Skip to content

UCSB Activity

Jump to bottom
samiradaneshgar edited this page Nov 8, 2019 · 35 revisions

11/08/2019

The quality of the training data was improved using the second u-net model. A ticket was opened on XSEDE regarding the issue of the second u-net model on Bridges. The second u-net model provides results with more than at least 70% probability compare to the first u-net model. AGU poster must be prepared. I need to work on the information needed for CoPI status on Dr. Chu's grant.

08/23/2019

Pan-sharpened WV images of the southwestern Greenland will be requested from PGC today. Once I receive them, they will be transferred to bridges to be processed by the prediction program.

I have updated the Matlab programs and train_unet.py and predict.py scripts so they accept input arguments. I am able to run all of the programs on Bridges. I ran train_unet.py considering some of the suggestions from Bento regarding the initial weights and batch size. I still need to add more augmentations. I have uploaded about 30 images to the Bridges to test the algorithm on them.

I am also working on the area threshold which should be applied to the outputs to remove the slush and disconnected pieces of the rivers. I will continue working on the draft of the paper next week.

07/12/2019

We decided to move forward with the current weights, and try the code on WV images of different years in order to generate river map series and assess the performance of the model in delineation of the rivers from different images. We are currently focusing on the science goals and the structure of the paper.

05/31/2019

Samira Daneshgar: More training data has been generated. No data issue has been found with Near Infrared bands of WV during 2014 and 2015 while we were testing the trained U-Net algorithm. Therefore, we decided to use other bands instead for training. The results have been improved.

05/17/2019

Samira Daneshgar: Trying to improve the training data and define the best threshold for the output watermask to preserve connected rivers and get rid of slush.

Enbo Zhou: Pull request for SegNet codes has been open. Doing experiments in bridges. Developing postprocessing codes for RiverNet.

05/03/2019

Enbo Zhou: SegNet Codes are refactored and ready for pull request. Develop postprocessing codes. Prepare to transfer codes to bridges.

04/04/2019

Samira: U-Net river extraction algorithm was initially trained with 8-band WorldView satellite imagery. However, the results have been improved when the model was trained with 3-band imagery recently. Using the implemented U-Net algorithm, it only takes about 2 minutes to process and extract rivers from a WorldView satellite imagery.

03/22/2019

Enbo Zhou: A RiverNet adding more ingredients to SegNet is designed and trained for river classification. Some postprocessing procedures still need.

03/01/2019

Samira: Different layers of data has been passed through the algorithm. The average test accuracy of 85% was achieved without fine-tuning.

Enbo ZHou: Preliminary river extraction results have been gotten. Next step fine-tune the network and generate the vector file of river network.

02/15/2019

Samira: A Keras based deep CNN architecture is employed to perform pixel based segmentation of river and non-river in the WorldView satellite imagery. The model was trained on a small training set which was already digitized manually, each with 8 bands and 2 classification masks. Image augmentation is used for input images to increase training data. Training was done on a NVIDIA Tesla P100-16GB GPU. Accuracy will be updated soon.

01/02/2019

Enbo Zhou: Implement the Segnet to conduct image semantic labelling, i.e. remote sensing classification. Experiments using datasets provided by ISPRS and Greenland WorldView Images are now carried in Servers, which need a long time. The method's accuracy can achieve 85% now.

12/07/2018

Enbo Zhou: Prepare the training data to test pixel-based classfication.

Build a CNN network to train the classfication model.

Potential Pipeline: One image is fed into one pipeline. Building different scale's image --> Feature extraction for every pixel --> classify each pixel -> smooth the classfication result -->large river and small streams extraction.

10/26/2018

Initial image classification results using a pretrained DenseNet201 convolutional neural network.

10/12/2018

In the second attempt, windows of 227 × 227 pixels were defined in the WorldView2 satellite imagery. The defined 332 regions of interest (ROI’s) were consisted of four classes:

  1. Large river (134 ROI’s)
  2. Small river (119 ROI’s)
  3. Crevasse (40 ROI’s)
  4. Ice/slush (39 ROI’s)

We fine-tuned a pretrained AlexNet convolutional neural network to perform classification in the WorldView2 satellite imagery. The network requires input images of size 227-by-227-by-3 (3 is the number of color channels), and outputs a label for the feature in the image together with the probabilities for each of the classes.

We performed the following steps:

  • Loaded the 333 images as an image datastore.
  • Divided the data into training (80%) and validation (20%) data sets.
  • Replaced the last three layers of the pretrained AlexNet with a fully connected layer, a softmax layer, and a classification output layer.
  • Set the fully connected layer to have the same size as the number of classes (four classes).
  • Increased the WeightLearnRateFactor and BiasLearnRateFactor values of the fully connected layer to 20.
  • Set the initial learning rate to a small value of 0.0001.

We reached the classification accuracy of 91% on the validation set.

09/28/2018

The goal of the first attempt was to design a supervised method to estimate probabilities of actively flowing supraglacial streams and glacio-hydrological features such as crevasses, slush, and ice in high-resolution multispectral WorldView2 satellite imagery collected over the Greenland ice sheet ablation zone on July 2012.

Firstly, 565 windows of 25 × 25 pixels were defined in the WorldView2 satellite imagery. The defined 565 regions of interest (ROI’s) were consisted of six classes:

  1. Large river (170 ROI’s)
  2. Lake (28 ROI’s)
  3. Small river (139 ROI’s)
  4. Crevasse (35 ROI’s)
  5. Dust/slush (37 ROI’s)
  6. Ice/slush (156 ROI’s)

For each ROI (25 × 25 pixels) that we defined in the WorldView2 satellite imagery, we extracted the mean and standard deviation of the following 21 features:

  • Multi-spectral signatures (8 features)
  • Modified normalized difference water index (MNDWI) (one feature)
  • Gaussian filtering in four scales (four features)
  • Morphological transformation (six features)
  • Phase congruency (one feature)
  • Hessian (one feature)

Therefore, a set of 42-D feature vectors was generated for each ROI. Then, we used the Neural Network Pattern Recognition app in MATLAB https://www.mathworks.com/help/deeplearning/gs/classify-patterns-with-a-neural-network.html to classify input data into a set of target categories. Input data to the Network were 565 42-D feature vectors and the target data of the Network were 6 classes of large river, lake, small river, crevasse, dust/slush, ice/slush. Input vectors and target vectors were randomly divided into three sets as follows:

  • 70% are used for training.
  • 15% are used to validate that the network is generalizing and to stop training before overfitting.
  • The last 15% are used as a completely independent test of network generalization.

After we trained the Network, we evaluated the performance of the Network on some test areas of 50 × 50 pixels. The output probabilities indicated the high accuracy of the trained Network in detection of large river, lake, and small river.

Clone this wiki locally