Assessing Machine Learning and Deep Learning Approaches for Wood Volume Prediction in Forest Plantations Using Hyperspectral Imagery: Limitations and Challenges

Abstract

This study evaluates four machine learning algorithms, Random Forest Regressor, Support Vector Regressor (SVR), Multilayer Perceptron, and a 1-D Convolutional Neural Network, for estimating Pinus taeda L. wood volume using EnMAP hyperspectral data. Spectral preprocessing included PCA dimensionality reduction, with permutation-based importance analysis identifying the NIR and SWIR regions as the most influential for component generation. Models were trained and validated using 153 plots in southern Brazil under two validation frameworks: nonspatial autocorrelation (NSAC) using random cross validation, and spatial autocorrelation (SAC) using clustering–based spatial cross validation. Moran’s I analysis confirmed significant SAC in the volume data. Under NSAC, no statistically significant differences were observed among the four algorithms. In contrast, under SAC, none of the models achieved satisfactory performance, consistently yielding R2 < 0 and substantially high errors, indicating complete failure to generalize to spatially distinct areas. This stark discrepancy underscores how SAC inflates performance metrics under nonspatial validation. Despite this, SVR was selected for the mapping due to its performance and stability across different training and test set proportions, achieving R2 = 0.617, RMSE = 72.59 m3 ha−1, and MAE = 40.18 m3 ha−1 on test set at the selected 80/20 proportion. The results reveal that when spatial dependence is properly accounted for, spectral–based models exhibit substantially limited predictive power for wood volume estimation. While the findings partially confirm the feasibility of integrating EnMAP data and machine learning for forest monitoring, the generated maps should be interpreted as exploratory decision–support tools rather than operational predictions, given the critical limitations imposed by SAC and the failure of models to generalize beyond the training domain.