Hyperspectral remote sensing and machine learning in the evaluation of the growth and quality of eucalyptus seedlings

Ensuring high-quality forest seedlings is fundamental for successful silviculture and reforestation initiatives. Conventional assessment methods for seedling quality are often destructive and labor-intensive, motivating the development of rapid and non-destructive techniques. The integration of machine learning (ML) with hyperspectral data can markedly advance the estimation of morphological and quality variables in forest seedlings, overcomes the limitations of traditional methods. This study aimed to evaluate the effectiveness of ML algorithms combined with hyperspectral data in predicting morphological variables and the quality of Eucalyptus deglupta Blume seedlings. Ninety samples were evaluated, providing an adequate basis for the development and validation of the models. However, further studies involving larger and more diverse datasets are needed to broaden the representativeness and generalizability of the results. Spectral variables (360 - 2350 nm) were measured using a hyperspectral spectroradiometer, along with morphological variables such as height, diameter, dry masses, and ratios among these traits. To estimate the morphological variables, the following machine learning algorithms were trained: Support vector machine (SVM), artificial neural networks (ANN), decision tree (DT), Zero R, M5P model tree, and random forest (RF). Two input combinations were evaluated during training: the wavelength provided by the sensor (WL) and data aggregated into 24 spectral bands (SB). Ten-fold cross-validation was adopted. Model performance was assessed using the correlation coefficient, mean absolute error, and root mean square error. The accuracy metrics were expressed in bar graphs containing error bars centered on the bars, allowing visualization of data variability, in addition to letters indicating statistical comparisons between means using the Scott-Knott test at a 5% probability level. Model performance varied with input type; the SVM model with full range inputs yielded the highest predictive accuracy overall. Use of the full the wavelength (WL) generally resulted in superior performance compared to aggregated bands (SB). Although the models showed good predictive performance on the evaluated dataset, generalization to other environmental conditions or seedling production systems still needs to be investigated in future studies, such as using a larger database with seedlings at different stages of development and employing deep machine learning to obtain more robust and generalized information. Combining ML algorithms with hyperspectral sensing represents approach for non-destructive estimation of forest seedling quality, offering efficiency and precision. Such integration can streamline nursery management and enhance large-scale forest plantation success.