Computer vision: detecting 2D structures (filaments) in satellite images

The task is to detect and somehow measure certain 2D structures called “filaments” that are narrow finger-like streams and in our case are oriented mostly diagonally (from upper right to lower left). The preferred software is OpenCV and its Microsoft .NET implementation EMGU [login to view URL] . You can use any software in the background but the front end should be a Visual Studio (VS2015 is preferred) solution of Windows Console application using C#. If you must use Python then it is OK but the front end must be a C# console application. Satellite images are 2D matrices of UINT8 numbers in either HDF4 or netCDF formats. For startup I can provide PNG (grayscale or RGB) bitmaps but the final solution must read datasets in HDF4 or netCDF. OpenCV and other software have extensions for reading HDF4 or netCDF datasets. Filaments are finger-like structures that extend from coast to offshore and are recognized by their COLD signature in SST (sea-surface temperature) images and HIGH values of CHL (chlorophyll) images. They can also bend and form eddies. Regions inside filaments have low SST (cold) and high CHL. Filament edges have HIGH gradients. Using typical color lookup palette filaments typically are blue in SST and green to yellow in CHL images – however, just using color or thresholding by value is NOT sufficient to detect filaments and their spatial structure (shape) must be used. I can easily do thresholding and gradient finding with my existing software. There is often cold and high CHL water along the coast that is NOT filaments. Along with the shape detection we also need some measures of the filaments, e.g. dimensions, directions, etc and some measure of intensity. An important issue is that most 2D datasets are only partly valid as clouds cover parts of the images. These invalid pixels (typically white or 255 in UINT8 datasets) have to be eliminated from the analysis. Note that the narrow areas BETWEEN filaments can be separated from filaments because the values inside must have low SST and high CHL. I have thousands of images. Once the method is adapted on sample images, I need to apply it to thousands of images, i.e. the method MUST be scalable to MANY images. From you I need a concrete plan: 1. Can you read 2D UINT8 datasets in HDF4 or netCDF? 2. Which methods (in OpenCV or other libraries) are you going use to detect and measure filaments? 3. What are the milestones that you suggest for the project? I don’t like videos and prefer a Visual Studio solution that does something useful. Link to sample images and task as Word file: [login to view URL]!AvbCFsQrBn0Sgm-_mnKuBqw_5xAF Note: Filaments are those colored fingers going away from coast but not the possibly similarly colored areas ALONG the coast. Colors are created using a particular color palette (lookup table) and is not fundamental. Fundamental is to use the UINT8 numbers that are the true variable.