Machine Learning Prediction

Machine learning prediction uses examples from the past to learn a statistical relationship between input variables and an outcome. The trained model then applies this learned relationship to new inputs, seeking generalisation, rather than merely recalling the training examples.

Learning patterns from historical data

Historical observations are converted into measurable features, such as rainfall, river level, slope or temperature. In supervised learning, each training example also has a known target, such as flood occurrence or cyclone intensity.

• An algorithm selects a model and adjusts its internal parameters so that predicted outcomes approach the observed outcomes.
• The discrepancy is measured by a loss function, which training seeks to minimise across the examples.
• Depending on the target, the model may perform classification, such as predicting whether flooding will occur, or regression, such as estimating expected water level.

From the fitted model to a prediction

After training, the fitted mathematical relationship receives new feature values and produces an estimated outcome. It may return a category, a numerical value or a probability, depending on the task.

• A validation dataset helps select model settings and control overfitting, where a model learns training details but performs poorly on new cases.
• A separate test dataset estimates how well the final model works on previously unseen data.
• Prediction assumes that the new situation remains sufficiently similar to the conditions represented in the training data.

Use and limitations in disaster management

Models can combine historical hazard and environmental observations to support early warning, risk mapping, damage assessment and resource planning. Their outputs should inform, rather than automatically replace, expert judgement and official warning systems.

• Incomplete, biased or unrepresentative data can produce unreliable predictions for poorly represented places or communities.
• A learned association is not necessarily a causal relationship.
• Rare extremes, changing climate conditions and faulty sensors can weaken performance, so models require monitoring and updating.
• Uncertainty, false alarms and missed events must be evaluated because disaster decisions have high consequences.