Abstract
Hyperspectral tree species classification in tropical forests is often hindered by spectral variability caused by changing environmental conditions such as sun angle and atmospheric haze. This paper presents a self-supervised learning (SSL) approach using the Barlow Twins algorithm to improve classification robustness. By leveraging temporal pairing—treating pixels from the same location across different flight acquisitions as positive pairs—the model learns to ignore abiotic noise. Results demonstrate a 10-point increase in classification accuracy across different dates compared to standard reflectance data, offering a promising path for automated spectral calibration in remote sensing.
Introduction
Airborne hyperspectral imaging is often considered the "holy grail" for tropical forest management, enabling the identification of tree species across vast, inaccessible areas. However, spectral signatures are not fixed; they vary significantly based on sun angle, atmospheric conditions, and moisture levels. This sensitivity to abiotic variability makes it difficult to scale botanical identification models across different times or locations without constant, expensive ground-truth recalibration. This study addresses the challenge of spectral variability by developing a robust classification framework using self-supervised learning.
Methods
Self-Supervised Learning Workflow
The authors employ a Self-Supervised Learning (SSL) workflow based on the Barlow Twins algorithm. The core innovation lies in the definition of "positive pairs" through Temporal Pairing. Instead of relying solely on synthetic augmentations, the model is trained using two hyperspectral images of the same forest captured on consecutive days. Pixels from the exact same coordinates across these two dates are treated as positive pairs, forcing the model to learn representations that are invariant to the lighting and atmospheric differences between the two acquisitions.
Data Augmentation and Redundancy Minimization
To further enhance robustness, the model incorporates spectral augmentations, including random band swapping, additive noise, and "Scaling Domain" transforms that mimic changes in light intensity. The Barlow Twins objective function decorrelates features by forcing the cross-correlation matrix between the two views toward the identity matrix, making the resulting encodings more resilient to noise.
Experimental Setup
The study was conducted at the Paracou research station in French Guiana. The dataset includes hyperspectral data from Hyspex VNIR and SWIR sensors (343 spectral bands) and a ground-truth survey of 1,376 trees across 40 species. Acquisitions were made on September 19 and 20, 2016, under different timing and coverage conditions.
Results and Discussion
Performance Gains
The most significant finding was that inter-date temporal pairing consistently outperformed single-date training with synthetic augmentations. Features generated via the SSL temporal pairing method achieved a 10-point boost in classification accuracy (reaching approximately 40%) compared to the baseline Linear Discriminant Analysis (LDA) performed on pre-processed reflectance data (30%).
Key Insights
- Automated Calibration: SSL acts as a form of "self-supervised calibration," learning to ignore atmospheric noise to isolate the underlying biological signal.
- Synthetic vs. Natural Variability: Natural variability between flight days (capturing complex effects like BRDF) is more effective for training robust models than simple mathematical transforms.
- Scaling Effectiveness: Among synthetic methods, domain scaling proved most effective, aligning with the physical reality that light intensity is a primary error source in remote sensing.
Conclusion and Future Work
While the 10-point accuracy boost is a significant step forward, the method requires near-perfect pixel-to-pixel alignment between acquisitions. Future research is needed to determine if these learned features remain stable across different geographic sites with varying soil types and phenological states. Despite the logistical challenge of requiring repeated overflights for training, this approach offers a powerful tool for mitigating the impact of abiotic factors on botanical identification.
Summary updated