acceptodds
Under review as a conference paper at ICLR 2027

PHENOBIID: Rethinking In-Season Crop Yield Prediction from a World Model Perspective

Abstract

Crop yield is the terminal outcome of a dynamic growth process, where weather drives vegetation evolution and accumulated seasonal development determines harvest productivity. However, during the growing season, only part of this process is observable before harvest, making it challenging to predict yield from incomplete growth information. Existing in-season yield prediction methods mainly aggregate available observations and directly regress final yield, leaving the hidden evolution of vegetation dynamics insufficiently modeled. In this work, we reformulate in-season yield prediction as a predictive world-modeling problem and introduce PhenoBIID, a task-specific weather-conditioned vegetation world model that learns weather-driven remote sensing vegetation state transitions and predicts hidden growth trajectories. Instead of directly mapping partial seasonal observations to yield, PhenoBIID reconstructs the unobserved portion of vegetation development and uses the completed trajectory for yield estimation, explicitly connecting environmental forcing, vegetation dynamics, and crop outcomes. To evaluate this perspective, we introduce CropDynamicsBench, a global benchmark covering maize, rice, soybean, and wheat that jointly evaluates vegetation trajectory forecasting and yield prediction under varying observation availability. In retrospective evaluations under realized weather, PhenoBIID reduces seed-mean yield RMSE by 0.12–4.74% relative to the strongest evaluated direct baseline at each crop–cutoff setting with 10–70% unobserved windows. Our findings highlight the potential of predictive world modeling to connect crop growth dynamics with yield assessment and support agricultural planning. Code, processed benchmark inputs, an executable inference sample, and evaluation artifacts are available at https://anonymous.4open.science/r/PhenoBIID-5EBB

Primary area:
applications to physical sciences (physics, chemistry, biology, etc.)
Keywords:
Remote Sensing, Crop Yield Prediction, World Models, Benchmark

Reviews

Among:
Shape:
1234this paper, Top 49%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.33 over 3 reviews, above 51% of the others.
Reviewer_J8C5OpenReview ↗
Rating
2
Confidence
4
Informativeness
2
Presentation
1
Soundness
2
Read review
Paper summary

This paper investigates in-season crop yield prediction from a world model perspective, aiming to predict crop yields under incomplete observations of vegetation development, and proposes PhenoBIID, which models vegetation growth dynamics to reconstruct unobserved vegetation trajectories and improve downstream yield prediction. The paper also introduces a benchmark for evaluating vegetation trajectory forecasting and crop yield prediction under partially observed growing seasons. The proposed PhenoBIID reduces RMSE by 0.12–4.74% relative to the strongest ML-based baselines with 10–70% unobserved windows.

Critical strengths
  1. In-season crop yield prediction is an important task for agricultural monitoring and decision-making.

  2. Modeling intermediate vegetation under a world model-based paradigm to predict final yield is an interesting direction.

Critical weaknesses
  1. The motivation and novelty of the world model formulation are unclear. In-season yield prediction has been studied previously [1], but the paper does not clearly identify the specific challenges that motivate its proposed framework and benchmark. If the primary objective is to forecast vegetation dynamics for yield prediction, the distinction from a temporal forecasting model with a yield prediction head needs to be established.

  2. The experimental evidence does not sufficiently justify the additional architectural complexity of PhenoBIID. Comparisons are primarily against traditional machine learning methods, yet the reported performance improvements appear relatively modest (e.g., Maize: RMSE decreases from 0.8689 to 0.8584; Soybean: RMSE decreases from 0.3243 to 0.3118), considering the PhenoBIID introduces substantially more architectural complexity.

  3. Some minor weaknesses:

  • Presentation and readability issues: Figure 2 is not in SVG format; Figure 11 in the appendix appears to contain AI-generated graphical artifacts (i.e., distorted symbols in the LightGBM residual head module)
  • Essential methodological details, such as the Bidirectional relevance modulation module, should be included in the main paper rather than relying heavily on the appendix.

[1] You, J., Li, X., Low, M., Lobell, D., and Ermon, S. Deep Gaussian Process for Crop Yield Prediction Based on Remote Sensing Data. AAAI 2017.

Questions for authors
  1. Is explicit vegetation trajectory forecasting necessary for yield prediction? As the proportion of masked observations increases from 10% to 70%, the RMSE for rice changes only slightly (0.8440 --> 0.8584), with similarly limited variation across other crops. How should these results be interpreted in relation to the role of trajectory reconstruction?

  2. How does PhenoBIID compare with deep learning-based baselines? In particular, how does it perform against a model that directly predicts yield from available observations without explicitly reconstructing missing vegetation trajectories?

  3. What is the unique contribution of the proposed benchmark? Could the same evaluation setting be established by masking selected observation windows in existing crop yield datasets? Moreover, given that the benchmark is a main contribution, it's better to introduce its construction details in the main paper rather than primarily in the appendix.

  4. The variables included in the context are not clearly specified, and initinitialization of (S_0), particularly how (V^{prev}) and the context are fused, is not adequately described.

  5. In Equation (1), the transition operator (i.e., C_p) is introduced without a clear explanation of its underlying operations. The variables (u_{p,k}) and (F_p) in Equation (2), as well as (X) in Equation (3), are not sufficiently defined.

  6. Do the encoders and BIID modules share parameters across stages, or does each stage use a separately parameterized module? How are these modules trained?

  7. How does the modified hidden-window completion rule in Section 5.3.1 differ from the original setting? The reported RMSE reductions for rice are consistently negative under this setting. Could the authors discuss the potential reasons for this deterioration?

6 comments

Dear Reviewer J8C5,

Thank you for your thoughtful feedback on our submission. With the discussion period coming to a close, we wanted to follow up on our reply to your review. If there are any points you would like us to elaborate on or any questions we have not fully addressed, we would be glad to provide further clarification.

Thank you again for your time and for helping us improve the paper.

Best regards, The Authors

Dear Reviewer J8C5,

Thank you again for your thoughtful feedback. We wanted to briefly follow up on our response, particularly the capacity-matched comparison and the rice diagnostics we added in response to your questions. If you have a chance to review these results, we would appreciate hearing whether any points about the value of explicit trajectory forecasting still need clarification.

Thank you for your time.

Reviewer_J8C5OpenReview ↗

Response to author

Thank you for the detailed response and additional experiments. The authors have addressed some of my concerns, and the new experimental results further strengthen the contributions of the proposed method.

However, several concerns remain unresolved:

  1. Motivation for using a world model. I am still not fully convinced of the necessity of a world model for this task. Conventional prediction models can also use recurrent conditioning to update hidden states based on observed feedback and perform multi-step rollouts to progressively predict future vegetation states. Therefore, the authors should further clarify what unique capabilities the proposed world model provides beyond existing temporal prediction frameworks.

  2. Performance on rice. The proposed method still does not achieve state-of-the-art performance on rice. Could this be related to differences in the rice data distribution compared with other crops? A more detailed analysis would help explain this performance gap and clarify the limitations of the proposed method.

So I keep my original socre. and I believe addressing these remaining concerns would make the paper more convincing and strengthen it for a future submission.

Reviewer_Cv49OpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper proposed a World Model-based solution, called as PhenoBIID for crop yield prediction. It employed bidirectional relevance modulation to combine the vegetation states and weather information. In the observation feedback, the predicted state and the ground-truth state are combined as the input of the next time step. The output trajectory is then used to prediction crop yield. The extensive experiments were conducted for evaluation.

Critical strengths
  1. Crop yield prediction is an important research area for food security.
  2. The proposed BIID looks novel, as it enables the vegetation states and weather information condition each other bidirectionally.
  3. The experiment settings are clearly described, with all parameters and settings well illustrated. Extensive experiments were conducted for evaluating the effectiveness of the proposed model.
Critical weaknesses
  1. Eq. (2) is a key formula, but some of its elements are not defined, which makes it difficult for readers to follow.

  2. The authors compare their model with GRU, Random Forest, and LightGBM, but they cannot represent the current state of the arts. The authors should include the comparison with the recent and advanced models, to better demonstrate its advantages and effectiveness.

  3. The main experiment setting looks unrealistic. Real weather data is used for the masked "future". In reality deployment, it is impossible to get the future real weather data.

Questions for authors

Please see my comments above.

2 comments

Dear Reviewer Cv49,

Thank you again for taking the time to review our paper and for sharing your feedback.

As the discussion period draws to a close, we wanted to briefly follow up on our response. If any of your concerns remain unresolved, or if any part of our response would benefit from further clarification, we would be happy to address it during the remaining discussion period.

We appreciate your time and consideration.

Best regards, The Authors

Dear Reviewer Cv49,

Thank you again for your helpful comments. We wanted to briefly follow up on the stronger baselines and the matched forecast-weather evaluation we added in response to your review. If you have a chance to look at the updated results, we would be grateful to know whether any questions about the comparisons or evaluation setup remain.

Thank you for your time.

Reviewer_ejAtOpenReview ↗
Rating
3
Confidence
4
Informativeness
2
Presentation
2
Soundness
1
Read review
Paper summary

PhenoBIID predicts missing vegetation observations from weather and uses the completed trajectories to estimate crop yield. The system trains vegetation prediction and yield prediction separately. It also uses historical yield predictors.

CropDynamicsBench evaluates vegetation forecasting and yield prediction across four crops and several observation cutoffs. The main experiments assume known weather over the full season. I assess these results as conditional prediction results.

Critical strengths
  1. The evaluation covers several crops and prediction stages.

The main comparison covers four crops and thirteen evaluation years. It repeats the complete system and strong direct predictors with three random seeds. Model selection uses preceding years, followed by evaluation on later years. [Section 5.1, pp. 5–6; Appendix A.2, pp. 14–15]

The complete system has the lowest average error across the reported crop-cutoff settings. Improvements over the strongest tested direct predictor range from 0.12% to 4.74%. Their size varies, but the breadth of the comparison supports the complete system’s value. [Abstract, p. 1; Table 1, p. 6]

  1. The controls examine the proposed mechanism.

I particularly like the tests that keep the yield predictor fixed while changing only vegetation completion. They separate the value of predicted growth from other parts of the system. [Section 5.3 and Figure 4, p. 7]

The paper also tests weather inputs, observed feedback, interaction direction, and simpler completion rules. BIID improves vegetation prediction over GRU in all fifteen reported crop-product-cutoff conditions. Evaluating vegetation error alongside yield error reveals when better vegetation predictions provide limited yield benefits. [Section 5.3, pp. 7–8; Tables 18–20, p. 37]

  1. Independent labels and forecast weather add useful evidence.

The regional experiments use independent yield statistics from Germany, France, and Poland. PhenoBIID improves over direct predictors and historical-mean completion in five of six crop-country groups across three cutoffs. These tests support benefits beyond the main reconstructed yield dataset. [Appendix C.5, pp. 33–35]

The SEAS5 experiment also tests sensitivity to forecast weather. It preserves gains for several crops under its specified setup. The authors explain that setup’s limits instead of treating it as a complete deployment test. These additions strengthen the empirical contribution. [Section 5.5, p. 9; Appendix C.6, pp. 35–36]

Critical weaknesses
  1. Predicted growth explains only part of the improvement.

The complete system’s improvement does not establish the value of learned vegetation dynamics by itself. The fixed-predictor controls give a narrower answer. Historical-mean completion beats learned completion for rice at all four cutoffs in Figure 4. Several settings with short missing windows also give small or negative gains for other crops. [Figure 4, p. 7; Table 21, p. 40]

Rice is the awkward case here. Better vegetation prediction does not translate into better yield prediction under the proposed completion rule. Longer missing windows give stronger support for maize, soybean, and wheat. The method therefore provides a useful but selective benefit. [Appendix C.3, p. 32]

The paper correctly acknowledges this pattern. My concern is the size and consistency of the added benefit from the more complex state model. The complete system’s gains should remain separate from claims about that specific mechanism.

  1. The complete comparison changes several parts of the system.

PhenoBIID and the direct predictors use different historical predictors and feature encoders. Their overall difference therefore includes more than vegetation completion. The fixed-predictor controls help, but they do not fully explain the complete comparison. [Appendix B, pp. 19–26; Section 5.2, p. 6]

The yield predictor also trains on observed vegetation and receives predicted vegetation during evaluation. Historical predictions on its training rows come from models fitted on those same rows. The readout may therefore learn from unusually accurate training predictions. [Section 4, p. 5; Appendix B.4, p. 23]

Appendix D.5 provides useful controls using vegetation predictions from earlier models. However, historical predictions remain in-sample, and that experiment changes historical predictors and training periods. A fully matched comparison would make the source of improvement clearer. This concern does not establish test leakage. [Appendix D.5, pp. 41–42]

  1. Practical forecasting and independent validation remain incomplete.

The forecast-weather test covers nine years and one seed. It changes the vegetation models and uses earlier yield predictors instead of the Table 1 refits. Periods outside the forecast horizon use historical weather averages. The test therefore does not directly validate the exact main system with forecast weather. [Appendix C.6, pp. 35–36]

The main sample selection and crop calendars also use retrospective information. A practical evaluation should restrict every input to information available at prediction time. [Appendix A.1–A.2, pp. 14–16]

The global GDHY targets are reconstructed yields. Independent regional tests help, but cover only maize and wheat in three European countries. Evidence using independent rice and soybean yields remains limited. These qualifications narrow the demonstrated scope without invalidating the conditional experiments. [Appendix C.5, pp. 33–35]

Questions for authors

How much improvement remains when all methods use the same historical predictor, inputs, and training periods? Can every training prediction come from a model fitted only on earlier data? Can you test the exact Table 1 system and direct predictors using the same forecast-weather inputs? What explains rice’s weaker yield results despite better vegetation prediction?

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.