A farm collects overlapping images of a 100-acre field using tractor-mounted cameras and has labeled images of common weed classes. They need a model that detects specific weed types and their locations in the field and will serve real-time inferences on SageMaker endpoints. Which approach will provide accurate detections of weed types and positions?

The correct approach is to convert images to RecordIO and store them in S3, then train an object detection model using a Single Shot Detector (SSD) in SageMaker. The problem requires detecting specific weed types and their locations within the images, which is a classic object detection task. An SSD model is well-suited for real-time inference due to its efficiency. RecordIO is a common and efficient data format for training deep learning models on AWS, especially for image data, as it optimizes I/O operations. Image classification models (options 1 and 4) only assign a single label to an entire image, failing to provide location information for individual weeds. Apache Parquet (options 2 and 4) is a columnar storage format typically used for structured or semi-structured data, not optimized for raw image data in this context.