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

TL;DR

This paper tackles LAI forecasting under cross-site, data-scarce conditions by evaluating the Sundial time-series foundation model in a strict zero-shot setting on HiQ data (2000–2022). The authors show that, with a sufficiently long historical context ($T_{ ext{in}}$ spanning multiple seasons), Sundial can outperform a fully trained LSTM on both one-step and multi-step horizons, challenging the view that task-specific training is always necessary. The work analyzes the mechanisms behind global priors versus instance-level adaptation and discusses the plug-and-play potential of pretrained time-series models for remote sensing. Limitations include the need for long contextual history, univariate focus, and higher inference cost, with future directions toward few-shot fine-tuning, multi-modal inputs, and broader geographic validation.

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. This demonstrates, for the first time, 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.

Figures (5)

• Figure 1: Spatial distribution of LAI across the study domain, shown as four seasonal snapshots from 2001. Higher LAI values (darker green) correspond to denser vegetation.
• Figure 2: Temporal LAI evolution at the representative center pixel (red dot in Fig. \ref{['fig:spatial_lai']}). The series shows clear seasonal cycles and interannual variability.
• Figure 3: Performance of simple statistical baselines on the final 150 time steps of a representative LAI pixel. The first 46 time steps show ground-truth LAI, followed by 104 forecasted steps. The Mean baseline underestimates seasonal amplitude, the Last baseline lags the seasonal cycle, and the Trend baseline tracks short-term tendencies but fails at seasonal turning points.
• Figure 4: Effect of input window size ($T_{\text{in}}$) on forecasting accuracy at horizon $H=4$. RMSE comparison across different window sizes. Larger input windows consistently improve Sundial performance, while ARIMA and LSTM saturate at moderate window lengths. Metrics are averaged per pixel.
• Figure 5: Effect of forecasting horizon ($H$) on forecasting accuracy. Multi-horizon forecasting comparison between Sundial and the supervised LSTM at a fixed context window ($T_\text{in}$ = 512). The x-axis denotes the forecasting horizon H $\in$ {1, 4, 8, 12}, and the y-axis shows the corresponding average RMSE. Although both models degrade as $H$ increases, the zero-shot Sundial model exhibits a substantially slower performance deterioration, outperforming the supervised LSTM consistently for all $H \geq 4$. This demonstrates that Sundial retains long-range temporal structure more effectively, making it particularly suitable for extended-horizon agricultural forecasting.