sum1lim / sea_ice_remote_sensing Goto Github PK

Write a script that predicts and visualizes the image segmentation using the classification models.

Using the SOBEL edge detection, segment images into objects and extract features per each object.

Currently, create-dataset script extracts data relying solely on randomness, thus train and test data can share the same source images. However, this does not allow us to test the accuracy fairly since the resulting model can be biased or overfitted on the features that indicate what image a sample came from. To solve this, test data should be extracted from the images that are not used for creating the training dataset.

After the package installation, python scripts are supposed to run as a command instead of python path/to/script, but Windows cannot find the commands. Find an ideal installation strategy for the Windows environment.

For reliable results, perform K-fold validation with K=10 [1].

[1] P. Refaeilzadeh, L. Tang, and H. Liu, “Cross-Validation,” Encyclopedia of Database Systems, pp. 532–538, 2009.

Enable CNN script to add additional attributes to the convolution layers, which might improve the performance of the model.

We want to test the effect of using metadata features for training on the model's performance. Run trials 1~5 without GLCM features.

Add a script to generate GLCM texture features.

Thresholding is important for object segmentation and extracting other stuff. Implement a strategy to accomplish this.

Features for machine learning should be extracted for each sample pixel. Derive a strategy for the extraction. Suggested features: RGB bands, year, month, day, AOI location, etc.

The expert data are provided in the mask directory of the provided dataset. Add functionality to generate the data distribution statistics. csv output preferred. Refrer to https://github.com/asylve/Sea-Ice

Implement a python script for SOBEL filter to segment images into objects with homogeneous pixels.

GLCM features do not work well with the regular neural network and degrade the performance of the model. It is suspected that the addition of the features makes the network too complicated, which prevents finding the global minimum of the loss function. Write a script for a 1D convolutional neural network as a potential solution.

Define a strategy for pixel-based sea ice classification using neural network

Add padding at the end of the feature sequences "to involve every possible combination into convolution" [1].

[1] R. Kestur, S. Farooq, R. Abdal, E. Mehraj, O. Narasipura, and M. Mudigere, “UFCN: a fully convolutional neural network for road extraction in RGB imagery acquired by remote sensing from an unmanned aerial vehicle,” Journal of Applied Remote Sensing, vol. 12, no. 01, p. 1, 2018.

Except for the training data, normalization should have standard min & max values instead of calculating such values within the dataset. The standard values are usually from the training dataset.

Add GLCM correlation features to the dataset.

Normalizing data can result in a better performance of the model. Try min-max normalization