hanxLi/Ghana-Drone-Image-Processing

# Ghana Drone Image Processing _Hanxi Li, Clark university_ ## Intorduction This repository contains the image processing pipeline for the Ghana Drone Images. The pipeline is for rasterize the image based on the bounding box geojson, and then create label rasters from the labels geojson and the range of bounding box, then chip it based on locations of targeting classes. ## Data - The original data is in geotiff format - The bounding box and crop type labels are in geojson formats - There are a total of 182 drone image tiles - Each dron image tiles contain 3 channels as Red, Green and Blue. However there are tiles contians the alpha channel. In this pipeline the alpha channel is ignored when processing the chips - The primary crop types are Maize and Fallow. Using one of these as the targeting crop or using both to perform multi-crop classification is recommended - The resulting image chips are 3 * 256 * 256 by default, with overlaps of 32 pixels - Here is the coverage distribution of the labels after cropped to the bounding box area: - ## Environment Setup - Change directory to the root directory of the cloned repo - Make sure Conda or Miniconda is installed on your device - RUN ```conda env create -f environment.yml``` - RUN ```jupyter lab -ip=0.0.0.0``` to launch jupyterlab - RUN the preProcessing notebook from the beginning - RUN the chipping notebook form the beginning ## Features ### Pre-Processing - Converting the low readability naming format of the original data into new format with grid name of the tile as the identifying name - Generate new catalog that replaces naming in the original catalog ### Chipping - User could define the primary crop type they wish to target. - Both binary classification chips and multi-class classification chips are available to generate - Parameter "Positive_Class_threshold" is used and can be defined by the User to better filter out chips - Various output formats including geotiff (tif), numpy compressed file ( …