cs.LGOct 19, 2026

Bombadil: A Foundation Model for Plant Electrophysiology

Authors: James Kirk, Piotr Bednarski, Filip Twardy, Justin Mast, Lee von Kraus

Organizations: Electric Plant Co, electricplant.co

Read PDF
Ask AI
Cite

Copy citation

DOINot available for this paper

Abstract

Living plants, like animals, use electrical signals to carry information about their physiological state across their body, from roots to branches. These signals can be recorded, as they are in animals, from a pair of differential extracellular electrodes inserted into signal-conducting tissue. While machine-learning models have been trained to predict specific plant stress states from these signals, each of these models has been trained for only a single task and none have been designed to transfer across tasks, species, or environments. These approaches do not allow easy scaling across the wide range of species and environments that are relevant to agriculture, forestry, and horticulture for both commercial and consumer plant care. Here we present Bombadil, a foundation model for plant electrophysiology, pretrained without labels on over 4 million hours of electrical signals recorded from 100 plant species in ten distinct environments, including laboratories, greenhouses, commercial orchards, and natural forests. In this way, Bombadil is able to learn the fundamental features of plant electrical signals across a wide array of species and environments. Then, from those same electrical signals, Bombadil tracks changes in seven physiological indicators of plant stress, even among species and stressors with no labels in training.

Figures & tables

Figure 1: System overview. A pair of electrodes on a plant records the electrical signal, and one Transformer backbone learns to represent it. Two task-specific heads read that representation: the patch-reconstruction head, trained without labels during pretraining, and the stress-factor head, trained on paired ground-truth readings.
FactorGroupDefinition
Φ\Phi IIFluorescenceOperating efficiency of photosystem II (PSII)
Φ\Phi NPQFluorescenceLight dissipated by regulated non-photochemical processes
Φ\Phi NOFluorescenceLight dissipated by unregulated non-photochemical processes
Fv ′ /Fm ′FluorescenceMaximum photochemical efficiency of PSII (light-adapted)
LEFFluorescenceRate of linear electron transport through PSII
LTDThermalLeaf temperature differential (leaf vs. air)
Table 1: The seven stress factors
Figure 2: Multi-head architecture schematic. A shared Transformer backbone over signal patches feeds task-specific heads: the patch-reconstruction head used in pretraining and the stress-factor head, with room for further heads.
Figure 3: The pretraining corpus by species. Each circle is one species, its area proportional to that species’ recorded electrode-hours.
Figure 4: The labeled corpus inside the pretraining corpus. Each hollow circle is one species’ electrode-hours of pretraining recording, and the filled circle within it is the share of those hours that carry ground-truth labels. Both circles are drawn to the same area scale as in the pretraining corpus figure.
Figure 5: Masked-patch pretraining. A fraction of input patches are hidden. The backbone and patch-reconstruction head train together to reconstruct them from the patches that remain visible. Grey marks the hidden spans, and the reconstruction is drawn across them against the observed signal.
Figure 6: Masked patch modeling on one four-day window from a held-out Glycine max plant, one sensor. Grey marks the patches hidden from the model at evaluation, half the window in 64-minute blocks (§2.6.4). The observed signal runs the whole window, and the model’s reconstruction covers the hidden spans with the estimated 95 percent interval shaded. The dashed line is the linear interpolation reference predictor.
Figure 7: The same window under forecasting. Grey marks the final day, hidden and forecasted from the three days before it. The forecast is drawn against the observed signal with its 95 percent interval shaded, and the dashed line is the seasonal naive reference predictor.
Figure 8: Stress-factor head training. The backbone has been pretrained with MPM and is frozen. The stress-factor head trains on the paired ground-truth readings of the labeled corpus (§2.4). The head reads a window of the electrical signal and estimates every stress factor’s state at the end of the window.
Figure 9: The three backbone training schemes. Each row is one complete pipeline, running left to right from the electrical signal to the estimated stress factors. The rows differ only in the backbone, and each backbone is labeled with what it starts from and whether it trains.
PredictorWhat it predictsRMSE-w σ\sigma
Foundation model −- 0.343
Foundation model, species held out −- 0.361
Linear interpolationa line across each gap−- 0.390
Table 2: Reconstruction error over the hidden patches, against linear interpolation
PredictorWhat it predictsRMSE-w σ\sigma
Foundation model −- 0.912
Foundation model, species held out −- 0.968
Seasonal naivethe value 24 hours earlier−- 1.011
Persistencethe last value in the context−- 1.085
Window meanthe mean of the context−- 1.126
Table 3: Forecast error over the masked day, against reference predictors
LeadFoundation modelSpecies held outSeasonal naivePersistence
0–1h−- 0.202−- 0.279−- 0.800−- 0.210
1–3h−- 0.431−- 0.529−- 0.821−- 0.483
3–6h−- 0.700−- 0.775−- 0.840−- 0.764
6–12h−- 0.838−- 0.893−- 0.884−- 1.068
12–24h−- 0.946−- 0.980−- 0.960−- 1.132
Table 4: Forecast RMSE-w σ\sigma by lead time
Figure 10: Tracking at two corpus sizes. Φ\Phi NPQ over 45 days for one droughted plant from the test split of each of two densely labeled species. The head’s estimate and 95 percent interval are drawn against the observed ground-truth trace, and grey bands mark watering. Each circle is that species’ labeled corpus on the same area scale as §2.2’s figures, and the trajectory r and AUC beside it are the species’.
FactorPlantsSensorsTrajectory rCCCRMSE-s σ\sigmaAUC
No skill  −- 0−- 0−- 1.000−- 0.500
Φ\Phi II−- 35−- 70−- 0.431−- 0.293−- 1.021−- 0.697
Φ\Phi NPQ−- 35−- 70−- 0.452−- 0.293−- 0.989−- 0.686
Φ\Phi NO−- 35−- 70−- 0.448−- 0.393−- 0.968−- 0.714
Fv ′ /Fm ′−- 35−- 70−- 0.450−- 0.311−- 0.983−- 0.704
LEF−- 29−- 58−- 0.707−- 0.554−- 0.831−- 0.807
Table 5: Per-factor results on the held-out test split
SpeciesPlantsΦ\Phi IIΦ\Phi NPQΦ\Phi NOFv ′ /Fm ′LEFLTDCWSI
Acer rubrum ‘October Glory’−- 7−- 0.549−- 0.475−- 0.452−- 0.430−- 0.633−- 0.480−- 0.373
Malus domestica ‘Braeburn’−- 7−- 0.678−- 0.496−- 0.578−- 0.422−- 0.788−- 0.271−- 0.250
Picea abies−- 7−- 0.350−- 0.738−- 0.794−- 0.825−- 0.901  
Dianthus barbatus−- 6−- 0.246−- 0.204−- 0.162−- 0.237 −- 0.320−- 0.335
Codiaeum variegatum−- 4−- 0.030−- 0.074−- 0.283−- 0.202−- 0.110−- 0.092−- 0.001
Phoenix roebelenii−- 4−- 0.480−- 0.474−- 0.352−- 0.412−- 0.629  
Table 6: Per-species within-plant tracking (trajectory r)
Figure 11: Corpus size against tracking, species by species. Φ\Phi NPQ trajectory r over the test split for each of the six densely labeled species, plotted against each species’ labeled electrode-hours. The dashed line marks the no-skill value.
FactorPlantsSensorsFrozen randomTask-trainedFoundation model
No skill  −- 0−- 0−- 0
Φ\Phi II−- 91−- 182−- 0.092−- 0.134−- 0.193
Φ\Phi NPQ−- 91−- 182−- 0.083−- 0.119−- 0.191
Φ\Phi NO−- 91−- 182−- 0.131−- 0.192−- 0.246
Fv ′ /Fm ′−- 91−- 182−- 0.068−- 0.114−- 0.198
LEF−- 76−- 151−- 0.217−- 0.195−- 0.283
Table 7: CCC by backbone on every labeled species
GroupSpeciesPlantsΔ\Delta CCCΔ\Delta trajectory r
Densely labeled species−- 6−- 35−- 0.043−- 0.025
Every other species−- 29−- 56−- 0.115−- 0.024
Table 8: The foundation model’s gain over a frozen random backbone
Figure 12: Treatment effect against a contemporaneous control. Cohen’s d for four of the head’s estimated stress factors in salt-treated and glyphosate-treated Coleus scutellarioides , against the untreated cohort, over twelve days from treatment onset. Shading is the 95 percent bootstrap interval, and the dashed lines mark the conventional large-effect threshold of ±\pm 0.8.
FactorUncalibratedCalibratedGainGain 95% CI
Φ\Phi II−- 0.362−- 0.431−- 0.069−- 0.063 to 0.214
Φ\Phi NPQ−- 0.238−- 0.452−- 0.214−- 0.086 to 0.343
Φ\Phi NO−- 0.183−- 0.448−- 0.265−- 0.119 to 0.386
Fv ′ /Fm ′−- 0.178−- 0.450−- 0.272−- 0.102 to 0.434
LEF−- 0.629−- 0.707−- 0.078−- 0.010 to 0.138
LTD−- 0.176−- 0.292−- 0.116−- 0.095 to 0.285
Table 9: Trajectory r with and without calibration
FactorUncalibratedCalibratedGainGain 95% CI
Φ\Phi II−- 0.251−- 0.293−- 0.042−- 0.062 to 0.159
Φ\Phi NPQ−- 0.127−- 0.293−- 0.166−- 0.043 to 0.304
Φ\Phi NO−- 0.108−- 0.393−- 0.285−- 0.158 to 0.393
Fv ′ /Fm ′−- 0.079−- 0.311−- 0.232−- 0.089 to 0.385
LEF−- 0.453−- 0.554−- 0.101−- 0.035 to 0.168
LTD−- 0.008−- 0.234−- 0.226−- 0.045 to 0.388
Table 10: CCC with and without calibration
 Φ\Phi IIΦ\Phi NPQΦ\Phi NOFv ′ /Fm ′LEFLTDCWSI
Worse sensor−- 0.258−- 0.312−- 0.319−- 0.312−- 0.620−- 0.134−- 0.106
One sensor−- 0.431−- 0.452−- 0.448−- 0.450−- 0.707−- 0.292−- 0.269
Two, blended−- 0.498−- 0.507−- 0.502−- 0.504−- 0.747−- 0.368−- 0.330
Better sensor−- 0.577−- 0.573−- 0.561−- 0.569−- 0.777−- 0.436−- 0.418
Table 11: Single-sensor and blended tracking (trajectory r)
Figure 13: Φ\Phi NPQ over 45 days for one plant from the test split of Acer rubrum ‘October Glory’. The two sensors on the plant and their blend are drawn against the observed ground-truth readings, each labeled with its trajectory r. The circle is the species’ labeled corpus on the same area scale as §2.2’s figures, and the trajectory r and AUC beside it are the species’.
More in this topic

Plant Phenotyping

Plant phenotyping uses imaging, sensing, and computational methods to measure plant traits, growth, health, and responses to the environment.

See all 68 papers

Explore similar work

CardsList
Apr 30, 2026

Early Detection of Water Stress by Plant Electrophysiology: Machine Learning for Irrigation Management

Purpose: Fast detection of plant stress is key to plant phenotyping, precision agriculture, and automated crop management. In particular, efficient irrigation management requires early identification of water stress to optimize resource use while maintaining crop performance. Direct physiological sensing offers the potential to detect stress responses before visible symptoms appear. Methods: In this study, we recorded electrophysiological signals from greenhouse-grown tomato plants subjected to water stress and developed a framework based on machine learning for online stress detection. The recorded time-series data were processed using a processing pipeline that includes statistical feature extraction and selection, automated machine learning or alternatively deep learning, and probability calibration. Results: Across multiple input time horizons, we found that a 30-minute look-back window strikes the best balance between rapid decision-making and classification performance. Using automated machine learning, the framework achieved classification accuracies of up to 92%, outperforming deep learning approaches. Sequential backward selection reduced the feature set while maintaining performance. Importantly, the framework detects transitions from healthy to stressed states in recordings that were not included in the training set. Conclusion: Overall, we provide a decision-support tool for farmers and establish a foundation for biofeedback-driven irrigation control to improve resource efficiency in (semi-)autonomous crop production systems.

Eduard Buss, Till Aust, Heiko Hamann

Department of Computer and Information Science, University of Konstanz, Universitätsstraße 10, Konstanz, 78464, Baden-Württemberg, Germany.

Topic pending
AskPDF
Mar 29, 2026

SPROUT: A Scalable Diffusion Foundation Model for Multi-Crop Plant Phenotyping

Image-based plant phenotyping depends on dense structural understanding of crops, yet pixel-level annotation remains expensive across species, organs, growth stages, and field conditions. General-purpose vision foundation models offer a natural route to label efficiency, but their web-scale pretraining objectives transfer weakly to agricultural imagery, where semantics are often determined by fine organ geometry inside repetitive, texture-dominated scenes. We introduce SPROUT, a diffusion foundation model for multi-crop plant phenotyping. SPROUT learns from 2.6 million unlabeled open-field images (MCD-2.6M) using a pixel-space Diffusion Transformer, and selects transferable features with a label-free effective-rank criterion over denoising timesteps. This design shifts pretraining from crop-based invariance to structure-preserving denoising, making the representation better aligned with dense phenotyping tasks. We evaluate SPROUT across dense phenotyping tasks, including organ segmentation, crop-weed parsing, depth estimation, and counting. SPROUT consistently improves over strong web-pretrained baselines, with the largest gains on dense structural prediction, and shows favorable label and compute efficiency compared with general-purpose and crop-specific foundation models. The source code and MCD-2.6M dataset are publicly available.

Shuai Xiang, James Burridge, Shouyang Liu +4

Graduate School of Agricultural and Life Sciences, The University of Tokyo, 1-1-1 Midori-cho, Nishitokyo, Tokyo 188-0002, Japan · Engineering Research Center of Plant Phenotyping, Ministry of Education; Jiangsu Collaborative Innovation Center for Modern Crop Production; Academy for Advanced Interdisciplinary Studies, Nanjing Agricultural University, Nanjing 210095, China · National Key Laboratory of Multispectral Information Intelligent Processing Technology, School of Artificial Intelligence and Automation, Huazhong University of Science and Technology, Wuhan 430074, China +3

Agricultural Image AnalysisPlant PhenotypingRepresentation LearningVision Foundation ModelsSemantic Segmentation
AskPDF
May 2, 2026

Developing a Strong Pre-Trained Base Model for Plant Leaf Disease Classification

Plants, crops and their yields are essential to our very existence, but diseases and pests cause large losses every year. As such it is vital to ensure that diseases can be spotted early and treated accordingly and stopping the spread while still possible. Manual and traditional methods require personal to walk through the field and check for symptoms 'by hand'. This is very laborious and very time consuming, so ML methods have been applied as a result and they have garnered promising results. CNN models are especially efficient as they can automatically extract features from images without any manual feature construction before then feeding the features to a classifier. Datasets are largely influential to the final performance of the model. Despite the importance that datasets pose to the field, there still seems to be somewhat of a discrepancy between what is publicly available for use and what would be required to sufficiently train fully capable models. To overcome these shortcomings, as part of this thesis open datasets for the field of plant leaf disease classification have been identified as well as models that can be trained on them and extensive benchmarks have been carried out to identify their suitability. Then a new dataset was constructed based on those findings as well as on the findings of a augmentation applicability study, which will be used to train a new Base Model based on the DenseNet201 architecture, which managed to outperform the baseline model on said new dataset as well as outperforming it on plant leaf disease classification domain specific Transfer-Learning experiments on another new dataset. This new model manages to train models through Transfer-Learning (TL) faster, more robust, more stable, and with less data than general model would, overcoming a large number of issues that the field still suffers from.

David J. Richter

Department of Artificial Intelligence Convergence Graduate School, Chonnam National University

Plant Disease ClassificationTransfer LearningConvolutional Neural NetworksData-Efficient Learning
AskPDF