Research map / Reinforcement learning / Visual and representation RL
To Measure or Not: A Cost-Sensitive, Selective Measuring Environment for Agricultural Management Decisions with Reinforcement Learning
Read the paperCheck your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Applies recurrent PPO in a WOFOST-based crop environment where the agent jointly learns nitrogen fertilization and costly, feature-specific measuring decisions, showing adaptive measuring near flowering and better yield than standard practice under realistic costs.
Problem
Crop management RL usually assumes crop state observations are free and always available, but measuring crop features is labor-intensive and expensive, so optimal fertilization must be learned with selective, cost-sensitive measurements.
Approach
Formulates nitrogen management as an AFA-POMDP with explicit per-feature measurement costs in the reward, and trains an LSTM-PPO agent with masked observations to learn when to measure which crop feature alongside fertilization.
Key results
- Realistic-cost agent reached median yield 7.46 t/ha vs 7.30 t/ha for standard practice, versus 7.86 t/ha for No-cost.
- Higher measurement costs lowered yield: Exp-cost median 6.63 t/ha with largest variance.
- Under Realistic costs the agent measured LAI (5.5) and SM (4.8) most often and rarely TAGP/NAVAIL/Random (about 0.2-0.3 per year), with measuring concentrated near flowering.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Add realistic feature-measurement costs (ablation) | LSTM-PPO crop management agent · Netherlands winter wheat | yield (tons/ha), median: 7.86 -> 7.46 | Hurt |
| Raise flat measurement costs from 10 to 60 (ablation) | LSTM-PPO crop management agent · Netherlands winter wheat | yield (tons/ha), median: 7.59 -> 6.63 | Hurt |
| Use realistic feature-specific measurement costs instead of a flat cost of 10 (ablation) | LSTM-PPO crop management agent · Netherlands winter wheat | yield (tons/ha), median: 7.59 -> 7.46 | Hurt |
| Remove crop-state observations, retaining only DVS and weather (ablation) | Fertilization-only RL agent · Netherlands winter wheat | yield (tons/ha), median: 7.86 -> 7.21 | Hurt |
| Use standard-practice fixed-date fertilization instead of cost-sensitive measuring and control | LSTM-PPO crop management agent · Netherlands winter wheat | yield (tons/ha), median: 7.30 -> 7.46 | Helped |
| Measure at no cost rather than train with complete observations and no measuring task (ablation) | LSTM-PPO crop management agent · Netherlands winter wheat | yield (tons/ha), median: 7.86 vs 7.86 | No clear effect |
Surprising findings
- The agent measured the Random distraction feature as often as TAGP under the Realistic cost scenario, despite large cost differences.
- Under No-cost, measuring showed no clear temporal pattern.
Limitations
- Simulation only; no sim2real validation and noiseless, non-destructive, instantaneous measurements are assumed.
- Single crop, site and fixed genotype; retraining needed for different sites.
- No pretraining of LSTM-PPO, so performance could improve with pretraining.
Experimental setup
Datasets: NASA POWER weather data WOFOST simulated crop data
Benchmarks: WOFOST simulated Netherlands winter wheat (even years 1990-2022)
Metrics: yield (tons/ha) measurement action counts
Baselines: No-cost Flat-cost Exp-cost All-observed None-observed Standard-practice Random-spread
Builds on: AFA-POMDP PPO WOFOST CropGym ACNO-MDP
Closest papers by meaning
- Learning Long-Term Crop Management Strategies with CyclesGym (NeurIPS 2022 · similarity 0.89)
- When to Sense and Control? A Time-adaptive Approach for Continuous-Time RL (NeurIPS 2024 · similarity 0.82)
- AgMMU: A Comprehensive Agricultural Multimodal Understanding Benchmark (NeurIPS 2025 · similarity 0.81)
- Adaptive Sampling to Reduce Epistemic Uncertainty Using Prediction Interval-Generation Neural Networks (AAAI 2025 · similarity 0.80)
- Empowering and Assessing the Utility of Large Language Models in Crop Science (NeurIPS 2024 · similarity 0.80)
- Abstract Reward Processes: Leveraging State Abstraction for Consistent Off-Policy Evaluation (NeurIPS 2024 · similarity 0.80)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.