LISA Framework

Light-Invariant Spectral Autoencoder for precision viticulture

Light-Invariant Spectral Autoencoder (LISA)

LISA is a domain-adversarial deep learning framework designed to overcome the fundamental challenge of illumination variance in hyperspectral imaging for agricultural applications.

The Challenge

Field-deployed optical sensing systems face a critical problem: data captured under different lighting conditions (morning sun, afternoon, cloudy, lab-controlled) look spectrally different even for identical samples. This “domain shift” severely degrades the performance of standard machine learning models when deployed in real-world conditions.

Our Solution

LISA employs domain-adversarial learning to force the model to learn illumination-invariant features. The network is trained to:

  1. Accurately predict grape quality (Brix and acidity levels)
  2. Be unable to distinguish between different lighting conditions

This adversarial approach ensures the learned representations are robust to environmental changes.

Key Results

  • 20% improvement in generalization compared to baseline models
  • Successfully deployed in real vineyards for in-field grape quality assessment
  • Part of a complete IoT-enabled robotic system for precision viticulture
  • Published in IEEE Internet of Things Journal (2025)

Applications

  • Non-destructive grape quality prediction
  • Real-time yield mapping
  • Spatially-resolved vineyard management
  • Harvest optimization

Related Publication: In-Field Mapping of Grape Yield and Quality with Illumination-Invariant Deep Learning (IEEE IoT Journal, 2025)