Technical Summary: Agent-Based Post-Hoc Correction of Agricultural Yield Forecasts

Problem Statement

Accurate crop yield forecasting is critical for operational planning in commercial agriculture (e.g., labor allocation, cold storage, and retail contracts). However, state-of-the-art forecasting approaches typically rely on data modalities unavailable to most commercial growers, such as dense sensor networks, satellite imagery, or high-resolution meteorological inputs. The UK soft fruit industry, where 95% of strawberries are grown under polytunnels, renders satellite and UAV-based data collection impractical. Consequently, growers often possess only limited historical records: weekly yield figures, basic plot metadata, and externally sourced weather data.

Existing forecasting models trained on these limited inputs often exhibit systematic errors that contradict biological reality. For instance, models may predict yield spikes before flowering or continued production after harvest clearance because they lack an explicit representation of the crop's seasonal growth cycle. The core problem addressed is how to correct these systematic, context-aware errors in existing models without requiring additional data collection or model retraining.

Methodology

The authors propose a structured Large Language Model (LLM) agent framework that performs post-hoc correction on top of existing forecasting models. Rather than replacing the base model, the agent acts as a reasoning layer that refines predictions using agricultural domain knowledge encoded in a fixed toolset.

System Architecture

The system operates via a ReAct (Reason + Act) loop, iterating up to 10 times per prediction to refine the output. The pipeline consists of two main phases:

1. Knowledge Graph Construction: Historical training plots are encoded as nodes containing metadata and "curve shape features" (e.g., volatility, peak position, growth pattern). This allows the agent to retrieve analogous historical seasons via Dynamic Time Warping (DTW) similarity search.
2. Prediction Loop: For each test plot and time step, the agent receives a state vector containing the raw prediction, recent actuals, and historical context. It iterates through a tool library to diagnose and correct the prediction.

The Agent Toolset

The agent utilizes a suite of expert-informed tools to diagnose and correct forecasts:

• Similarity Search: Retrieves historical plots with similar curve geometries to provide analogical context.
• Bias Learning: Calculates directional bias by comparing recent predictions against lagged actuals, applying a decay factor for post-harvest periods to prevent hallucinated yields.
• Phase Detection: Segments the yield curve into operational phases (pre-season, ramping, peak, declining, ended). It enforces constraints such as zero yield during off-seasons or decay rates during the decline phase.
• Range Validation: Clamps predictions that violate empirical bounds (e.g., weekly growth rates exceeding historical percentiles).
• Apply Correction: Synthesizes outputs from the above tools using a priority cascade: Physical Limits > Phase Constraints > Statistical Fine-Tuning.
• Safety Verifier & Trajectory Evaluation: Acts as a sanity check to ensure corrections do not cause catastrophic drift and align with recent trends.

Persistent Memory

The system includes Cross-Plot Persistent Memory, where the agent maintains a "Position Bias Table" (recording errors at specific calendar milestones) and a "Plot History" archive. This allows the agent to learn systematic errors across the dataset and refine its heuristics (Meta-Strategy Reflection) over the course of the season.

Experimental Setup

The framework was evaluated on two datasets:

1. Angus Soft Fruits: A proprietary dataset of strawberry yields from 123 plots (polytunnel-grown) in the UK.
2. USDA Corn Harvest: A public dataset of weekly corn harvest progress across 38 US states.

Three base forecasting models were tested: XGBoost, Random Forest, and Moirai2 (a zero-shot foundation model). Three LLMs were evaluated as the reasoning engine: Llama 3.1 8B, Qwen 2.5 7B, and LLaVA 13B (a vision-language model included to test sensitivity to model choice).

Key Results

The agent-based refinement consistently improved forecast accuracy across all base models and datasets, with the magnitude of improvement correlating to the baseline error.

• Strawberry Dataset (Angus):
  • XGBoost: Llama 3.1 8B reduced Mean Absolute Error (MAE) by 20% (0.0131 to 0.0105) and Mean Absolute Scaled Error (MASE) by 56% (3.93 to 1.75).
  • Random Forest: Showed the most dramatic improvement, with MAE reduced by 28% and MASE by 66% (from 5.90 to 2.03), largely correcting the model's tendency to predict non-zero yields during off-seasons.
  • Moirai2: MAE reduced by 24% and MASE by 22%.
• Corn Dataset (USDA):
  • Improvements were more modest due to the lower baseline error of the base models but remained consistent. XGBoost MAE was reduced by roughly 4-5%, and MASE by 2.5%.
• Model Sensitivity: Llama 3.1 8B produced the strongest and most consistent corrections across all configurations. Qwen 2.5 7B followed with moderate gains. LLaVA 13B showed inconsistent performance, often degrading results (e.g., increasing error on corn), highlighting that the choice of refinement model is critical and that vision-language pretraining does not necessarily aid structured numerical reasoning.

Ablation Study

An ablation study confirmed that Phase Detection is the "load-bearing" component of the pipeline. Removing it caused MAE to increase by 72% on the strawberry dataset, rendering corrections unreliable. Other tools (Bias Learning, Range Validation) contributed incrementally but were not as critical as the ability to recognize the seasonal phase.

Significance and Claims

The paper claims that this work demonstrates the viability of LLM-based post-hoc correction as a practical solution for data-constrained agricultural forecasting. Key contributions include:

1. A New Paradigm: Moving beyond the assumption that better forecasts require more data or more complex base models. Instead, the approach reasons more effectively over existing, "messy" data.
2. Domain Knowledge Encoding: Successfully translating agricultural heuristics (seasonal phases, growth limits) into a structured toolset that an LLM can execute, rather than relying solely on the LLM's parametric knowledge.
3. Generalizability: The framework is dataset-agnostic. Without modification to tools or prompts, the same agent improved forecasts for both UK strawberries and US corn, proving it can handle different crops, scales, and geographies.
4. No Retraining Required: The method operates as a plug-in layer, requiring no retraining of the base forecasting model and no additional data collection.

The authors conclude that while the approach introduces computational overhead (due to the ReAct loop and LLM calls), it offers a deployable, explainable, and effective complement to existing forecasting pipelines, particularly where base models fail to capture seasonal biological constraints.