Detecting the presence of clouds using satellite image data

Task

• Modeling cloud detection in polar regions based on radiances recorded automatically by the MISR sensor aboard the NASA satellite Terr

Data

• 3 satellite images
• “Expert labels” used for model training for each point in image
• Features
  • NDAI, SD, CORR (features based on subject knowledge, see Yu2008.pdf)
  • Radiance angles (DF, CF, BF, AF, AN, see http://www-misr.jpl.nasa.gov/)

Data Processing

• Created 9 cross validation sets by splitting each image into 9 approximately equal sized blocks
• Each training split of a cross validation set contains 7 of 9 blocks of each image
• Each testing split of a cross validation set contains 2 of 9 blocks of each image
• Extracted feature sets of 4 closest neighbors for each point (neighbors)

Cloud Repo Data

• Cross validation sets contained within Cloud/src/main/resources/
  • Testing splits contained in Cloud/src/main/resources/test
  • Training splits contained in Cloud/src/main/resources/train
• Cross validaiton sets using neighbor data contained within Cloud/src/main/resources/
  • Testing splits contained in Cloud/src/main/resources/neighborTest
  • Training splits contained in Cloud/src/main/resources/neighborTrain
  • n1 contains features and label for a point
  • n2 contains features for the closest neighbor
  • n3 contains features for the 2nd closest neighbor
  • n4 contains features for the 3rd closest neighbor
  • n5 contains features for the 4th closest neighbor

Neural Networks

• Cloud.java: Contains code for a regular feed forward network using the 8 features using MultiLayerNetwork
• Cloud_neighbor.java: Contains code for a feed forward of 8 original features + features of closest neighbors using ComputationGraph + mergeVertex