Reconstruction of species ecological characteristics from chlorophyll a fluorescence signatures

Predictive models were fitted using the 17 OJIP parameters as predictors and each ecological characteristic as a separate response variable to assess whether species ecological characteristics could be reconstructed from their chlorophyll a fluorescence signatures. The analyses included light, temperature, continentality, edaphic moisture, topographic moisture, acidity, nutrient availability, salinity, naturalness, hemeroby, and the two isometric log-ratio coordinates representing CSR strategy (CSR ILR1 and CSR ILR2). Four regression approaches representing different model structures were compared: Ridge regression, Elastic Net, partial least squares (PLS) regression, and Random Forest. Model performance was evaluated using repeated out-of-sample validation. In each of 100 repetitions, species were randomly partitioned into a 75% training set and a 25% test set. Models were fitted exclusively to the training species, and predictive performance was evaluated on the withheld species. The same training–test partition was used for all models within each repetition, allowing their performance to be compared under identical sampling conditions. Predictive performance was quantified using the out-of-sample coefficient of determination (predictive R²), calculated as , together with the root mean squared error (RMSE), mean absolute error (MAE), and the correlation between observed and predicted values. The distribution of predictive R² across the 100 independent partitions was used to characterize both the magnitude and stability of predictive performance. The contribution of individual OJIP parameters to ecological reconstruction was additionally evaluated using two complementary approaches. For Random Forest models, permutation importance was calculated within each repeated training partition. For PLS models, variable importance in projection (VIP) was calculated for each OJIP parameter. Predictor importance was summarized across the 100 repetitions using the median and the 2.5th and 97.5th percentiles. The reproducibility of predictor rankings between Random Forest and PLS was quantified separately for each ecological characteristic using Spearman’s rank correlation.