Retrieving plant secondary metabolites across biomes using remotely sensed data and radiative transfer models.

Abstract

Secondary metabolites are diverse defence organic chemical compounds produced by plants against biotic and abiotic environmental stresses. Examples of secondary metabolites include terpenes, phenolic and nitrogenous compounds. Because these metabolites absorb electromagnetic energy in specific wavelengths within the 400 to 2500 nm spectral range, they can be detected non-destructively using remote sensing. Methods for quantifying plant secondary metabolites are generally divided into empirical and physical approaches. Empirical methods rely on finding the statistical relationship between plant attributes and remote sensing. Although they have been praised for producing desirable results over the past decades, they are trained and tested on a specific geography or dataset, sensor-specific, hence tend to lack generalization and transferability. On the other hand, physical models, such as radiative transfer models (RTMs) simulate the interactions of light with plant canopies, making them more robust across different seasons, locations, and sensor types. Despite their potential, most RTM applications focus on retrieving more common plant attributes like Leaf Area Index (LAI) and chlorophyll, particularly in agricultural, grassland, and forest landscapes. There is a notable gap in the literature regarding the use of RTMs to directly estimate secondary metabolites such as flavonoids and carotenoids, especially in ecosystems rich in phenolic compounds like the fynbos biome. Hence, this study aimed to address this gap by exploring how remotely sensed data and RTMs can be used to retrieve plant secondary metabolites across various biomes. The first objective examined the application of RTMs on different terrestrial ecosystems for estimating both biophysical and biochemical properties. The review highlights that most studies have employed 1D models such as PROSPECT and PROSAIL, while for more complex 3D models they have used DART and FLIGHT. Furthermore, most of the existing research has utilized sensors like MODIS and Landsat on parameters like LAI and chlorophyll through inversion techniques such as look-up tables (LUT). However, research efforts are concentrated mainly in the Global North, with limited studies conducted in regions like Africa. The review reveals that PROSPECT and PROSAIL have remained the dominant RTMs over the past twenty years.Building upon this, the second objective used the PROSPECT-5 model to estimate secondary metabolites such as carotenoid content using different wavelengths at the leaf level across three datasets (i.e., savanna, tropical forest, and a combination). Results showed that the green spectral region (510–580 nm) provided the best performance for estimating carotenoids, with R² values up to 0.90 and RMSE of 0.71 in the tropical forest dataset. Wavelengths selected by the Random Forest algorithm, particularly in the green, yellow, and red-edge regions, improved estimation accuracy, especially in the savanna dataset, with R² values exceeding 0.80. The visible spectrum (400–700 nm) also played a crucial role in the detection, emphasizing targeted wavelength selection for better model performance. These findings highlight the importance of focusing on specific wavelengths rather than the full spectrum. The third objective, also at the leaf level, restricted the wavelengths in the form of vegetation indices to study the relationship between fire and secondary metabolites like phenolics and flavonoids in the phenolic-rich fynbos biome. The data revealed that areas burned in2015, 2016, and 2021 had higher phenolic levels, compared to recently burned areas, indicating fire’s role in stimulating secondary metabolite production and ecosystem regeneration. Spectral techniques such as continuum removal identified significant signals around 1660 nm for phenols and 415 nm for flavonoids. Vegetation indices like FI415,710 and FI420,710 showed correlations (R² ≈ 0.55) with phenolic content, highlighting the importance of specific wavelengths. The fourth objective tested whether including secondary metabolites in the PROSAIL model improves canopy-level estimates of plant cover in the phenolic-rich fynbos region. The results demonstrated that the PROSAIL-D model, which accounts for secondary metabolites, outperformed the traditional PROSAIL-4 model, achieving R² values around 0.72–0.78 and lower NRMSE errors (approximately 3.86–3.99%) when using AVIRIS and field spectral data dataset. This emphasizes the importance of secondary metabolites in the PROSAIL-D model to effectively simulate areas rich in phenolic compounds. Finally, the fifth objective used the PROSAIL-D model to estimate flavonoid content using AVIRIS hyperspectral data. The vegetation index FI417,693 yielded the highest accuracy, with an R² of 0.63 and RMSE of 0.24 mg CAE/g. This demonstrated the potential of using narrow spectral bands and targeted indices for reliable retrieval of flavonoids, even within complex canopies. The study underscores that focusing on specific spectral features enhances the capacity of remote sensing to effectively quantify plant biochemical traits. In conclusion, this thesis demonstrates that RTMs, especially when incorporating secondary metabolites such as carotenoids and anthocyanins, are valuable tools for monitoring plant biochemical and biophysical properties across diverse ecosystems. Targeted spectral band selection and advanced modelling approaches can significantly improve the accuracy of remote sensing estimates. While most studies are concentrated in developed regions, expanding research into underrepresented areas like Africa is essential for global ecology.