מושגי ליבה
Contrastive learning improves regression on hyperspectral data.
תקציר
The content discusses the application of contrastive learning in regression tasks for hyperspectral data. It introduces a framework with various transformations to augment hyperspectral data and improve regression performance. The experiments conducted on synthetic and real datasets show significant enhancements in regression models compared to state-of-the-art techniques. The paper also explores related work, proposed methods, and presents results indicating the effectiveness of the contrastive learning approach.
Abstract:
Contrastive learning effective for representation learning.
Shortage in studies targeting regression tasks on hyperspectral data.
Proposed framework enhances regression model performance.
Introduction:
Hyperspectral imagery valuable for object analysis without physical contact.
Gain attention for classification, regression, unmixing, and object detection tasks.
Self-supervised learning gaining popularity due to limited labeled data.
Method:
Proposed framework for pixel-level regression on hyperspectral data.
Spectral data augmentation methods introduced.
Contrastive loss integrated into training process.
Experiments & Results:
Synthetic Data:
Various spectral transformations applied to enhance model performance.
Shift and elastic transformations provided top results.
Combination study shows improved metrics with multiple transformations.
Real Soil Data:
Dataset used for soil pollution analysis with hydrocarbon concentration.
Shift, flip, and elastic transformations yield best results.
Conclusion:
Contrastive learning improves regression tasks on hyperspectral data.
Clear enhancement seen in both synthetic and real datasets.
סטטיסטיקה
This work is funded by Tellux Company and ANRT (Association Nationale de la Recherche et de la Technologie).