Anjola99/feijoa-regressor

Feijoa Model 2 - Quality Parameter Regressor

Purpose

This model card documents the scoped Agent 3 Random Forest regressor for predicting six feijoa post-harvest quality parameters from environmental, package, and color measurements.

Model

• Estimator: MultiOutputRegressor(RandomForestRegressor)
• Hyperparameters: n_estimators=300, max_depth=None, min_samples_split=5, min_samples_leaf=2, max_features='sqrt', random_state=42, n_jobs=-1
• Features: time_days, temperature, perforations, a_star, b_star, lightness_L
• Targets: weight_loss_pct, wvtr_g_m2_day, o2_pct, co2_pct, rh_pct, firmness_N
• Scaling: saved Phase 1 StandardScaler loaded from models/model2/scaler.pkl; it was not refit.

Test Metrics

target	r2	rmse	mae
weight_loss_pct	0.995967	0.121755	0.088991
wvtr_g_m2_day	0.992334	0.512590	0.378364
o2_pct	0.994867	0.192750	0.146789
co2_pct	0.994457	0.177318	0.126378
rh_pct	0.995488	0.211157	0.170330
firmness_N	0.998825	0.595307	0.435982

All R2 thresholds passed: True

Mean Feature Importance

rank	feature	importance
1	perforations	0.307760
2	lightness_L	0.175677
3	b_star	0.138072
4	temperature	0.131147
5	a_star	0.126363
6	time_days	0.120981

Artifacts

• random_forest.pkl: trained multi-output model
• scaler.pkl: Phase 1 feature scaler, preserved
• evaluation_metrics.json: test-set metrics and metadata
• feature_importance.png: averaged feature importance chart
• actual_vs_predicted_*.png: one test-set scatter plot per target