TRUSTWORTHY AI FOR PRECISION AGRICULTURE: WEATHER DATA IMPUTATION AND RETRIEVAL-GROUNDED CROP ANOMALY ADVICE

Awanish Khanal

Precision agriculture increasingly depends on computational systems that must be accurate, efficient,and trustworthy under real operational conditions. Weather stations provide the environmental context used for irrigation scheduling, frost protection, disease-risk modeling, and crop-stress assessment. At the same time, multimodal and retrieval-augmented AI systems are beginning to provide crop and plant anomaly explanations to human users. These two settings raise a shared systems question: how can agricultural AI remain reliable when inputs are incomplete, visual cases are ambiguous, and generated advice must be grounded in evidence rather than fluency alone? This thesis studies trustworthy AI for precision agriculture through two complementary contributions. The first contribution evaluates agricultural weather-data imputation as both a machine-learning and performance-engineering problem. Using high-frequency observations from three Jena, Germany meteorological towers and two AgWeatherNet stations in Sunnyside, Wash- ington, the study compares twelve imputation methods across statistical models, classical machine learning, and deep neural networks. The benchmark evaluates mean absolute error, root mean squared error, robustness under realistic synthetic outages and stress regimes, runtime, throughput, peak memory, GPU use, energy consumption, and monetary cost. The main result is that a tuned Random Forest provides the strongest overall default for operational gap filling. It achieves 0.46◦C MAE on the WS Saaleaue Jena tower and 0.72◦C MAE on the Sunnyside-In AgWeatherNet station, while also offering lower deployment cost than deep sequence models. Lightweight GraphSAGE–GRU embeddings provide modest accuracy gains, especially for the longer Jena record, but their additional GPU preprocessing cost makes them an optional enhancement rather than a default requirement. The second contribution develops and evaluates a trust-aware crop anomaly identification pipeline. The system combines known-class plant disease recognition, open-set anomaly routing, metadata-aware retrieval, source-bounded generation, claim-level citation verification, and a con- troller that can answer, answer cautiously, retrieve top-2 alternatives, provide unknown-anomaly triage, ask for better input, or abstain. Results show that supervised vision models reach strong known-class performance on controlled apple-leaf images, with DeiT-Small achieving 98.9% ac- curacy on AL9EE, but field-style robustness remains lower. A combined open-set gate improves unknown-anomaly detection to 0.904 AUROC and reduces forced disease diagnoses on unknown cases to 8.3%. For retrieval, BGE-small with metadata filtering and cross-encoder reranking achieves 0.84 P@5, 0.91 R@5, and 0.88 MRR. For generation, the trust-aware RAG configuration reaches 0.90 faithfulness and 0.91 citation correctness while reducing unsupported and risky outputs to 0.04 and 0.03, respectively. Together, the two contributions show that trustworthy agricultural AI is not a property of a single model. It emerges from reliable data repair, performance-aware deploy- ment, evidence-grounded generation, open-set routing, and explicit mechanisms for uncertainty and human review.

TRUSTWORTHY AI FOR PRECISION AGRICULTURE: WEATHER DATA IMPUTATION AND RETRIEVAL-GROUNDED CROP ANOMALY ADVICE

(1)