acceptodds
Under review as a conference paper at ICLR 2027

SPEAR-SIM: Process-Guided Weak Supervision for Agricultural Foundation Models Using Mechanistic Crop Simulations

Abstract

Earth observation foundation models have demonstrated strong transferability across remote sensing tasks, yet adapting them to agricultural applications remains challenging because large-scale supervision for intermediate crop physiological processes is largely unavailable. Existing downstream adaptation therefore relies primarily on sparse labels such as crop yield, providing limited guidance for learning physiologically meaningful representations. An observation-constrained process-guided adaptation framework, SPEAR-SIM, is proposed to leverage mechanistic crop simulations as weak supervision for Earth observation foundation models. The SPEAR-SIM framework first generates ensembles of APSIM crop growth trajectories and selects simulations that best agree with satellite-observed crop development, producing observationally validated teacher signals for multitask adaptation of a pretrained Earth observation foundation model. The resulting representations are evaluated through downstream U.S. county-level crop yield prediction across multiple forecasting cutoffs. Experimental results demonstrate consistent improvements over both handcrafted-feature baselines and conventional foundation model adaptation. The proposed framework increases early-season corn prediction performance from an of 0.308 to 0.623 when evaluated beyond the regions where simulator-derived supervision was available, while providing consistent improvements for soybean during mid-to-late season forecasting. Additional analyses further demonstrate that both observation-constrained teacher selection and physiological auxiliary supervision contribute substantially to learning informative crop representations. More broadly, these findings demonstrate that observation-constrained process-guided weak supervision provides an effective paradigm for integrating mechanistic crop simulations with foundation models and offers a general strategy for combining scientific simulators with modern self-supervised learning.

Primary area:
unsupervised, self-supervised, semi-supervised, and supervised representation learning
Keywords:
Process Based Models, Crop Simulations, Weak Labeling

Reviews

Among:
Shape:
1234this paper, Bottom 10%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 1.67 over 3 reviews, above 10% of the others.
Reviewer_fYBJOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper proposes SPEAR-SIM, a process-guided adaptation framework that leverages mechanistic crop simulations as weak supervision for learning physiologically meaningful representations. The framework is evaluated on early-season crop yield prediction, demonstrating improvements over handcrafted feature baselines and standard foundation model adaptation. And the improvement is especially obvious in the early stage with fewer labels, which suggests the effectiveness of the proposed framework.

Critical strengths
  1. The paper addresses the limited availability of physiological process supervision in agricultural foundation model adaptation, and proposes a novel adaptation framework that introduces crop simulation as weak supervision to improve intermediate physiological representation learning.

  2. The experiments demonstrate consistent improvements over several baselines, while the representation visualizations provide an intuitive illustration of improved physiological representations.

Critical weaknesses
  1. The proposed framework relies on crop simulations as weak supervision, but the reliability of the simulated trajectories and their spatio-temporal alignment with real observations are not sufficiently evaluated. Figures 2 and 5 suggest that the observed soybean satellite trajectories are not well aligned with those simulated by APSIM, and Table 2 also shows that the variant without the simulator performs better in May and June. These observations raise concerns about whether misaligned simulations could introduce misleading supervision and limit the framework's generalization capability across different scenarios.

  2. The main contribution appears to be the simulation-guided adaptation framework rather than a new model architecture. Thus, it would be better to apply this adaptation framework to other models to demonstrate whether the improvements arise from the proposed framework and whether it generalizes across different architectures.

Questions for authors
  1. How sensitive is performance to the crop simulation filtering strategy? Could the authors provide ablations comparing different trajectory selection strategies to isolate the contribution of the proposed filtering procedure?

  2. Is there a metric to assess the reliability of simulated crop trajectories and can be used as guidance for weak supervision?

  3. Can the proposed process-guided adaptation framework be applied to other foundation models?

  4. How is the filtered trajectory constructed, and what preprocessing is applied before it is used in the loss in Equation (2)?

  5. How is the peak window defined, and how is peak normalization performed?

Reviewer_iXFhOpenReview ↗
Rating
1
Confidence
2
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper proposes SPEAR-SIM, a process-guided agricultural representation learning framework that uses satellite-filtered crop simulations as weak supervision to adapt SPEAR-NeXT, aiming to improve early-season crop yield prediction without requiring yield labels during adaptation. Experiments demonstrate the effectiveness of SPEAR-SIM.

Critical strengths

1.The observation-constrained teacher selection and process-guided adaptation are interesting designs, leveraging satellite-filtered APSIM simulations as weak supervision to enhance agricultural representations without requiring yield labels.

Critical weaknesses
  1. The methodological novelty is limited, and the work appears incremental, mainly combining observation-filtered crop simulations as weak supervision with continued adaptation of the existing SPEAR-NeXT model.
  2. The contribution of simulator supervision is not isolated. A baseline with matched continued self-supervised training but without simulator supervision is needed.
Questions for authors
  1. For early-season yield prediction, does the teacher selection process use only satellite observations available before each forecasting cutoff, or does it incorporate future observations?
Reviewer_K2vzOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
3
Read review
Paper summary

SPEAR-SIM adapts a pretrained Earth-observation model (SPEAR-NeXT) for county-level crop yield prediction using APSIM crop simulations as training-only supervision. Simulations are filtered by agreement with a Sentinel-2 canopy proxy, averaged into a teacher trajectory, and used as an auxiliary target during continued self-supervised pretraining. A frozen-feature MLP probe then predicts yield under leave-one-year-out evaluation. The method improves over the unadapted backbone at all corn cutoffs and at five of seven soybean cutoffs.

Critical strengths
  1. The problem is well motivated, and using a simulator as privileged training-time signal is a sensible framing.

  2. The paper is candid about limitations, including proxy circularity, the lack of spatial generalization evidence, and weaker soybean results.

  3. The comparison against the same backbone under an identical protocol is the right controlled design.

Critical weaknesses
  1. The central claim is not isolated. Stage A changes two things at once: continued pretraining on target-crop data and the simulator loss. The baseline is the released checkpoint with no continued pretraining, so the gains cannot be attributed to the simulator.

  2. Evaluation may leak location information. LOYO holds out years, not counties, and yields are strongly autocorrelated across years. Since the probe sees other years from every evaluated county, the reported gains may reflect location priors rather than transferable crop representations. A state- or county-disjoint evaluation is needed to support the generalization claims.

  3. The headline result is not fold-clean. Stage A is trained once on all seasons of the teacher states, including the held-out year’s imagery and year-specific teachers derived from that year’s weather. Because end-of-season APSIM output under true weather is itself a noisy yield proxy, the protocol is not year-clean.

  4. Cross-state generalization is asserted, not shown. The abstract claims gains in regions without simulator supervision, but results are aggregated across states. Without a per-state breakdown separating teacher and non-teacher states, the gains could be concentrated in the teacher region.

Questions for authors
  1. How do results change under a state-disjoint split, with Stage A and Stage B restricted to teacher states?

  2. Does the April corn gain persist when Stage A is rerun per LOYO fold, excluding the held-out year’s imagery and teachers?

  3. What are the per-state results separating teacher and non-teacher states?

From OpenReview, CC BY 4.0. Authors as signed there; shown unedited.

Related papers

Loading the map…

Discussion (0)

Sign in to comment.