Skip to content

Self-Organizing Maps (SOM)

CDTS includes a highly optimized, multi-threaded C++ implementation of the Batch Self-Organizing Map (Batch SOM) algorithm, based on the model originally proposed by Kohonen (1990) (see References). It is specifically designed to handle large Earth Observation data cubes efficiently using OpenMP and Eigen (SIMD vectorization).

Self-Organizing Maps are unsupervised neural networks used to reduce the dimensionality of your data, clustering similar temporal profiles into a 2D grid. It is incredibly useful for discovering land cover patterns without prior training data.

Training a SOM

You can train a SOM directly on a 2D numpy array (Pixels x Features). The underlying C++ engine handles the heavy lifting safely.

import numpy as np
from cdts.ai import train_som_batch, predict_bmus

# Simulate 10,000 pixels with 24 features (e.g., 4 bands x 6 time steps)
X_train = np.random.rand(10000, 24).astype(np.float64)

# Train a 10x10 SOM
# C++ OpenMP automatically uses available cores (leaving 1 free to keep the OS responsive)
som_weights = train_som_batch(
    data=X_train,
    grid_rows=10,
    grid_cols=10,
    num_epochs=100,
    initial_learning_rate=0.5,
    n_jobs=-1 # Automatically uses all cores minus 1
)

print("Trained SOM Weights Shape:", som_weights.shape) # (100, 24)

Predicting Best Matching Units (BMUs)

Once trained, you can classify new pixels by finding their Best Matching Unit (BMU) on the 2D grid. 

# Simulate new data
X_new = np.random.rand(5000, 24).astype(np.float64)

# Predict the BMU index for each pixel (values from 0 to 99)
bmus = predict_bmus(X_new, som_weights, n_jobs=-1)

print("BMUs Shape:", bmus.shape) # (5000,)

Hardware & Threading (n_jobs)

Both train_som_batch and predict_bmus expose the n_jobs parameter to control the level of parallelization via C++ OpenMP.

  • n_jobs = -1 (Default): Uses max_threads - 1. Your machine will stay responsive during heavy training.
  • n_jobs = 4: Forces the use of exactly 4 cores.
  • n_jobs = 1: Disables OpenMP (runs sequentially, great for debugging).

References