Sonoma-LCC-Gen: A Neural Network Project for Land Cover Classification Based on Google Earth Engine Embeddings

Core Overview of the Sonoma-LCC-Gen Project

This project was co-developed by NickJC01, Broderick Noyes, and Nicholas Slankard, and released on GitHub on June 4, 2026 (link: https://github.com/NickJC01/Sonoma-LCC-Gen). Its core goal is to generate a high-precision 10-meter resolution land cover classification map for Sonoma County, California, USA. The unique feature is that it only uses 64-dimensional embedding vectors provided by Google Earth Engine (GEE) as input data without any additional supplementary data sources, demonstrating an innovative application of combining remote sensing data with deep learning.

Introduction to Google Earth Engine Embeddings

GEE is a cloud-based geospatial data processing platform. In recent years, it has introduced deep learning embedding functionality, which converts satellite images into 64-dimensional numerical vectors (embeddings). This is an abstract compression of raw remote sensing data and contains rich information about surface cover.

Advantages of Using Embeddings

1. Lightweight Data: Significantly reduces the volume of raw image data;
2. Efficient Computing: Directly input into downstream models, eliminating complex preprocessing;
3. Rich Features: Contains surface features learned by pre-trained models;
4. Standardized Input: Uniform format facilitates cross-time/region comparative analysis.

Neural Network Model Architectures

The project explores multiple models:

• Convolutional Neural Network (CNN): Extracts local spatial features, suitable for ground object texture/shape recognition;
• U-Net and its variants: Encoder-decoder structure for pixel-level classification;
• Other classifiers: May try fully connected networks, random forests, support vector machines, etc., to compare performance differences.

Project Structure and Workflow

• Infrastructure module: Responsible for data acquisition and preprocessing, including Drive mounting, tile generation (label shapefile + GEE embeddings to generate TIF), and Parquet file generation (with/without labels for inference);
• Models module: Contains multiple neural network implementations, trained using cloud GPUs in the Google Colab environment, with the latest test results retained.

Data Access Restrictions

The GEE project ('sonoma-lcc') used in the project is private, so other users will encounter permission errors when running the code. Solutions:

1. Create your own GEE project;
2. Modify the project ID configuration;
3. Generate the dataset according to the process.

Google Colab Dependency Issues

The project depends on Colab-specific features (e.g., Drive mounting). For local execution:

1. Adjust data path configuration;
2. Ensure correct GEE authentication;
3. Prepare GPU resources (recommended).

Precision Agriculture

Helps farmers understand farmland conditions, identify crop types, estimate planting areas, monitor growth, and support precise decision-making.

Forest Management

Sonoma County is rich in forest resources, and the classification results can be used for:

• Monitoring forest cover changes;
• Assessing health status;
• Planning sustainable forestry activities;
• Preventing/monitoring forest fires.

Urban Planning

10-meter resolution data clearly shows the distribution of built-up areas, green spaces, and water bodies, providing a basis for urban planning.

Climate Change Research

Long-term land cover change data is an important input for climate change impact research, helping to understand the impact of human activities and natural factors on the earth's surface.

Technical Insights

The project demonstrates a new paradigm for remote sensing data processing:

1. Cloud Computing: Uses GEE and Colab cloud resources to reduce local computing needs;
2. Embedding Representation: Pre-trained features simplify downstream data preparation;
3. Open Source Collaboration: Open-source code facilitates reproduction and expansion.

Future Outlook

• Integrate multi-source data (optical, radar, elevation) to improve classification accuracy;
• Implement time-series analysis to monitor dynamic changes in land cover;
• Develop lightweight models to support edge device deployment;
• Expand to larger regions or even global-scale land cover mapping.