Chlorophyll content estimation in radiata pine using hyperspectral imagery: A comparison between empirical models, scaling-up algorithms, and radiative transfer inversions

Abstract

Radiata pine (Pinus radiata D. Don) is a widely planted tree species. Fertilizers, especially those containing leaf nitrogen (N) and phosphorous (P), are essential for maximizing growth. Nutrient deficiencies and excessive fertilization can limit growth, so monitoring is crucial. Leaf pigments such as chlorophyll a+b (C_a+b) can be used to assess plant nutrition, specifically leaf N. Remote sensing approaches can be used to monitor forest condition by estimating C_a+b content as a proxy for leaf N. Conventional methods for C_a+b estimation are based on empirical relationships using sensitive spectral indices or inversions of Radiative Transfer Models (RTMs). However, the structural complexity of tree crowns composed of multiple layers of clumped leaves/needles and background and shadow effects challenge the use of the indices proposed for both leaf C_a+b and leaf nitrogen assessment. This study compares the accuracy of methods for C_a+b estimation in radiata pine using hyperspectral data collected from a greenhouse experiment over the growing season and from a field trial representing a stand with a complex structure. The methods used to predict needle C_a+b from tree-crown spectra included: 1) empirical relationships between C_a+b measurements and hyperspectral indices; 2) scaling-up of hyperspectral index-based C_a+b predictive relationships through RTM simulations; and 3) RTM inversions of C_a+b content. These methods were tested over two different segmentation strategies, including sunlit-vegetation and full-crown spectra, to assess the effects of the increased structural complexity.

Predictions of C_a+b from the greenhouse experiment were generally higher for empirical models that used TCARI/OSAVI (Transformed Chlorophyll Absorption in Reflectance Index normalized by the Optimized Soil-Adjusted Vegetation Index) and CI (Chlorophyll index) hyperspectral indices when looking at full-crown rather than sunlit-vegetation pixels. RMSE measurements for full-crown models based on TCARI/OSAVI and CI across the three seasons ranged between 3.60 and 8.71 µg/cm² and between 3.70 and 7.86 µg/cm², respectively. Using the scaling-up methodology, the TCARI-OSAVI-derived models were more stable across different methods of pixel extraction than the CI-derived models were, showing the smallest variations across measurement dates. Predictions of C_a+b in the field trial showed that PRO4SAIL2, which combines the PROSPECT-D model with the 4SAIL2 model and accounts for clumping and a more complex tree structure, was more accurate than PRO4SAIL, which couples PROSPECT-D with the original 4SAIL model, across both crown segmentation methods. Using PRO4SAIL2, predictions were more accurate for the full-crown spectra (R² = 0.82; RMSE = 3.35 µg/cm²) than for the sunlit-vegetation pixels (R² = 0.69; RMSE = 4.03 µg/cm²).These results obtained in greenhouse and field trials reinforce the superior performance with forest species of simpler RTM strategies like 4SAIL2, as compared to more complex 3-D approximations, to accurately characterize pine tree-crown traits by integrating multi-layer and clumping effects.