Inferring processes within dynamic forest models using hybrid modeling

Maximilian PichlerTheoretical EcologyUniversity of Regensburg93053 RegensburgGermany [email protected], Yannek KäberBiometry and Environmental System AnalysisUniversity of Freiburg79098 FreiburgGermany [email protected] · 2025

Read Inferring processes within dynamic forest models using hybrid modeling by Maximilian PichlerTheoretical EcologyUniversity of Regensburg93053 RegensburgGermany [email protected], Yannek KäberBiometry and Environmental System AnalysisUniversity of Freiburg79098 FreiburgGermany [email protected] (2025) in the Tower of Babel, a free online library of public-domain works preserved in clean, searchable text. No sign-up. The full text is served as a plain-text file at a predictable URL, where text is present: https://antarctic-labs.com/tower-of-babel/arxiv-2508-01228.txt.

Introduction

Forest models that simulate tree growth from fixed formulas can go wrong when nature follows a different curve. Maximilian Pichler and Yannek Käber introduce FINN, a hybrid that swaps parts of a forest gap model for neural networks. The whole system trains in one step, so the data itself shapes the process.

On a famous 50-hectare plot in Panama, replacing the growth process with a network improved predictions and the forest's succession path. Better still, the network had learned a plausible, improved growth curve, which the authors extracted with explanation tools and could read like an ecological formula. They offer the approach as a way to forecast forests under climates with no modern precedent. The paper appears in 2025.

Catalog record

CREATORMaximilian PichlerTheoretical EcologyUniversity of Regensburg93053 RegensburgGermany [email protected], Yannek KäberBiometry and Environmental System AnalysisUniversity of Freiburg79098 FreiburgGermany [email protected]
YEAR2025
COLLECTIONPAPERS
FORMATTXT
FILE SIZE62 KB
VERSION1.0
LICENSEhttp://creativecommons.org/licenses/by/4.0/
RIGHTSCC BY
SOURCEarXiv

Source: https://arxiv.org/abs/2508-01228

Browse the full library · About the Tower of Babel

Full text

The opening of the work, from the library's hosted text:

Inferring processes within dynamic forest models using hybrid modeling Maximilian PichlerTheoretical EcologyUniversity of Regensburg93053 RegensburgGermany [email protected], Yannek KäberBiometry and Environmental System AnalysisUniversity of Freiburg79098 FreiburgGermany [email protected] Abstract Modeling forest dynamics under novel climatic conditions requires a careful balance between process-based understanding and empirical flexibility. Dynamic Vegetation Models (DVM) represent ecological processes mechanistically, but their performance is prone to misspecified assumptions about functional forms. Inferring the structure of these processes and their functional forms correctly from data remains a major challenge because current approaches, such as plug-in estimators, have proven ineffective. We introduce Forest Informed Neural Networks (FINN), a hybrid modeling approach that combines a forest gap model with deep neural networks (DNN). FINN replaces processes with DNNs, which are then calibrated alongside the other mechanistic components in one unified step. In a case study on the Barro Colorado Island 50-ha plot we demonstrate that replacing the growth process with a DNN improves predictive performance and succession trajectories compared to a mechanistic version of FINN. Furthermore, we discovered that the DNN learned an ecologically plausible, improved functional form of the growth process, which we extracted from the DNN using explainable AI. In conclusion, our new hybrid modeling approach offers a versatile opportunity to infer forest dynamics from data and to improve forecasts of ecosystem trajectories under unprecedented environmental change. * * footnotetext: These authors contributed equally to this work K eywords Forest Ecology \cdot Dynamic Vegetation Models \cdot Process Models \cdot Hybrid Modeling \cdot Deep Learning Main Forest ecosystems play a central role in Earth’s carbon and water cycle, and in climate regulation, acting as key components in both local and global ecological systems. Understanding and accurately predicting how forests respond to environmental changes is crucial for addressing climate change, conservation, and resource management ( Forzieri et al. 2022 ; McDowell et al. 2020 ) . To this end, we rely on Dynamic Vegetation Models (DVM) that serve as repositories of integrated ecosystem knowledge ( Botkin et al. 1972 ) and simulate mechanistic processes, from small scale such as photosynthesis to large scale such as tree community composition across space and time ( Bugmann and Seidl 2022 ; Moorcroft et al. 2001 ; Seidl et al. 2012 ; Smith et al. 2001 ) . However, the reliability of DVMs depends strongly on how well they represent the underlying processes ( Beven 2006 ; Cailleret et al. 2020 ; Käber et al. 2024 ; Van Oijen et al. 2005 ) . Misspecified processes cause structural model bias, which decreases the accuracy of the DVM, especially when predicting outside the observed data ( Oberpriller et al. 2021 ) . But choosing the correct functional form of a process has proven to be non-trivial. Existing functional forms are often educated guesses based on isolated empirical correlations that do not consider the broader context (i.e. other processes). Inverse calibration methods (manual or systematic) ( Hartig et al. 2011 ; Hartig et al. 2012 ) are usually limited by the practical infeasibility of evaluating a large number of candidate models ( Delpierre et al. 2019 ; Käber et al. 2024 , e.g.) . Even worse, structural bias often goes undetected because other processes can compensate for it, thereby obscuring its source and recovering predictive accuracy (i.e., predicting the right things for the wrong reasons) ( Cameron et al. 2022 ; Oberpriller et al. 2021 ) . Therefore, an approach that can automatically infer the functional form of a process empirically from data would be much more appealing. However, this requires a highly flexible empirical model, as well as joint calibration with the other mechanistic components of a DVM. To address this challenge, the empirical model of the process must be sufficiently flexible to approximate any realistic functional form. Recent advances in deep learning (DL), particularly deep neural networks (DNN), offer a promising solution ( Pichler and Hartig 2023a ) . DNNs can approximate any nonlinear function and automatically capture complex interactions between variables ( Pichler et al. 2020 , e.g.) , offering advantages over more rigid alternatives and reducing the need for cumbersome model selection or manual formulation. However, their high flexibility comes at the cost of non-interpretability ( Pichler and Hartig 2023b ) , which is at odds with our interest in understanding the functional form learned by the DNN. Tools of explainable Artificial Intelligence (xAI) were developed to address this problem ( Ryo et al. 2021 ) . Explainable AI tools analyze how model predictions change with controlled permutations of the inputs, analogous to sensitivity analysis ( Murdoch et al. 2019 ; Malchow and Hartig 2024 ) . Together, the flexibility of DNNs and the post-hoc interpretability of xAI create a synergy that can advance scientific discovery ( Davies et al. 2021 , e.g.) , which could include the derivation of forest processes. While flexibility addresses the functional form of a misspecified process, it has been shown that joint calibration of all components (mechanistic or not) improves the reliability of DVM. Calibrating processes externally, without the same context (i.e., different inputs and not in the presence of the other processes), can yield misleading inferences, similar to spurious correlations in statistics ( Wood 2001 ) . Even when sharing the same input-output structure, plug-in models of processes often perform worse than target processes calibrated jointly with other DVM processes ( Hülsmann et al. 2018 ; González et al. 2016 ; Cailleret et al. 2020 , e.g.) . This is likely because DVMs are designed to produce emergent patterns of forest dynamics ( Moorcroft et al. 2001 ) where process-level observations are interrelated over time and with each other, which is difficult to replicate with external independent models. Another reason could be that, since our understanding of nature will never be perfect, structural model bias in DVM is probably unavoidable. This causes the processes to compensate for each other during joint calibration ( Oberpriller et al. 2021 ) , resulting in better performance than external plug-in models of processes ( Cailleret et al. 2020 , see also) . To meet both criteria, we propose Forest Informed Neural Networks (FINN) (Fig. 1 ), a DVM that replaces selected processes (e.g., growth, mortality, and regeneration) with DNNs (Fig. 1 ). Unlike other hybrid modeling approaches with pre-calibrated plug-in models ( Chen et al. 2023 ; Reichstein et al. 2019 ) , FINN calibrates the replaced processes jointly with the remaining mechanistic processes end-to-end. This differs from process-informed neural networks, where process-based constraints adjust the model outputs ( Wesselkamp et al. 2024 ) ; in FINN the replaced processes are learned within and constrained by the surrounding DVM scaffold, which helps calibrate the DNN to predict emergent patterns of forest dynamics (Fig. 4 ). Consequently, FINN can learn the functional form of processes from stand- and population-level without direct process-level observations. After calibration, processes approximated by a DNN can be explained by xAI (Fig. 5 ). Consequently, FINN can learn the most suitable functional form for each process during a single, unified calibration step. This effectively combines classic inverse calibration and modern DL methods ( Boussange et al. 2024 ; Sapienza et al. 2024 , see also) . Figure 1: Overview of the Forest-Informed-Neural-Network (FINN) Model. FINN provides a

Read the complete text