Using deep learning for land classification within the konza prairie, 1985 – 2011

Abstract

Machine learning has been around for decades, but deep learning is the new focus of study within machine learning. The goals of implementing deep learning into remote sensing are resulting in much faster and more accurate results for much larger datasets. The field of remote sensing has focused on increasing the accuracy of land classification. A possible solution for increasing accuracy is the use of convolutional neural networks. The goals of this study were to determine whether convolutional neural networks can be used on moderate-resolution imagery to accurately classify land. The study site of focus was the Konza Prairie in Geary County, Kansas. The image data are from Landsat 4 and 5 spanning the years 1985-2011. The Konza was split into 4 x 4-pixel size fishnet of cells that were classified as either burnt or non-burnt. To better examine the convolutional neural network, it was compared to machine learning and other neural network models. The machine learning models explored were logistic regression, k-nearest neighbor, decision tree, and linear support vector machine. The neural networks implemented included the basic neural network, shallow neural network, flatten time window neural network, convolutional neural network, and deep convolutional neural network. The results show that the k-nearest neighbor produce the highest overall accuracy compared to all the machine learning and neural networks but consist of high errors of omission proving that the classification is not represented accurately. The deep convolutional neural network has the best results for classifying burnt and non-burnt cells and low errors of omission and commission, which best represents the classification.