: 02/09/2025
Author(s): Matthew Kutugata

1. Overview

Model Name: DeepLabV3+ (ResNet-152 Backbone)
Purpose: Segment and classify pixels in SemiF images into background and plants for automatic annotation purposes.
Pipeline Integration: Used in the preprocessing stage of the AgIR SemiField pipeline to generate semantic masks.
Status: Deployed

2. Model Details

2.1. Input & Output

• Input: RGB images (JPEG) at 512x512 resolution

• Output: Segmentation mask with integer class labels

2.2. Training Data

• Dataset Source: SemiF Agricultural Image Repository (AgIR) and Synthetic data

• Train Images: 75,399

• Validation Image: 9,425

• Test Images: 9,418

• Annotations Used: Pixel-wise labeled masks for plants and background

• Preprocessing Steps:

  • Normalization :

  • (mean=[0.398, 0.380, 0.293] std=[0.186, 0.183, 0.153])

2.3. Model Architecture & Hyperparameters

• Base Model: DeepLabV3+ with a ResNet-152 backbone

• Fine-tuned or Trained from Scratch? Fine-tuned on agricultural datasets

• Key Hyperparameters:

  • Learning Rate: 1e-3

  • Batch Size: 32

  • Epochs: 50

  • Optimizer: Adam

  • Loss Function: Dice Loss

3. Model Performance & Evaluation

3.1. Metrics

• mIoU (Mean Intersection over Union): 92.6%

• Pixel Accuracy: 92.4%

• Precision / Recall / F1-score:

  • Crops: P=85.3%, R=80.7%, F1=82.9%

  • Weeds: P=74.2%, R=71.6%, F1=72.9%

3.2. Limitations & Known Issues:

• Struggles with shadows and occlusions in dense crop fields

• Requires high-resolution images for best results

4. Deployment & Integration

4.1. Deployment Details

• Environment: HPC Cluster (SLURM) and Local GPU Workstation

• Inference Hardware: NVIDIA A100 (HPC) / RTX 4090 (Local)

• Model Format: PyTorch .pth file

• Inference Speed: ~85ms per 1024x1024 image

4.2. Pipeline Integration

• Preprocessing Dependencies:

  • Image resizing to 1024x1024

  • Normalization

• Postprocessing Steps:

  • Convert mask to class-wise labeled PNG

  • Overlay with input image for visualization

• Where is it Used in the Pipeline?

  • Step 2: Semantic segmentation of crops and weeds

• Integration Code Snippet:

5. Maintenance & Future Improvements

• Retraining Requirements: Every 6 months based on new field imagery

• Planned Improvements:

  • Fine-tune with more diverse lighting conditions

  • Implement real-time inference with TensorRT for Jetson devices

  • Add multi-spectral image support

• Responsible Team / Contact: [Your Name] ([Your Email])

6. References & Additional Resources

• Model Training Logs: [Link to TensorBoard logs]

• Code Repository: [GitHub or internal repo link]