In order to join virtual sessions, you must be registered and logged-in(Were you registered for the in-person meeting in Denver? if yes, just log in.) 
Note: All session times are in Mountain Daylight Time.

A Novel Spatiotemporal Deep Learning method for Crop Classification in the US using Sentinel-2 data

Authors: Yumiao Wang*, Wuhan University, Qingyun Du, Wuhan University
Topics: Agricultural Geography
Keywords: Crop classification, deep learning
Session Type: Poster
Presentation File: No File Uploaded


Accurate crop type maps are critical for various agricultural applications. As satellite image data keeps improving its spatial and temporal resolution, it has become a powerful support to create crop type maps. However, crop classification in large areas is still challenging due to the existence of spatial heterogeneities. This study aims to develop a deep learning model that considers spatiotemporal effects. Specifically, there are two major parts in our method: 1) Long Short-Term Memory (LSTM) layers to model the dynamic temporal dependencies in multi-spectral time-series information. 2) an attention mechanism to emphasize spatial heterogeneities. The experiment was carried out in the US, which is a major crop export country with extensive territory. With Google Earth Engine, we extracted time-series spectral information over 2016-2018 from Sentinel-2 as model inputs. Moreover, the location and climate data were also collected for modeling. We validated our model using the US Department of Agriculture's Cropland Data Layer (CDL). The results indicated our model could classify crops accurately. In addition, the results showed consistent performance over space with less spatial dependency, which means our model can provide a relatively accurate and robust crop classification in large areas.

To access contact information login