Cloud Detection

The goal of cloud detection is to develop a classification model that effectively distinguishes the presence of cloud over the presence of ice/snow surfaces.

Usage

Clone the repositories

All works are done in cloudDetection.Rmd and CVgeneric.R Every R chunk has a brief introduction

$ git clone https://github.com/caojilin/cloud-detection
$ cd cloud-detection/cloudDetection.Rmd

Load Required Libraries / Packages

library(caret)
library(ggplot2)
library(scales)
library(e1071)
library(class)
library(randomForest)
library(ggfortify)
library(MASS)
library(corrplot)
library(rpart)
library(PRROC)
library(rpart.plot)

1 Data Visualization

• Load datasets and add column names
• Add a column identifying which image they are from (for later convenience when splitting and merging data)
• Remove unlabeled values (zeros) and convert -1 to 0
• Combine data frames for all three images into one
• Plot well-labeled map of the images based on x and y coordinates

# E.g. Image 1
ggplot(data = image1) + geom_point(aes(x = x_coordinate, y = y_coordinate, color = as.factor(expert_label))) + scale_y_reverse() + scale_color_manual("Expert Labels", labels = c("Not Cloud", "No Label", "Cloud"), values = c("grey", "black", "white")) + ggtitle('Image 1: MISR Orbit 13257')

• Observe pair-wise relationships between variables by creating correlation plots on all features

# E.g. 
corrplot(cor(images_all[,c(7, 8, 9, 10, 11)]), 
         method = 'number', 
         type = 'lower',
         tl.srt = 45,
         title = "Correlation For Radiance Angles")

• Observe relationships between different features and expert labels by creating histograms of individual variables color-coded with expert labels

# Distribution of CORR In Relation to Expert Label
ggplot(data = image1_nozero) + geom_histogram(aes(x = image1_nozero$CORR, fill = expert_label))

2. Data Preprocessing

a) Split the entire dataset using two non-trivial methods Method 1: split each image into 16 squares, extract 80% training, and 20% testing from each square Method 2: split each image into many smaller squares and randomly select some squares to be testing and others to be training b) a baseline classifier and its accuracy c) generating PCA features d) generic CV function, see CVgeneric.R Input: classifier is a string such as 'svm', 'logistic' training features, training labels, number of folds K and a loss function(optional) Output: a dataframe containing accuracy and loss (if given a loss function) in each fold

3. Modeling

In this section,

• Logistic Regression, LDA, QDA, knn, svm, decision tree
• ROC curves for each model
• Precision Recall curves for each model

4. Diagnostics

• A for loop for finding the optimal the number of trees in randomforest model. For each i in given ntree, train the randomforest model and evaluate the accuracy on validation dataset.
• A function for generating maps where the misclassification pixels are plotted in different color for further investigation