Optimizing deep learning architectures for remote sensing image analysis
Topics: Land Use and Land Cover Change
, Remote Sensing
, Geographic Information Science and Systems
Keywords: Convolutional Neural Network, Land Use Land Cover, Deep Learning
Session Type: Virtual Poster
Day: Wednesday
Session Start / End Time: 4/7/2021 01:30 PM (Pacific Time (US & Canada)) - 4/7/2021 02:45 PM (Pacific Time (US & Canada))
Room: Virtual 52
Authors:
Matthew T DeWitte, Department of Geography & Anthropology University of Wisconsin-Eau Claire
Rahul Gomes, Department of Computer Science
Mohan Pavithra Devy, Department of Computer Science University of Wisconsin-Eau Claire
Papia F Rozario, Department of Geography & Anthropology University of Wisconsin-Eau Claire
,
,
,
,
,
,
Abstract
Deep Learning tools have become very efficient in high resolution image analysis compared
to traditional classification models. One such example is the implementation of semantic
segmentation using a Convolutional Neural Network (CNN). Unlike image labeling, we can
use semantic segmentation to identify different class labels in an image. CNNs also facilitate
an object-oriented approach when several types of raw data are used as input. This makes
CNN an ideal tool for Land Use Land Cover (LULC) modelling. The dynamicity of Land Use
Land Cover (LULC) change forms an integral part of land management and optimization.
This information is essential in monitoring and assessing the various impacts of human
activities on natural resources. This project attempts to create a deep learning architecture
by reducing complex mathematical operations that plague the deployment of CNN in
resource-constrained environments. Preliminary results indicate that there is a potential to
achieve higher accuracy by using optimized kernel variants along with pre-trained weights
and variable dilation rates for image processing. The proposed model will be used to
compare and assess the LULC change dynamics for the lower Chippewa Valley watershed
region in Wisconsin.
