Deep Learning for Reconstructing Historical Precipitation from Growth Rings in Baobab Trees

This study explores the application of deep learning for reconstructing historical precipitation patterns using growth rings of Adansonia digitata L. Malvaceae (Baobab) trees in Recife, Brazil. Baobabs, due to their longevity and adaptability to arid conditions, provide valuable dendrochronological records suitable for climate reconstruction in regions with limited meteorological data. High-frequency densitometry was used to measure annual wood increment and density from 482 growth rings across 11 Baobab specimens. These features were statistically correlated with official precipitation records from 1979 to 2024, forming the basis for a multiple regression model. A modified LeNet-5 convolutional neural network was then trained to predict historical precipitation from growth ring images, achieving over 99% classification accuracy using only 25 augmented images. The strong inverse correlation between wood density and precipitation and the direct correlation between increment and precipitation confirm the utility of Baobab rings as bioindicators. This integration of dendroclimatology and deep learning presents a robust, non-destructive method for reconstructing past climate, offering valuable insights for ecological monitoring and land management in data-scarce tropical and semi-arid regions.