RETA-ML

A Machine Learning Framework for Automated Anomaly Detection in Precision Agriculture Geospatial Data

Zachary Komarnisky · Felippe H. S. Karp — Olds College of Agriculture & Technology Presented at the 17th International Conference on Precision Agriculture (ICPA) / 11th Brazilian Congress on Precision and Digital Agriculture (ConBAP), 2026.

This is the public repo to the RETA-ML project. It ships the trained models and the framework code — enough to inspect the models, run them on your own sensor data, and build on the work.

---

What is the idea?

On-the-go agriculture sensors are fast and useful but produce errors or include data that is not desirable for further analysis, things like headland turns, overlaps, speed changes, localized and global errors. These quietly reduce the quality in precision-ag maps. Cleaning that data is manual, slow, and different for every sensor. We ask whether one model can learn cleaning once and apply it across sensors, and we measure exactly where that transfer holds and where it breaks.

Annotated data vs. ML-filtered data

(a) every expert-labeled point on a held-out field — Clean, Operational, Global and Local classes. (b) the same field after running it through the shipped random_forest.joblib: everything the model doesn't predict as Clean is filtered out (shown faint grey), leaving the value surface behind.

What we found (leave-one-sensor-out)

Trained on two sensors, tested on the held-out third, macro-averaged over three folds:

Class	RF P	RF R	RF F1	XGB P	XGB R	XGB F1
Clean	0.96	0.94	0.95	0.96	0.85	0.90
Operational	0.46	0.63	0.52	0.30	0.65	0.40
Local	0.44	0.24	0.30	0.32	0.38	0.34

Random Forest macro-accuracy 0.90, XGBoost 0.83. Clean data and operational errors (the kinematic ones) transfer across sensors; local spatial anomalies do not — they look different on every sensor, which motivates a spatially-explicit (graph) model next. (Global-outlier class excluded from reporting: only 14 examples across all fields.)

---

Repository layout

reta_ml/     the framework — preprocessing, features, models, evaluation
models/      trained models + a model card
app.py       Streamlit app to load a field, run a model, and inspect results
requirements.txt / environment.yml   dependencies

Trained models (models/)

File	What it is
random_forest.joblib	Random Forest (400 trees, class-balanced), the poster's headline model. joblib.load(...).
xgboost.json	XGBoost (400 trees, depth 6). XGBClassifier().load_model(...).
model_card.json	Feature list, config, training data, and per-class counts.

The RF/XGBoost models are trained on all expert-labeled points across the three sensors (49,665 points; the annotated dataset itself is not distributed in this repository) with the locked-in configuration: 13 sensor-agnostic features (field-relative motion + fixed-radius 15/30/45 m spatial z-scores), seed 42, no scaler.

---

Quick start

# 1. install
pip install -r requirements.txt          # or: conda env create -f environment.yml

# 2. use a trained model in a few lines
python - <<'PY'
import joblib, json
from reta_ml.load import load_table
from reta_ml.preprocess import preprocess_pipeline, recompute_value_stats
from reta_ml.lofo_benchmark import _feature_matrix

feats = json.load(open("models/feature_columns.json"))
rf = joblib.load("models/random_forest.joblib")

raw = load_table("your_field.gpkg", validate=False)   # bring your own sensor data
raw["value"] = raw["<your_value_column>"]
proc = preprocess_pipeline(raw, main_variable="value", local_stats_mode="fixedbands")
proc = recompute_value_stats(proc, main_variable="value", local_stats_mode="fixedbands")
pred = rf.predict(_feature_matrix(proc, feats))     # 0=Clean 1=Operational 2=Global 3=Local
print(pred[:20])
PY

# 3. or explore interactively
streamlit run app.py

---

Method, briefly

• Features (13, sensor-agnostic): field-relative speed / distance / acceleration, turn & short-segment flags, whole-field z-score, fixed-radius spatial z-scores at 15 / 30 / 45 m, and ratio to the transect mean. Deliberately relative and fixed-radius so the same feature vector is comparable across a soil sensor and a combine.
• Models: Random Forest and XGBoost (class-balanced), with a HeteroGAT graph neural network as the spatially-explicit next step.
• Evaluation: Leave-One-Sensor-Out — train on two sensors, test on the third — the honest test of cross-sensor transfer.

Citation

If you use this work, please cite the ICPA/ConBAP 2026 paper (Komarnisky & Karp, 2026).

License

Released under the MIT License.

Link to Github

https://github.com/Dozer3530/RETA-ML/tree/main

Acknowledgments

Supported by an NSERC CCI Mobilize Grant (CCMOB-2024-00038), the Olds College Office of Research Services, and the Werklund School of Agriculture Technology.