Testing machine learning algorithms on a binary classification phenological model

Aim Phenological models have become a vital tool for predicting future phenological responses to global climate change. Recently, machine learning (ML) has been used successfully to develop phenological models based on ground observations. However, fitting an observation series that has been observe...

Full description

Saved in:
Bibliographic Details
Published in:Global ecology and biogeography Vol. 32; no. 1; pp. 178 - 190
Main Authors: Dai, Wujun, Jin, Huiying, Zhou, Lin, Liu, Tong, Zhang, Yuhong, Zhou, Zhiqiang, Fu, Yongshuo H., Jin, Guangze
Format: Journal Article
Language:English
Published: Oxford Wiley Subscription Services, Inc 01.01.2023
Subjects:
Algorithms
autumn
binary classification
biogeography
Classification
Climate change
ecophysiological model
ecophysiology
Flowering
ground observations
Learning algorithms
Machine learning
Model accuracy
model validation
phenological model
Phenology
plant phenology
prediction
Predictions
Root-mean-square errors
sample size
Spring
Spring (season)
ISSN:1466-822X, 1466-8238
Online Access:Get full text
Tags: Add Tag
No Tags, Be the first to tag this record!
Description
Summary:Aim Phenological models have become a vital tool for predicting future phenological responses to global climate change. Recently, machine learning (ML) has been used successfully to develop phenological models based on ground observations. However, fitting an observation series that has been observed for only a few years (or even a decade) can easily lead to overfitting, and it is still a great challenge to predict future phenology accurately based on a short observation series. Here, based on historical ground phenological observations, we construct an ML‐based binary classification phenological model that can be applied to temperate trees. Innovation We thoroughly describe the construction process of a species‐specific ML‐based binary classification phenological model that is suitable for phenological predictions in both spring and autumn. Through experiments, we evaluate 18 commonly used ML classification algorithms and the effects of two parameters on the prediction performance of the model. Finally, we compare the performance between the binary classification model and six widely used ecophysiological models for accuracy of spring phenological prediction. Main conclusions The median root mean square error (RMSE) of the binary classification model for the first flowering date (93 observation series) was only 2.99 days, which proves that it is a good method of phenological prediction for temperate trees. This model can effectively overcome the insufficient sample size of ground observations for specific species and provide new insights of phenological predictions for tropical and subtropical trees. The comparison of different ML algorithms showed that the median root mean square errors of the six algorithms were <3.5 days, and lower than those of six ecophysiological models (>3.7 days) in prediction of spring phenology. In addition, this model can help us to infer plant physiological mechanisms and drivers of phenological change and provide more accurate predictions of plant phenological responses to climate change.
Bibliography:Handling Editor
Stephanie Pau
ObjectType-Article-1
SourceType-Scholarly Journals-1
ObjectType-Feature-2
content type line 14
content type line 23
ISSN:1466-822X
1466-8238
DOI:10.1111/geb.13612