nasa-impact/marine_debris_ml ? reverse-engineered prompt

Build me a working marine debris detection project from satellite imagery.

I want to be able to take Planetscope style ocean images, train an object detection model on the labeled debris boxes, and then run inference on new tiles to find floating trash in coastal waters. Please wire up the data prep so the image scenes can be tiled, converted into TFRecords, and used for training with TensorFlow Object Detection. Then make the inference flow export predictions with bounding boxes, confidence scores, and GeoJSON output so they can be used in a map or dashboard.

Please include clear setup instructions, a simple way to train with the provided configs, and a way to test the model on sample images. If anything needs current docs or a missing dependency, look it up online and use the most stable approach. I mostly want this to run end to end without me having to figure out all the glue between the data, model, and inference pieces.