Commodity Crop Model Documentation

Model Overview

Model Architecture

• Base Architecture: Attention U-Net with multi-input fusion
• Original Framework: TensorFlow 2.14.0

Hardware requirements

Software requirements

• Python >= 3.9
• Mamba (recommended for environment management)
• CUDA-compatible GPU

Key Dependencies for local users

• Tensorflow and Keras: Deep learning framework
• Scikit-learn: Machine learning
• Numpy: Data manipulation
• Rasterio: I/O georeferencing
• Google Earth Engine: For satellite data and elevation data access
• Geemap: Data download
• Matplotlib: For plotting

You can run the model either locally or via Colab. Below I provide both options. With the Colab option, you can run the model prediction interactively, and step-by-step code is provided. Of course, you can adapt it to your own environment.

Installation on local machine

Running the model on an interactive notebook in Google Colab (Everything is set and ready to run)

#The notebook allows you to:

🖼️ Draw or upload a Region of Interest (ROI) on an interactive map

🧠 Automatically selects AI model based on location (Africa, Southeast Asia, Latin America)

🛰️ Downloads and preprocesses Sentinel-1 + Sentinel-2 + elevation + indices

🌾 Predicts land use categories over deforested areas only using ONNX models

🗺️ Side-by-side map of RGB imagery + land use prediction

📤 Export predictions as GeoTIFF for GIS analysis

Model Input Specifications

• Input Name: "input"
• Shape: [1, 64, 64, 17]
  • Batch size: 1 (fixed)
  • Height: 64 pixels
  • Width: 64 pixels
  • Channels: 17 (combined features)
• Data Type: float32 (elem_type: 1)

Input Channel Organization

1. Sentinel-2 Bands (Channels 0-8):
   • Blue, Green, Red
   • Red Edge 1, 2, 3
   • NIR
   • SWIR 1, 2
   • Note: These bands are normalized using log-transformation and percentile-based scaling
2. Radar Data (Channels 9-10):
   • VV polarization (normalized to [-25, 0] range)
   • VH polarization (normalized to [-30, -5] range)
3. Geographical Information (Channels 11-13):
   • Altitude (normalized to [-400, 8000] range)
   • Longitude (normalized to [-180, 180] range)
   • Latitude (normalized to [-60, 60] range)
4. Additional Features (Channel 14):
   • Derived indices (NDVI, EVI, NDRE)

Output Specifications

• Output Name: "activation_65"
• Shape: [1, 64, 64, 22]
  • Batch size: 1 (fixed)
  • Height: 64 pixels
  • Width: 64 pixels
  • Classes: 1 (probability distribution over crop type)
• Data Type: float32 (elem_type: 1)

Land use following deforestation classes

Usage Example (Actual usage refers to interactive notebook)

Models

• Africa: {Africa Model}
• Southeast Asia: {SEA Model}
• Latin America: {LA Model}

Deployment Considerations

1. The model requires normalized input data according to the specified ranges for each channel
2. Memory requirements: Approximately 150MB for model weights
3. Compatible with any ONNX Runtime supported platform
4. Recommended to use ONNX Runtime version 1.20.0 or higher

ONNX Model Description

1. Channels 0–11 (12 channels): Processed by a U-Net-like structure for feature extraction and segmentation.
2. Channels 12–13 (2 channels): Handled by a second U-Net-like structure, mirroring the first.
3. Channels 14–16 (3 channels): Fed into a dense (fully connected) network to generate features, likely used as attention maps to guide the U-Nets.

Components

• Convolutional Layers: Extract spatial features using filters (e.g., 3x3 kernels), typically followed by batch normalization and ReLU activation for stability and efficiency.
• Max Pooling: Reduces feature map sizes (e.g., 64x64 to 32x32) during the encoder phase, capturing essential patterns while lowering computational load.
• Transpose Convolutions: Upsample feature maps in the decoder phase, restoring resolution (e.g., 16x16 to 32x32).
• Attention Mechanisms: Weight features based on attention maps from the dense network, focusing on critical areas.
• Concatenation: Merges features from different network parts, such as skip connections in U-Nets or the final fusion step.
• Sigmoid: Outputs class probabilities for the final segmentation map.

Detailed Structure

1. Input Splitting

• Channels 0–11: Concatenated into [1, 64, 64, 12].
• Channels 12–13: Concatenated into [1, 64, 64, 2].
• Channels 14–16: Concatenated into [1, 64, 64, 3].

2. Dense Network (Channels 14–16)

• Input: [1, 64, 64, 3], transposed to [1, 3, 64, 64].
• Structure: Mirrors the U-Net for channels 10–11:
  • Encoder: Downsamples to [1, 512, 8, 8] through levels with 64, 128, 256, and 512 filters.
  • Decoder: Upsamples back to [1, 64, 64, 64] with attention and skip connections.
• Output: [1, 64, 64, 64] feature map.
• Purpose: Provides an attention signal or context for the U-Nets.

3. U-Net for Channels 12–13

• Input: [1, 64, 64, 2], transposed to [1, 2, 64, 64].
• Encoder (Downsampling):
  • Level 1: Two 3x3 conv layers (64 filters), output [1, 64, 64, 64], then max pooling to [1, 64, 32, 32].
  • Level 2: Two 3x3 conv layers (128 filters), output [1, 128, 32, 32], then max pooling to [1, 128, 16, 16].
  • Level 3: Two 3x3 conv layers (256 filters), output [1, 256, 16, 16], then max pooling to [1, 256, 8, 8].
  • Bottom: Two 3x3 conv layers (512 filters), output [1, 512, 8, 8].
• Decoder (Upsampling with Attention):
  • Level 3: Upsampled to [1, 256, 16, 16], combined with encoder features via skip connection and attention, processed with conv layers.
  • Level 2: Upsampled to [1, 128, 32, 32], attention applied, concatenated, and conv layers.
  • Level 1: Upsampled to [1, 64, 64, 64], attention applied, concatenated, and conv layers.
• Output: [1, 64, 64, 64] feature map.

4. U-Net for Channels 0–11

• Input: [1, 64, 64, 10], transposed to [1, 10, 64, 64].
• Structure: Mirrors the U-Net for channels 10–11:
  • Encoder: Downsamples to [1, 512, 8, 8] through levels with 64, 128, 256, and 512 filters.
  • Decoder: Upsamples back to [1, 64, 64, 64] with attention and skip connections.
• Output: [1, 64, 64, 64] feature map.

5. Fusion and Output

• Concatenation: Outputs from both U-Nets and the dense network (resized to [1, 256, 64, 64]) are combined into [1, 384, 64, 64] (64 + 64 + 256 = 384, though channel counts may vary).
• Final Convolution: A 1x1 conv layer with 22 filters produces [1, 22, 64, 64].
• Softmax: Generates class probabilities.
• Output: Transposed to [1, 64, 64, 1], a segmentation map with 1 class per pixel.

Workflow Summary

1. Splitting: The 17-channel input is divided into three parts for specialized processing.
2. Feature Extraction:
   • U-Nets extract hierarchical features from channels 0–11 and 12–13 using encoder-decoder paths with skip connections.
   • The dense network processes channels 12–14 into a guiding feature map.
3. Attention: The dense network’s output weights U-Net features during upsampling, highlighting key areas.
4. Fusion: All features are combined to integrate multi-channel information.
5. Segmentation: A final layer classifies each pixel into one 1 class.

Design Philosophy

• Multi-Channel Inputs: Handles diverse data types across 17 channels.
• Attention-Driven Focus: Improves accuracy by emphasizing important regions.
• Detailed Segmentation: Preserves spatial details via U-Net skip connections.

Summary

License

Contact

Citation