Sensor‐independent analysis method for hyperspectral data based on the pattern decomposition method

Lifu Zhang, Shinobu Furumi, Keiichi Muramatsu, N. Fujiwara, M. Daigo, Liangpei ZhangInternational Journal of Remote Sensing 2006 · DOI

TL;DR

This paper introduces a "Universal Pattern Decomposition Method" (UPDM) to analyze hyperspectral satellite data in a way that's independent of the specific sensor used. By standardizing spectral patterns and employing a modified decomposition technique, the UPDM allows for consistent analysis results across different satellite platforms, overcoming a major hurdle in comparing and integrating remote sensing data.

What Problem Does This Paper Address?

Satellite remote sensing is crucial for understanding Earth's surface, but a significant challenge arises from the diversity of satellite sensors. Each sensor has unique characteristics, such as the number of spectral bands and their specific wavelengths. This variability leads to sensor-dependent analysis results, making it difficult to compare data from different satellites or to build comprehensive, long-term datasets. For instance, analyzing land cover or estimating global primary production using data from, say, Landsat and then MODIS, would yield different outcomes, hindering large-scale studies and consistent monitoring. The paper aims to solve this by developing a method that produces "sensor-independent" results.

Methods

The core of the paper is the development of a Universal Pattern Decomposition Method (UPDM), a modification of the existing Pattern Decomposition Method (PDM).

  • Conventional PDM: This method decomposes a pixel's reflectance spectrum into a linear combination of standard spectral patterns (water, vegetation, soil). The coefficients represent the proportion of each component. However, these standard patterns were sensor-specific, leading to sensor-dependent coefficients.
  • UPDM:
    • Sensor-Independent Standard Patterns: The UPDM defines a set of continuous, "universal" standard spectral patterns (water, vegetation, soil) derived from ground-measured spectra across a broad wavelength range (350-2500 nm), excluding atmospheric absorption regions. These patterns are normalized using integration over the entire wavelength range, making them independent of any specific sensor's band structure.
  • Supplementary Pattern: To improve analysis, especially for nuanced features like vegetation changes, a supplementary "yellow-leaf" pattern is introduced. This pattern is designed to be orthogonal to the main three, preventing multicollinearity issues.
  • Decomposition Coefficients: The decomposition coefficients are calculated using a least-squares method. Crucially, the non-negative constraint often applied in spectral mixing analysis is removed in the UPDM's general interpretation, allowing for a more flexible multivariate analysis.
  • Normalization: The method uses a consistent normalization approach for all sensors, ensuring that the derived coefficients are comparable.

Datasets

The study utilizes ground-measured samples for developing and validating the method.

  • Ground-Measured Samples: Approximately 600 samples were collected, including green-leaf, yellow-leaf, dead-leaf, soil, water, and concrete. These were used to define the standard spectral patterns and to simulate data for various sensors.
  • Simulated Sensor Data: Data from several satellite sensors were simulated using the ground-measured samples:
  • Landsat/MSS
    • ALOS/AVNIR-2
    • Landsat/ETM+
    • Terra/MODIS
    • ADEOS-II/GLI
  • Model Sensors: The "virtual sensors" MODEL and CONTINUE were also used. The CONTINUE sensor, with 92 bands, served as a reference for comparison.

Key Insights and Ideas

  • The Power of Universal Standards: By defining a fixed, continuous set of standard spectral patterns that span a wide spectral range, the UPDM effectively decouples the analysis from the specific band configurations of individual sensors. This is the core innovation for achieving sensor independence.
  • Orthogonal Supplementary Patterns Mitigate Issues: The introduction of a "residual yellow-leaf" pattern, designed to be orthogonal to the primary components, is a clever way to incorporate more detailed spectral information (like vegetation stress) without introducing statistical noise or correlation problems.
  • Multivariate Analysis is More General: The authors argue that viewing PDM as a multivariate analysis, rather than strictly spectral mixing, is more appropriate, especially with supplementary patterns. This perspective justifies relaxing non-negativity constraints, leading to a more robust method.
  • Sensor-Dependent Errors Still Exist, But Are Quantified: While the UPDM aims for sensor independence in coefficients, the estimation errors for these coefficients are shown to vary by sensor. This provides valuable information about the reliability of results from different platforms.

Key Results

  • Sensor-Independent Coefficients: The UPDM demonstrated that the decomposition coefficients obtained for the same sample were nearly identical across different sensors, confirming sensor independence. The paper states that coefficients were within 0.77% of the root mean square (rms) for most sensors.
  • Reconstruction Accuracy: The reconstructed spectra for ETM+, MODIS, and MODEL sensors closely matched the original ground-measured CONTINUE data. AVNIR-2 and MSS showed some discrepancies, particularly at longer wavelengths.
  • Error Analysis: The reduced $\chi^2$ values, indicating fitting error, were generally low. For sensors like Landsat/ETM+, Terra/MODIS, and ADEOS-II/GLI, the values were around 0.00070, with a fitting error of about 2.6% per degree of freedom. Four-band sensors (MSS, AVNIR-2) showed larger errors.
  • Correlation of Coefficients: Linear regression analysis showed a strong correlation (close to 1) between the UPDM coefficients derived from different sensors and the CONTINUE sensor, further supporting the sensor-independent nature of the method.

Limitations and Future Work

  • Four-Band Sensor Discrepancies: The paper notes that four-band sensors (Landsat/MSS and ALOS/AVNIR-2) exhibited larger estimation errors compared to sensors with more bands. This suggests that the UPDM's performance might be more sensitive to the spectral resolution and number of bands in sparser spectral configurations.
  • Further Validation: The authors plan to apply the UPDM to data from Wuhan, China, and Kii Peninsula, Japan, using ETM+, MODIS, and GLI satellites to further validate its performance.
  • Sensor-Independent Vegetation Index and Classification: Future research will focus on defining a sensor-independent vegetation index based on UPDM coefficients and developing sensor-independent ground classification and NPP estimation methods.

Summary updated

Abstract

Abstract This paper describes a modified pattern decomposition method with a supplementary pattern. The proposed approach can be regarded either as a type of spectral mixing analysis or as a kind of multivariate analysis; the later explanation is more suitable considering the presence of the additional supplementary patterns. The sensor‐independent method developed herein uses the same normalized spectral patterns for any sensor: fixed multi‐band (1260 bands) spectra serve as the universal standard spectral patterns. The resulting pattern decomposition coefficients showed sensor independence. That is, regardless of sensor, the three coefficients had nearly the same values for the same samples. The estimation errors for pattern decomposition coefficients depended on the sensor used. The estimation errors for Landsat/MSS and ALOS/AVNIR‐2 were larger than those of Landsat/TM (ETM+), Terra/MODIS and ADEOS‐II/GLI. The latter three sensors had negligibly small errors. Acknowledgments The authors thank A. Ono of Otemae University for encouraging us to write this paper. This work was supported under the ADEOS‐II/GLI Project of the Japan Aerospace Exploration Agency (JAXA), the Academic Frontier Promotion Project of the Ministry of Education, Science, Sports and Culture of Japan, and the 973 Project of the People's Republic of China (project no. 2003CB415205).
BibTeX
@article{zhang2006sensorindependent,
  author = {Lifu Zhang and Shinobu Furumi and Keiichi Muramatsu and N. Fujiwara and M. Daigo and Liangpei Zhang and Liangpei Zhang and Liangpei Zhang},
  title = {{Sensor‐independent analysis method for hyperspectral data based on the pattern decomposition method}},
  journal = {International Journal of Remote Sensing},
  year = {2006},
  doi = {10.1080/01431160600702640},
}
Figure