Comprehension of carbon cycling, climate change, paleoecology, environmental reconstruction, and aquatic ecosystem health is essential for the environmental sciences, and one of the invaluable tools is the δ13C record of lake deposits. In this study, we propose a novel hybrid machine learning (ML) algorithm known as the ARAMT model, which combines two key components: a meta-classifier of additive regression (AR) and a base classifier of alternating model trees (AMT). The AR-AMT hybrid model significantly enhances the prediction of carbon isotopes (δ13C) by addressing limitations in existing methodologies. For the first time, this model is used to predict the spatial prediction of a stable isotope in Polish lakes. The δ13C for 30 Polish lakes is determined using the calcite (CaCO3) precipitated in the near-surface layers of the lakes (epilimnion). The chemical composition (Ca2+, HCO3−, Na, K, sulfates, fluorides, Cl, Mg, and P) and temperature of the surface water at a depth of 1 m is analyzed seasonally. The current approaches for predicting δ13C have demonstrated shortcomings in accuracy and precision. In this study, a random forest (RF), M5P, AMT, and Gaussian process (GP) are the four cutting-edge ML algorithms that are compared with the proposed hybrid model (ARAMT). In accordance with the results, the ARAMT hybrid model performed more effectively in predicting δ13C than the other benchmark ML methods (R2 = 0.9882, MAE = 0.456, and RMSE = 0.527), the others giving: AMT (R2 = 0.982, MAE = 0.558, and RMSE = 0.347), RF (R2 = 0.8014, MAE = 0.612, and RMSE = 0.550), M5P (R2 = 0.7508, MAE = 0.813, and RMSE = 0.701), and GP (R2 = 0.7315, MAE = 0.768, and RMSE = 0.683). Although the ARAMT hybrid doesn't directly preserve lake ecosystems, the enhanced accuracy of its δ13C predictions by providing a more detailed understanding of carbon cycling dynamics can indirectly inform and improve lake ecosystem health and management.
