Mapping miscanthus using multi-temporal convolutional neural network and google earth engine

Authors: XIN, YANAN - Pennsylvania State University; Adler, Paul

Submitted to: Workshop Proceedings
Publication Type: Proceedings
Publication Acceptance Date: 9/27/2019
Publication Date: 11/11/2019
Citation: Xin, Y., Adler, P.R. 2019. Mapping miscanthus using multi-temporal convolutional neural network and google earth engine. In proceeding of the 3rd ACM SIGSPATIAL International Workshop on AI for Geographic Knowledge Discovery. 81-84. https://doi.org/10.1145/3356471.3365242.
DOI: https://doi.org/10.1145/3356471.3365242

Interpretive Summary: No Interpretive Summary is required for this Proceeding. JLB.

Technical Abstract: Grasslands play an important role in ecology and agriculture. Accurately mapping the grasslands at a large scale is essential for productivity monitoring, policymaking, and environmental assessment. The advancements in remote sensing and machine learning technologies have enabled the generation of high accuracy national level crop layers. Although the national crop layer for the US includes grasslands, it doesn’t differentiate them into species. To fill the gap of mapping grasslands to the species level at the national scale with high accuracy, we propose a Convolutional Long Short-Term Memory (Convolutional-LSTM) neural network model for grass identification using multi-temporal Sentinel-2 images. Miscanthus is used as a case study for this short paper. The classification of Miscanthus using our model has yielded a 99.8% accuracy, which is significantly higher than the 92% accuracy produced by a benchmark model 3-layer fully connected neural network. Additionally, we leveraged the efficiency and effectiveness of cloud computing practices by implementing the entire analytical process in a cloud-based environment.