County-level soybean yield estimation based on Bayesian-CNN incorporating phenology dynamic
Topics: Remote Sensing
, Agricultural Geography
,
Keywords: Bayesian, CNN, deep learning, crop yield
Session Type: Virtual Paper Abstract
Day: Saturday
Session Start / End Time: 2/26/2022 08:00 AM (Eastern Time (US & Canada)) - 2/26/2022 09:20 AM (Eastern Time (US & Canada))
Room: Virtual 30
Authors:
Chishan Zhang, UNIVERSITY OF ILLINOIS URBANA-CHAMPAIGN
Chunyuan Diao, UNIVERSITY OF ILLINOIS URBANA-CHAMPAIGN
,
,
,
,
,
,
,
,
Abstract
Crop yield prediction is a fundamental way to study the relationship between crop yield, management practices, and environmental factors. Based on new advances in remote sensing technology, great progress has been made in this field by using the deep learning (DL) method, especially convolutional neural networks (CNN) and Long Short-Term Memory (LSTM). However, very few deep learning models quantifying the uncertainty associated with the predictions for crop yield have been reported. What’s more, the traditional deep learning approaches typically simplified the representation of the crop growth dynamics and integrated remote sensing and meteorological indices into a fixed timescale, which cannot reflect the phenological phases differentiation of crop response sensitivities toward environmental stress. To address these issues, a phenology dynamic based Bayesian-CNN was developed for soybean yield and uncertainty estimation. Based on the google earth engine, satellite data, heat stress variables, water stress variables, and soil properties variables are transferred to histogram-based tensors and used to train the Bayesian-CNN model. Then, the features and phenological phases' importance are analyzed by using the model inspection technique. The results of experiments show: (1) Incorporating vegetation phenology dynamics could significantly improve the model performance; (2) Multimodal data fusion in Bayesian CNN showed effectiveness in yield and uncertainty estimation; (3) Environmental variables improve the prediction accuracy but also increased predictive uncertainty. In this paper, we proposed a comprehensive phenology dynamic incorporating framework for the crop yield estimation and highlighted the deep understanding of prediction uncertainties and crop response sensitivities toward environmental stress.
County-level soybean yield estimation based on Bayesian-CNN incorporating phenology dynamic
Category
Virtual Paper Abstract
Description
This abstract is part of a session. Click here to view the session.
| Slides