Zero-Shot Transfer Capabilities of the Sundial Foundation Model for Leaf Area Index Forecasting

AI summary

Overview

Research area: Time series forecasting with foundation models applied to remote-sensing vegetation monitoring (Leaf Area Index prediction in agriculture).

Technical level: Intermediate. The paper is written accessibly and requires only general familiarity with machine learning, time series forecasting, and supervised baselines such as LSTMs; no specialized remote-sensing background is needed.

Scope: A systematic zero-shot benchmark comparing the Sundial time series foundation model against statistical baselines and a fully supervised LSTM for LAI forecasting over the continental United States using the HiQ dataset.

What This Paper Is About

Leaf Area Index (LAI) is a biophysical variable that quantifies canopy structure and is central to agricultural and ecological monitoring, but forecasting it is difficult because cropping systems are diverse, environmental conditions vary, and labeled data are limited. The paper asks whether a general-purpose, pretrained time series foundation model (Sundial) can, with no task-specific training at all, match or beat a supervised LSTM trained specifically for LAI prediction. The authors find that the answer depends on how much historical context the foundation model is given.

Key Contributions

1. A rigorous zero-shot benchmark of the Sundial foundation model for LAI remote-sensing prediction using the comprehensive HiQ dataset.
2. A demonstration that Sundial, in the zero-shot setting, outperforms a fully supervised LSTM provided the historical input context is sufficiently long, challenging the assumption that task-specific fine-tuning is necessary for optimal performance.
3. An analysis of the context window threshold at which zero-shot performance overtakes the supervised baseline, contrasting instance-level in-context adaptation in foundation models with dataset-level generalization in LSTMs.
4. Evidence that "plug-and-play" forecasting is feasible for pixel-level LAI, together with robustness tests under randomly missing observations.

Main Findings

• Zero-shot Sundial beats a trained LSTM with enough context: Sundial's accuracy improves monotonically with input window size and surpasses the supervised LSTM once the window reaches T_in >= 512 time steps (approximately 11 years given the 8-day cadence). The paper frames this threshold as needing a context covering more than one or two full seasonal cycles.

• One-step-ahead numbers: At horizon H = 1, Sundial reaches its best results at T_in = 1024 with MAE 0.096, RMSE 0.140, R² 97.32%, MAPE 19.15%, and CVRMSE 15.44%. The supervised LSTM peaks earlier, at T_in = 64, with RMSE 0.140, R² 97.34%, MAE 0.097, MAPE 18.83%, and CVRMSE 15.78%.

• The LSTM degrades with long windows: LSTM performance is best at short windows and worsens as the window grows (for example, RMSE 0.148 at T_in = 256 and 0.143 at T_in = 512 versus 0.140 at T_in = 64), which the authors attribute to a limited ability to exploit extended temporal context.

• Sundial's advantage grows at longer horizons: In the unified multi-horizon comparison, Sundial's RMSE is lower than the LSTM's for T_in >= 512 and the gap widens as H increases. At T_in = 1024 and H = 12, Sundial achieves RMSE 0.210, whereas the LSTM was not evaluated at T_in = 1024. At T_in = 512, Sundial's RMSE goes from 0.145 (H = 1) to 0.234 (H = 12), while the LSTM goes from 0.143 to 0.269.

• Statistical significance: A paired t-test at H = 4 confirms Sundial's improvement over the LSTM is highly statistically significant (p < 0.001, N ≈ 10,000), despite large per-pixel RMSE variance from spatial heterogeneity.

• Classical baselines are clearly weaker: Under the one-step-ahead benchmark, the Mean baseline (window 4) yields MAE 0.274, RMSE 0.375, R² 72.6%, MAPE 36.37%, and CVRMSE 32.90%; Last Value gives RMSE 0.210; Trend gives RMSE 0.210; ARIMA(2,1,0) improves from RMSE 0.204 at window 32 to 0.196 at window 512 but remains noticeably above the learning-based methods.

• Sundial tolerates missing data without interpolation: Sundial has a native pipeline for NaN inputs. Table 2 reports relative RMSE increases at H = 1 for missing ratios of 5%, 10%, 20%, and 40% across windows 1024, 512, 256, 128, and 64. At a 20% missing ratio the increases range from 7.5% (window 64) to 25.6% (window 1024); at a 40% missing ratio they range from 36.7% to 57.1%. The text describes masking at rates of 5%, 10%, and 20%, while the table also includes a 40% row.

• Simple heuristics fail on seasonality: The Mean baseline underestimates seasonal amplitude, Last Value lags the seasonal cycle, and Trend tracks short-term tendencies but fails near seasonal turning points.

Methodology in Plain English

The authors take an existing, publicly released Sundial model checkpoint and never modify or fine-tune it. They treat the historical LAI sequence as a "prompt" — the same way a language model is given text — and let the model autoregressively generate future values from that prompt, performing no task-specific loss optimization.

They evaluate on the HiQ dataset, a reprocessed version of the MODIS LAI product that improves spatiotemporal consistency. They restrict the study to the continental United States (CONUS) subset for computational reasons, using 5-km resolution pixels with 8-day sampling and roughly 1,050 time steps per location, covering 2000–2022.

Every model uses the same fixed sliding-window protocol: an input length T_in from {32, 64, 128, 256, 512, 1024} and a forecast horizon H from {1, 4, 8, 12}, corresponding to 8, 32, 64, and 96 days ahead. Evaluation uses a strict forward-in-time split so models are conditioned on history and tested on the most recent unseen data.

Comparison models include simple heuristics (Last Value, Mean(k), Trend), an ARIMA(2,1,0) fitted per test window using state-space estimation, and a supervised LSTM with two stacked layers of 64 hidden units and a fully connected output layer, trained with the Adam optimizer at learning rate 0.001, batch size 256, and early stopping on validation loss.

Preprocessing involves per-pixel normalization; for supervised models, the mean and standard deviation come from valid observations in the training period. Missing values are masked during loss computation or linearly interpolated where required, and pixels with a missing fraction greater than 0.1 are excluded from training and evaluation. Performance is measured with MAE, RMSE, R², MAPE, and CVRMSE, aggregated over 100 randomly sampled locations with roughly 100 rolling test windows per pixel, giving averages over about 10,000 × H prediction–ground-truth pairs.

Why This Matters

Impact on research: The paper challenges the default assumption that task-specific supervised training is necessary for strong performance on specialized scientific forecasting tasks. It identifies a concrete crossover point — context length — and offers a mechanistic explanation: the LSTM learns a "global average" pattern across the dataset and smooths over site-specific anomalies, whereas Sundial performs instance-level in-context learning and effectively builds a temporary local model for each pixel in real time.

Real-world applications:

• Agricultural monitoring and agronomic decision-making, where LAI informs canopy status and crop growth assessment.
• Crop yield estimation, which the paper links directly to accurate LAI forecasting.
• Land-surface and climate analysis, since vegetation influences regional climate through biogeophysical processes such as albedo and roughness and through biogeochemical cycles.
• Rapid deployment in data-scarce or under-monitored regions, where collecting and annotating training data is impractical but a pretrained model can be used off the shelf.

Industry relevance: The "plug-and-play" result is directly relevant to agtech, Earth-observation analytics, and climate-risk modeling, where it offers a scalable alternative to the resource-intensive "collect, annotate, and train" workflow. The authors are explicit that inference cost of foundation models remains higher than lightweight baselines, which matters for operational deployment; they also note the long-history requirement may limit use in regions with sparse records.

Future Directions

• Few-shot fine-tuning to reduce the context window requirement, since the need for T_in >= 512 (about 11 years) restricts use where historical records are short.
• Adding climatological drivers such as temperature and precipitation as input features to improve forecasts under varying environmental conditions.
• Multivariate modeling to jointly model multiple vegetation indices and ecosystem variables for more comprehensive ecological status prediction.
• Broader spatial transfer experiments at larger geographic scales to rigorously evaluate generalization and support global vegetation monitoring, since the present study is limited to the CONUS subset.

Open questions implied by the work include whether the context-length crossover generalizes to other foundation models and other regions, and how to close the gap between foundation-model accuracy and the inference cost of lightweight baselines.

Target Audience

This paper is most valuable to machine learning researchers working on time series foundation models and zero-shot transfer, to remote-sensing and geospatial scientists doing vegetation or LAI forecasting, and to agricultural and environmental practitioners evaluating whether pretrained models can replace bespoke supervised pipelines. It is also a useful case study for readers interested in the practical trade-offs between in-context learning and task-specific training, and for engineers assessing deployment costs versus accuracy in operational monitoring systems.

Authors’ abstract

This work investigates the zero-shot forecasting capability of time series foundation models for Leaf Area Index (LAI) forecasting in agricultural monitoring. Using the HiQ dataset (U.S., 2000-2022), we systematically compare statistical baselines, a fully supervised LSTM, and the Sundial foundation model under multiple evaluation protocols. We find that Sundial, in the zero-shot setting, can outperform a fully trained LSTM provided that the input context window is sufficiently long-specifically, when covering more than one or two full seasonal cycles. We show that a general-purpose foundation model can surpass specialized supervised models on remote-sensing time series prediction without any task-specific tuning. These results highlight the strong potential of pretrained time series foundation models to serve as effective plug-and-play forecasters in agricultural and environmental applications.