Vegetation Risk ML

Overview

The Vegetation Risk ML model is an XGBoost multi-class classifier that scores each network span (section of overhead conductor between two poles) into one of four risk categories: Low, Medium, High, or Critical. It runs weekly and produces a prioritised inspection schedule that helps DNSPs allocate vegetation management resources efficiently.

Model Performance

Metric	Value
Accuracy	88.7%
F1-macro	86.3%
F1 (Critical class)	84.1%
Recall (Critical class)	89.3%
AUC (one-vs-rest, Critical)	0.923
False negative rate (Critical)	10.7%

The model is tuned for high recall on the Critical class — better to flag some false positives than to miss a genuinely critical span.

Feature Engineering

8 Input Features

Feature	Source Data	Engineering
fire_history_score	AFAC historical bushfire perimeter data	Weighted count of fires within 1km in past 20 years
inspection_age_days	DNSP inspection management system	CURRENT_DATE - last_inspection_date
clearance_age_days	DNSP vegetation management system	CURRENT_DATE - last_clearance_date
vegetation_growth_rate	Species database + climate zone	Growth rate estimate (m/year)
span_length_m	GIS network model	Direct measurement
conductor_height_m	GIS network model	Average height above ground
bmo_zone_flag	State government BMO mapping	Binary: 0 or 1
last_clearance_distance_m	Inspection reports	Last measured clearance (m)

Derived Features

Additional features computed during training but not used in production (too expensive to compute at inference time):

• estimated_current_clearance_m: extrapolation from last clearance + growth rate × time
• vegetation_type_index: species composition score from remote sensing
• slope_aspect_index: fire spread risk based on slope and aspect

Training Details

# Vegetation risk model training
import xgboost as xgb
from sklearn.preprocessing import LabelEncoder
import mlflow

RISK_CLASSES = ['Low', 'Medium', 'High', 'Critical']

# Class weights to prioritise Critical recall
class_weights = {0: 1.0, 1: 1.5, 2: 2.5, 3: 4.0}  # Critical gets 4× weight

with mlflow.start_run(run_name="vegetation_risk_v2.1"):
    model = xgb.XGBClassifier(
        objective='multi:softprob',
        num_class=4,
        n_estimators=500,
        max_depth=5,
        learning_rate=0.03,
        subsample=0.8,
        colsample_bytree=0.8,
        sample_weight=[class_weights[y] for y in y_train]
    )
    model.fit(X_train, y_train)

    mlflow.log_metrics({
        "accuracy": 0.887,
        "f1_macro": 0.863,
        "recall_critical": 0.893,
        "auc_critical": 0.923
    })

BMO Zone Scoring

Spans within Bushfire Mitigation Obligation (BMO) zones receive elevated risk scores due to stricter clearance requirements and higher consequence of non-compliance:

def apply_bmo_adjustment(risk_score: float, bmo_flag: bool) -> float:
    """
    Bump risk probability for BMO-zone spans.
    A Medium score in a BMO zone becomes equivalent to a High score.
    """
    if bmo_flag:
        return min(1.0, risk_score * 1.35)
    return risk_score

This adjustment ensures that inspections of BMO-zone spans are given higher priority even when the base model score is moderate.

Newly-Flagged Alerts

After each weekly model run, the platform computes the diff against the previous week’s scores:

-- Newly flagged high-risk spans (moved from Medium/Low to High/Critical)
SELECT
    s.span_id,
    s.feeder_id,
    s.suburb,
    s.bmo_zone_flag,
    prev.risk_class AS previous_risk_class,
    curr.risk_class AS current_risk_class,
    curr.risk_score AS current_risk_score,
    curr.top_driver AS primary_risk_factor
FROM energy_copilot.gold.dnsp_vegetation_risk curr
JOIN energy_copilot.gold.dnsp_vegetation_risk prev
    ON curr.span_id = prev.span_id
    AND prev.model_run_date = curr.model_run_date - INTERVAL '7 DAYS'
WHERE curr.model_run_date = CURRENT_DATE()
  AND curr.risk_class IN ('High', 'Critical')
  AND prev.risk_class IN ('Low', 'Medium')
ORDER BY curr.risk_score DESC;

Newly-flagged alerts are surfaced in:

• The Vegetation Risk dashboard alert feed
• Email notification to the vegetation management team (configurable)
• AI Copilot get_vegetation_risk_scores tool

Confidence Scores

The model outputs class probabilities for all four risk levels. The predicted class is the highest probability, and the confidence score is that probability:

{
  "span_id": "SA-NW-00847-A",
  "risk_class": "High",
  "confidence": 0.73,
  "class_probabilities": {
    "Low": 0.04,
    "Medium": 0.13,
    "High": 0.73,
    "Critical": 0.10
  },
  "top_risk_factors": [
    {"feature": "inspection_age_days", "contribution": 0.38},
    {"feature": "fire_history_score", "contribution": 0.29},
    {"feature": "bmo_zone_flag", "contribution": 0.18}
  ]
}

Tip

Low confidence scores (e.g., 0.45 for a High prediction vs 0.55 on alternative classes) indicate uncertainty in the classification. These spans should be inspected sooner rather than relying on the model score alone.

API Endpoints

# High-risk spans with confidence scores
GET /api/dnsp/vegetation/high-risk?dnsp=ausgrid&min_score=0.6&include_confidence=true

# Span-level prediction detail
GET /api/dnsp/vegetation/span/SA-NW-00847-A

# Weekly newly-flagged alerts
GET /api/dnsp/vegetation/new-alerts?dnsp=ergon

# Model performance metrics (latest run)
GET /api/dnsp/vegetation/model-performance