Source-linked AI summary
Zero-Shot Forecasting of In-Situ Soil Moisture and Temperature with a Time-Series Foundation Model
Mohammad Talaei, Mohammad Goudarzi, Maille Todd, Vanessa N. L. Wong, Iadine Chadès
TL;DR
Newly instrumented restoration sites may lack hydraulic parameters for physical models and months of site-specific training data for data-driven forecasters. This study evaluates zero-shot TimesFM on in-situ telemetry alongside statistical and tree-based baselines. TimesFM performs better at short horizons but worse at longer ones, with inference costs that create a practical deployment trade-off.
Problem
Physical forecasting models need hydraulic parameters rarely measured at restoration sites, while data-driven models typically require months of per-site telemetry.
Source: §1 · pp. 1–2Method
The study evaluates zero-shot TimesFM on gap-aware rolling-origin forecasts from in-situ sensors, comparing it with rule-based statistical baselines and walk-forward CatBoost.
Source: §1 · p. 2, §2 · pp. 2–3Results
TimesFM beats the best classical baseline by 21-76% at 1–8 h but loses by 6-79% beyond 16 h, while CatBoost never wins pooled across the season.
Source: §3 · p. 3Takeaways & Limitations
TimesFM offers a cold-start option for useful sub-daily forecasts, but its higher inference-time and memory costs contrast with stronger seasonal and climatological baselines beyond one day.
Source: §4 · p. 4, §3 · p. 4Conclusions are limited to six months of observations at one pilot site and do not establish whether the performance patterns persist across seasons, climates, or sensor networks.
Source: §4 · p. 4
Abstract
from arXiv · showhide
Soil moisture and temperature govern plant-water availability and microbial activity, making them key indicators for evaluating land-restoration efforts on degraded landscapes. However, the physical and data-driven models normally used to forecast them both need information that a newly instrumented restoration site does not yet have: hydraulic parameters for the former and months of per-site training data for the latter. We evaluate TimesFM 1, a 500M-parameter time-series foundation model, zero-shot on 15-minute telemetry from 27 sensor stations across three land-management regimes (Control, Treatment, Scald) at a semi-arid restoration site, forecasting 5 cm and 30 cm soil moisture and enclosure temperature at seven lead times from 1 h to 168 h. Using gap-aware rolling-origin evaluation across sensors, we find that TimesFM beats the best rule-based statistical forecasting baselines (persistence, seasonal-naive, or 7-day climatology) by 21-76% at sub-daily lead times but loses by 6-79% beyond a day, a pattern explained by a diurnal-alignment mechanism that holds for both soil-moisture depths and temperature. A CatBoost baseline trained walk-forward on the same origins with full history and weather covariates never wins overall, and we show that its weakness is partly a cold-start effect specific to short deployments. We additionally quantify the inference cost of zero-shot forecasting on commodity CPU hardware. Our results highlight a practical trade-off between zero-shot forecasting capability, predictive performance, and computational efficiency.
1 Introduction
Forecasting in-situ soil moisture and temperature can help evaluate restoration interventions, but physical models need rarely measured site hydraulic parameters and data-driven models typically need months of per-site telemetry. The paper asks whether zero-shot TimesFM forecasting holds up on real, gap-ridden sensor data and where it breaks down, while evaluating accuracy and inference efficiency.
1 Introduction
- Restoration evaluation motivates in-situ forecasts because soil moisture regulates plant-water availability and solute transport, while soil temperature governs microbial and root activity.The paper frames forecasting these variables as useful for evaluating interventions such as landscape rehydration.Source: §1 · p. 1
- Physical models need hydraulic parameters rarely measured at restoration sites, while data-driven forecasters typically need months of per-site telemetry unavailable at newly instrumented reserves.Time-series foundation models are presented as a possible zero-shot alternative without per-site training.Source: §1 · pp. 1–2
- The study evaluates TimesFM zero-shot against statistical and tree-based baselines for accuracy and inference efficiency, examining short- and long-horizon regimes and limited-history effects on tree models.Its stated contributions include identifying performance regimes linked to diurnal alignment and examining training-history effects on tree-model performance.Source: §1 · p. 2
2 Data and methods
The study uses six months of 15-minute telemetry from 27 stations to forecast two soil-moisture depths and enclosure temperature. It evaluates TimesFM and statistical baselines across gap-aware rolling origins, with walk-forward CatBoost as a comparison.
2 Data and methods
- Telemetry spans 27 stations across Treatment, Control, and Scald regimes, with forecasts for 5 cm and 30 cm soil moisture and enclosure temperature.Enclosure temperature is treated as a proxy for air temperature, not soil temperature at depth.Source: §2 · p. 2
- Figure 1 shows sensor–gateway topology and a representative seven-day-ahead forecast against ground truth for enclosure temperature and 30 cm soil moisture.The topology marks the farthest sensor pair at 2.5 km.Source: Figure 1 · p. 2
- Gap-aware evaluation uses contiguous fully observed windows for context and horizon, yielding 206 origins for both moisture depths and 137 for temperature across seven lead times.Origins are at least seven days apart per sensor; raw series include gaps up to 69 days.Source: §2 · p. 2
- TimesFM v2.0 (498.8M parameters) is run strictly zero-shot against untrained, rule-based persistence (last value held constant), seasonal-naive (same 15-minute slot 24 hours earlier), and 7-day climatology (preceding week’s same-clock-time mean).Source: §2 · p. 2
- Figure 2 compares TimesFM with a target-specific best baseline selected by mean normalized MAE across all lead times and held fixed by horizon.Table 1 reports pooled-sensor MAE by method, target, and lead time, with lower values better.Source: Figure 2 · p. 3, Table 1 · p. 3
- CatBoost is retrained walk-forward using pre-origin target and feature timestamps, lagged target features, and weather covariates.It is an intentionally generous comparison because TimesFM uses only univariate history capped at 2048 steps.Source: §2 · p. 3
3 Results
TimesFM’s performance changes with forecast horizon: it leads the rule-based baselines at sub-daily leads but loses beyond a day. Its regime split appears across management groups, while CatBoost remains weaker overall and TimesFM incurs greater inference cost.
TimesFM wins short-term, loses long-term
- At 1–8 h, TimesFM beats the best classical baseline by 21–76%, but beyond 16 h loses by 6–79%, with 7-day climatology overtaking everything past four days.The diurnal-alignment explanation is consistent with peak-to-peak amplitude-to-drift ratios of 4.1 (5 cm), 3.4 (30 cm), and 6.6 (enclosure temperature), and with persistence nearly tying TimesFM at moisture 24 h and 168 h.Source: §3 · p. 3
Behaviour is consistent across land-management regimes
- The short/long horizon split holds across Control, Treatment, and Scald; Control generally has higher absolute error at longer moisture horizons.Scald results are descriptive only because the group has three sensors.Source: §3 · p. 3, Figure 3 · p. 4
CatBoost’s weakness is partly a cold-start effect
- CatBoost never wins overall; its 5 cm moisture MAE is 3.3× worse in the smallest-training-data quartile than the largest, and the gap narrows but remains in the best-trained half of origins.Error correlates negatively with training-set size for 5 cm moisture (r = −0.25).Source: §3 · pp. 3–4
Inference cost
- On commodity CPU hardware, TimesFM inference takes 6.4 s and uses +2.7 GB peak memory, while CatBoost incurs a modest training overhead.Table 2 reports local system performance; CatBoost retrains at each 14-day cutoff and per horizon, for 518 refits total.Source: §3 · p. 4, Table 2 · p. 4
4 Discussion, limitations, and future work
TimesFM offers a cold-start option for sub-daily forecasts, but its computational costs and horizon-dependent performance create a deployment trade-off. The conclusions are bounded by a six-month, single-site pilot; future work will test transfer, fine-tuning, data needs, and efficiency across settings.
Discussion
- TimesFM offers useful sub-daily forecasts for newly instrumented sites without site-specific training or GPU infrastructure, but seasonal and climatological baselines are more effective beyond one day.The zero-shot flexibility comes with higher inference-time and memory costs.Source: §4 · p. 4
Limitations
- Six months of observations from one pilot site do not establish whether the performance patterns persist across seasons, climates, or sensor networks.Scald results are descriptive with only three sensors; enclosure temperature is an air-temperature proxy, and moisture telemetry was not calibrated against a reference probe.Source: §4 · p. 4
Future work
- Future work will examine when TimesFM fine-tuning helps, how much data site-specific models require, whether forecasting knowledge transfers, and efficiency across deployment settings.The planned efficiency evaluation includes inference time, memory footprint, and training cost.Source: §4 · p. 4