Authors: Hao Hu*, University of Illinois at Urbana-Champaign, Shaowen Wang, University of Illinois at Urbana-Champaign
Topics: Spatial Analysis & Modeling, Geographic Information Science and Systems, Agricultural Geography
Keywords: crop yield, recurrent neural network, deep learning, weather, remote sensing
Session Type: Paper
Start / End Time: 1:20 PM / 3:00 PM
Room: Astor Ballroom I, Astor, 2nd Floor
Presentation File: No File Uploaded
Accurately estimating large-scale crop yields can bring numerous benefits to scientific research and practical applications. Traditional approaches to estimating crop yields at large-scale are usually done by calibrating regression models with predictive variables such as weather data, spectral information from remote sensing data, or a combination of them. One major issue associated with these traditional approaches is that the temporal characteristics of predictive variables are not fully leveraged. Particularly, time series observations of these predictive variables are often treated as independent observations as model inputs without accounting for the potential accumulative effects. In this work, we propose a deep learning method, namely long short-term memory (LSTM) neural networks, to fully utilize sequential and temporal characteristics of the predictive variables for crop yield estimation. Our assumption is that many hidden, deep connections between the time series of predictive variables and final crop yield cannot be revealed by simple linear or nonlinear regression. The proposed method is capable of estimating crop yields before harvesting, which can generate huge value for crop price projection and assessment of crop insurance. To better illustrate our method, we use weather data and spectral information from remote sensing imagery to predict corn yield in Midwest US at county-level with label information provided by National Agricultural Statistics Service (NASS) of the U.S. Department of Agriculture (USDA).