Geodaysit 2023

Federico Carotenuto


Investigating PLSR and RF for retrieving wheat crop traits in a field phenotyping experiment using full-range hyperspectral data: performance assessment and modelling interpretation
Ramin Heidarian Dehkordi, Mirco Boschetti, Gabriele Candiani, Federico Carotenuto, Carla Cesaraccio, Andrea Genangeli, Beniamino Gioli, Donato Cillis, Marina_Ranghetti

Crop traits monitoring is a fundamental step for controlling crop productivity in the context of precision agriculture and field phenotyping. At present, the usage of hyperspectral data in machine learning regression algorithms (MLRAs) has attracted increasing attention to alleviate the challenges associated with traditional crop trait measurements. However, the performance assessment of such hyperspectral-based MLRA models for crop trait retrievals with respect to the well-known natural variations in either structural or biochemical crop properties remains largely elusive. As such, this experiment was set up to assess whether full-range hyperspectral data, acquired by a handheld spectrometer (Spectral Evolution; 350 – 2500 nm), as inputs to partial least squares regression (PLSR) and random forest (RF) models are capable of modeling different wheat crop traits at the canopy level. The examined crop traits were leaf area index (LAI), canopy water content (CWC), canopy chlorophyll content (CCC), and canopy nitrogen content (CNC). This approach allowed us, as an overarching objective, to compare the performance of the two aforementioned MLRA models while also focusing on the physical interpretation of the modelling results for each particular crop trait.
Overall, PLSR provided remarkably higher accuracy, tested with a cross-validation strategy, as compared to RF for all the crop traits. More precisely, PLSR denoted R2 (resp. nRMSE%) values of 0.72 (11.97), 0.77 (10.89), 0.70 (14.61), and 0.74 (14.38) for LAI, CWC, CCC, and CNC, respectively. All PLSR models indicated robust prediction capability with RPD values greater than 1.4, and amongst them, CWC was found to have excellent prediction performance with an RPD higher than 2. However, RF yielded less predictive models with R2 (resp. nRMSE%) values of 0.59 (14.59), 0.42 (17.42), 0.50 (18.86), and 0.42 (21.41) for LAI, CWC, CCC, and CNC, respectively. RF models for LAI and CCC showed good prediction capabilities (RPD > 1.4), whilst RF models of neither CWC nor CNC were reliable (RPD < 1.4).
In general, RF band importance and PLSR regression coefficient results revealed physically- meaningful and consistent patterns for each specific crop trait. Specific wavelengths at SWIR (1716-1745 nm) and NIR (1057-1120 nm), Green, and the Red-Edge bands respectively showed the highest importance for LAI retrieval. Water absorption regions around 910 nm and 1200 nm as well as the Red-Edge and Visible parts were of higher importance for the retrieval of CWC. The best-performing bands were situated in Red-Edge and Green spectral channels for CCC retrieval. SWIR spectral regions between 1600-1800 nm and 2100-2300 nm appeared to be important (in particular with respect to the other traits) alongside the Red-Edge part of the spectrum to retrieve CNC.
We demonstrated that full-range hyperspectral data in combination with MLRA algorithms can provide accurate estimates of wheat crop traits at the canopy level. The success of utilizing hyperspectral data in MLRA algorithms was further highlighted by the physically-meaningful modelling performances in accordance with the subtle structural and biochemical crop properties. Our results suggest that such spectroscopic hyperspectral-based MLRA approaches could be a powerful tool to accurately monitor crop status throughout the cropping season to improve high-throughput phenotyping activities and to further aid precision agricultural practices.

AIT Contribution
Sala Videoconferenza @ PoliBa