Prediction of pine mistletoe infection using remote sensing imaging: A comparison of the artificial neural network model and logistic regression model


Usta A., Yilmaz M.

Forest Pathology, vol.53, no.1, 2023 (SCI-Expanded) identifier identifier

  • Publication Type: Article / Article
  • Volume: 53 Issue: 1
  • Publication Date: 2023
  • Doi Number: 10.1111/efp.12783
  • Journal Name: Forest Pathology
  • Journal Indexes: Science Citation Index Expanded (SCI-EXPANDED), Scopus, Academic Search Premier, Agricultural & Environmental Science Database, BIOSIS, CAB Abstracts, Environment Index, Geobase
  • Keywords: forest healthy, independent-samples t-test, Landsat 8, Scots pine forests, the predicted pseudo-probability, SOIL AGGREGATE STABILITY, WATER-CONTENT, SCOTS PINE, CANOPY, REFLECTANCE, TEMPERATURE, FORESTS, LEAF, STRESS
  • Karadeniz Technical University Affiliated: Yes

Abstract

© 2022 Wiley-VCH GmbH.In this study, the prediction of pine mistletoe distribution in Scots pine ecosystems was explored using remote sensing variables to compare the multilayer perceptron (MLP) artificial neural network (ANN) and logistic regression (LR) model performances. For this purpose, 109 sample plots were distinguished in pure Scots pine forests (natural) in the Eastern Black Sea Region of Turkey. Distinguishing mistletoe-infected stands (69) and uninfected stands (40) was performed with field observations. The variables acquired from Landsat 8 (Level 1) images were used as independent variables for independent-sample t-test, MLP ANN and LR models. Remote sensing variables indicated that mistletoe-infected stands were in drier areas with a lower vegetation-leaf area index. Based on the performance results of both models, the sensitivity (SEN), specificity (SPE), positive predictive value (PPV), negative predictive value (NPV) and accuracy of the MLP ANN model were superior to those of the LR model. The prediction percentages (SEN, SPE, PPV and NPV) of mistletoe-infected stands were better than the prediction percentages of uninfected stands. The prediction accuracies of LR and MLP ANN models were 74.3% and 89.6%, respectively. However, all remote sensing variables were included in the prediction equation of the MLP ANN model, while the thermal infrared 1 (TIRS1) variable was included in the LR model. In the MLP ANN model, the TIRS1 variable also had the highest normalized importance (100%). The area under the curve (AUC) value for identifying the mistletoe-infected stands of Scots pine forests used by the MLP ANN model (0.892 ± 0.034) was higher than in the LR model (0.838 ± 0.039), explaining the more accurate predictions obtained from the MLP ANN model. The MLP ANN model showed much better performance than the LR model. The results of this study are expected to make important contributions to the identification of potential mistletoe-infected areas.