Transfer Learning for Dead Fuel Moisture Prediction Using Time-Warped Recurrent Neural Networks
2026
Read Transfer Learning for Dead Fuel Moisture Prediction Using Time-Warped Recurrent Neural Networks (2026) 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-2605-08379.txt.
Introduction
Jonathon Hirschi, Jan Mandel, and Adam Kochanski adapt a fuel moisture model across fuel types by warping time. Dead fuel moisture drives wildfire behavior models. Fuels are classed by their lag time. The classes run from one-hour twigs to thousand-hour logs. Sensors report the ten-hour class richly. The other classes are observed sparsely.
Their method rescales the learned dynamics of a recurrent network with an LSTM layer. It stretches or compresses time. A model trained on ten-hour fuels can then predict the slower and faster classes. Validation uses a landmark Oklahoma field study. That study calibrated the standard Nelson model. The test asks whether time warping is a sound bridge between fuel classes.
Catalog record
| YEAR | 2026 |
|---|---|
| COLLECTION | PAPERS |
| FORMAT | TXT |
| FILE SIZE | 62 KB |
| LICENSE | http://creativecommons.org/licenses/by/4.0/ |
| RIGHTS | CC BY |
| SOURCE | arXiv |
Source: https://arxiv.org/abs/2605.08379
Browse the full library · About the Tower of Babel
Full text
The opening of the work, from the library's hosted text:
Transfer Learning for Dead Fuel Moisture Prediction Using Time-Warped Recurrent Neural Networks Abstract This paper proposes a time-warping transfer learning method, a technique for temporally rescaling the learned dynamics of a recurrent neural network (RNN) with a Long Short-Term Memory (LSTM) layer to enable task transfer across fuel moisture classes. Fuel moisture content (FMC) is divided into idealized classes based on characteristic lag time. Large quantities of real-time data are available for 10h fuels from sensors on weather stations, but observations of other fuel classes are sparse in space and time. We use transfer learning to adapt an RNN pretrained on 10h FMC to predict FMC for 1h, 100h, and 1000h fuels. We validate this method using data from a landmark field study conducted in Oklahoma that was used to calibrate the state-of-the-art Nelson fuel moisture model. † † authors: Jonathon Hirschi, \aff a Jan Mandel \aff a Adam Kochanski, \aff b \statement Dead fuel moisture data are widely available for 10-hour fuels but scarce for 1, 100, and 1000-hour fuel classes, making standard machine learning approaches difficult to train reliably for the sparsely observed classes. We develop a transfer learning method based on time-warping to adapt an RNN trained on 10-hour fuel data to predict 1-, 100-, and 1000-hour fuel moisture, where target data are scarce, while modifying only a small subset of model parameters. To our knowledge, this is the first application of transfer learning to fuel moisture prediction and the first published ML study to target 100-hour and 1000-hour fuel moisture. It shows that transfer learning can extend data-driven prediction beyond the fuel class with abundant sensor observations. 1 Introduction Fuel moisture content (FMC) is a measure of the water content of fuels that is important for understanding wildfire danger and active fire behavior, e.g. ( 35 ) . It is expressed as the ratio of the mass of water in the fuel to the dry mass of the fuel, typically given as a percentage. Live and dead fuels are typically modeled separately because they are governed by different processes. Dead fuel moisture content is controlled primarily by external environmental conditions such as temperature, relative humidity, solar radiation, and wind, through moisture exchange with the atmosphere (absorption and desorption). In contrast, live fuel moisture content is regulated by internal physiological processes within the plant, including water uptake from soil, storage, and transpiration ( 35 ; 25 ) . This paper focuses on predicting dead FMC as a response to weather. Operational wildfire management tools and wildfire research commonly use the idealized 1-hour, 10-hour, 100-hour, and 1000-hour fuel classes—denoted as FM1, FM10, FM100, and FM1000—which correspond to characteristic dead fuel size classes of approximately 0–0.25 inches (0–0.6 cm), 0.25–1 inch (0.6–2.5 cm), 1–3 inches (2.5–7.6 cm), and 3–8 inches (7.6–20 cm) in diameter, respectively. Fuel classes are defined by their time lag, the characteristic time required for a fuel’s moisture content to change by 1-e^{-1} (about 63%) of the difference between its initial moisture content and its new equilibrium value in response to a change in atmospheric conditions ( 3 ; 10 ) . Large quantities of data exist for fuels with a lag time of 10 hours, but observations for other lag times are sparse. A large network of Remote Automatic Weather Stations (RAWS) provides data from standardized sensors for FM10 ( 33 ) . The FMC sensor consists of a 1/2-inch pine dowel with a probe that estimates the moisture content in the woody interior ( 4 ; 11 ) . RAWS across the continental United States (CONUS) provide real-time data for FM10 ( 44 ) . Observations of the other fuel classes are sparse in space and time, as automated sensors are only available for FM10. For all fuel classes, FMC can be measured gravimetrically by weighing fuel samples before and after oven drying. Gravimetric measurement is accurate and direct, but it is slow and labor-intensive, so the available data is much sparser. These observations include field samples collected by wildfire management agencies and distributed through the Fire Environment Mapping System (FEMS) ( 46 ) , which can be accessed through the National Fuel Moisture Database (NFMDB) ( 52 ) . Controlled field studies provide higher-temporal-resolution observations, with fuel dimensions standardized to the conventional time lag fuel classes. A particularly important controlled field study is that of 5 , which measured FMC for multiple fuel classes up to twice daily from spring 1996 through 1997. From now on we refer to that study as the Oklahoma field study. Those data were used to calibrate the 34 fuel moisture model for use within the National Fire Danger Rating System (NFDRS), as well as other process-based models ( 47 ) , and we use them here because they are historically important and have higher temporal resolution than the FEMS field studies. Modern machine learning (ML) models use large amounts of FM10 data to learn spatial and temporal patterns and provide accurate forecasts. In contrast, ML approaches to predicting fine fuel moisture (FM1) and FMC for larger fuels can rely only on the use of relatively small amounts of data. Traditional physics-based models of FMC continue to be the standard approach to predicting FM1, FM100, and FM1000. This paper attempts to fill a gap in the FMC modeling literature by leveraging the vast network of FM10 sensors to improve predictions for other fuel classes. We describe a new approach to transfer learning based on modifying the learned temporal dynamics of a recurrent neural network (RNN), which we refer to as time-warping. We use a pretrained RNN that has been developed to predict FM10, then apply the time-warping method to predict FM1, FM100, and FM1000. We compare this method to several standard transfer learning methods and show that we achieve similar levels of accuracy despite modifying only a small fraction of the learned parameters relative to the other methods. This paper is based on the thesis 17 , which builds on forecasting of FMC by RNNs with weather inputs, introduced by 29 and developed into a spatiotemporal product by 15 . 1.1 Modeling Approaches for Dead Fuel Moisture Content Mathematical models of FMC can generally be divided into three categories: physical process-based models, time lag models, and ML models. Process-based models and modern ML approaches still use fuel classes to define output variables, as that remains the standard for data collection ( 54 ) . In some areas of the literature, researchers study the moisture content of fine fuels, which includes dead bark, twigs, and leaf litter. FM1 has a specific time lag definition and has been studied with standardized wooden dowels, while fine fuel moisture content is used more generally in wildfire science. 1.1.1 Physical Process-Based Models Physical process-based methods attempt to directly represent the transfer of energy and moisture between a wood stick and its environment. 34 developed a model for FMC as part of the NFDRS that solves coupled one-dimensional heat and moisture transport equations in the radial coordinate of a cylindrical wood stick. The model receives weather inputs of air temperature, RH, accumulated rainfall, and downward shortwave solar radiation. 47 modeled FMC using two homogeneous layers and added the effect of the wind to some