Regression models to predict continuous crop yield from sensor history is the correct choice because the cooperative needs to estimate a continuous numeric quantity for each season from historical sensor data.

Regression models fit this problem since supervised learning maps input features to a numeric target and time series or aggregated features from five minute telemetry can be used as predictors. You can perform feature engineering to summarize seasonal and field level signals and then train models such as linear regression, tree ensembles or neural networks using platforms like Amazon SageMaker to produce yield estimates.

Clustering of unlabeled sensor data to surface anomalies is unsupervised and it groups or highlights outliers rather than producing a labeled numeric prediction so it will not directly provide harvest tonnage estimates.

Reinforcement learning to dynamically optimize irrigation schedules is aimed at learning control policies from rewards and it is appropriate for decision making tasks rather than for supervised prediction of a continuous outcome from historical sensor features.

Image classification to detect leaf diseases in drone images targets visual inputs and categorical labels and it is unrelated to forecasting a numeric harvest quantity from tabular time series sensor data.

Full AWS Practitioner Certification Question