Spectral band selection for vegetation properties retrieval using Gaussian processes regression (2012.08640v1)

Abstract: With current and upcoming imaging spectrometers, automated band analysis techniques are needed to enable efficient identification of most informative bands to facilitate optimized processing of spectral data into estimates of biophysical variables. This paper introduces an automated spectral band analysis tool (BAT) based on Gaussian processes regression (GPR) for the spectral analysis of vegetation properties. The GPR-BAT procedure sequentially backwards removes the least contributing band in the regression model for a given variable until only one band is kept. GPR-BAT is implemented within the framework of the free ARTMO's MLRA (machine learning regression algorithms) toolbox, which is dedicated to the transforming of optical remote sensing images into biophysical products. GPR-BAT allows (1) to identify the most informative bands in relating spectral data to a biophysical variable, and (2) to find the least number of bands that preserve optimized accurate predictions. This study concludes that a wise band selection of hyperspectral data is strictly required for optimal vegetation properties mapping.

• The paper demonstrates that optimizing spectral bands with GPR-BAT significantly enhances vegetation property retrieval, achieving high accuracy metrics such as an R² of up to 0.95.
• It employs a sequential backward band removal process that efficiently reduces redundant bands while maintaining model performance and interpretability.
• Experimental validation on field and HyMap datasets highlights the method's potential for precision agriculture and the development of tailored sensor systems.

Spectral Band Selection for Vegetation Properties Retrieval Using Gaussian Processes Regression

The paper presents an advanced method for optimizing the retrieval of vegetation biophysical properties from hyperspectral data by employing Gaussian processes regression (GPR) specifically designed for spectral band selection. The methodology is encapsulated in the GPR-BAT tool, which is integrated within the ARTMO framework—a toolbox that hosts a suite of machine learning regression algorithms (MLRAs). The primary aim of this research is to identify the most informative spectral bands and reduce their quantity while preserving high predictive accuracy for the estimation of vegetation characteristics such as leaf chlorophyll content (LCC), green leaf area index (gLAI), leaf area index (LAI), and canopy water content (CWC).

Methodological Approach

The GPR-BAT tool performs a sequential backward band removal process. It begins with the full spectrum and iteratively removes the least contributing band based on Gaussian processes regression model sensitivity until only one band remains. This operation is facilitated by an automated graphical user interface (GUI) that significantly minimizes user intervention, allowing for efficient identification of critical spectral bands that contribute to the regression model's accuracy.

Experimental Validation

The validation of GPR-BAT is carried out through two distinct hyperspectral datasets: a field dataset spanning 400-1100 nm with measurements on maize and soybean in Nebraska, US, and an airborne HyMap dataset covering 430-2490 nm acquired from various crops in Barrax, Spain. The regression models, optimized through band selection, demonstrated marked improvements in performance metrics compared to using the entire spectrum.

• Field Hyperspectral Dataset: For LCC retrieval, optimal accuracies were achieved with nine well-chosen bands, yielding a cross-validated coefficient of determination ($R^2_{CV}$) of 0.79 and a normalized root mean square error (NRMSE$))_CV)$ of 12.9%. Similarly, gLAI retrieval benefited from a selection of seven bands, with an $R^2_{CV}$ of 0.94 and NRMSE$))_CV)$ of 7.2%.
• HyMap Dataset: In LAI retrieval, four bands provided a $R^2_{CV}$ accuracy of 0.95 and an NRMSE$))_CV)$ of 6.5%. CWC retrieval, facilitated by six crucial bands, reached an $R^2_{CV}$ of 0.95 and NRMSE$))_CV)$ of 7.2%.

Theoretical and Practical Implications

The findings underscore the importance of reduced band configurations in remote sensing application models, both simplifying the model while enhancing interpretability and processing speed. Each targeted vegetation property substantially benefited from optimally located spectral bands, specifically those within the red edge and near-infrared regions which coalesce crucial absorption features. This not only aids in modeling efficiency but also supports the design and deployment of novel sensors with tailored spectral configurations.

Future Directions and Applications

The approach outlined by this paper offers promising avenues for the development of sensor systems with reduced spectral requirements, tailored to specific vegetation properties, which can be pivotal in enhancing real-time remote sensing applications such as precision agriculture. Future work could involve scaling this method to global satellite imaging platforms, enhancing operational capabilities in mapping vegetation dynamics at broader scales and adjusting for more complex biophysical characteristics under varying environmental conditions.

In conclusion, this paper establishes a robust framework for spectral band selection using Gaussian processes regression. It significantly contributes to improving the accuracy and efficiency of hyperspectral data analysis for vegetation properties retrieval, facilitating practical implementation in Earth observation technologies.