Machine learning-based classification of Hyrcanian forests using PRISMA data
- 1Department of Forestry and Forest Economics, University of Tehran, Karaj, Iran
- 2Department of Agricultural, Food, Environmental, and Forestry Sciences and Technologies (DAGRI), University of Florence, Florence, Italy
Abstract
Forest type mapping is crucial for sustainable forest management, but field-based methods are costly and time-consuming. PRISMA hyperspectral imageryis promising for forest type classification. This study evaluates PRISMA data for mapping mixed Hyrcanian forests in northern Iran using classical supervised machine learning (ML) algorithms. Four classifiers, i.e. Random Forests (RF), Support Vector Machine (SVM), Neural Network (NN) and Linear Discriminant Analysis (LDA), were employed to analyze five datasets: (1) original PRISMA bands; (2) Minimum Noise Fraction (MNF)-transformed bands; (3) original bands combined with vegetation indices; (4) bands most correlated with selected components and (5) principal component–based bands within near-infrared and shortwave infrared regions. Results showed that the RF classifier with MNF-transformed data achieved the highest accuracy based on 5-fold cross validation, with an overall accuracy of 88.65% ± 1.02% and a kappa coefficient of 0.83, outperforming the original spectral bands (OA = 72.16% ± 2.22%, KC = 0.59). Among the other classifiers, SVM and LDA achieved their highest accuracies using Dataset 3 (OA = 73.58% ± 2.98%, KC = 0.61 and OA = 72.57% ± 1.76%, KC = 0.61, respectively), while NN showed a lower performance (OA = 66.72% ± 4.09%, KC = 0.52). The classification performed well for pure beech and mixed forest types, highlighting the potential of PRISMA hyperspectral data, particularly after MNF preprocessing, for forest type mapping in Hyrcanian forests.
Related
- Optical Remote Sensing and Machine Learning Enable Scalable Carbon Mapping in Himalayan Forests Despite Spectral Saturation Challenges
- How mixed are production-oriented planted forests? A global review and classification approach
- Sentinel-1 SAR time series for forest phenology: multimodal deep learning with meteorological fusion across deciduous broadleaf and evergreen needleleaf forests
- JUICE-MAJIS Earth observations during the 2024 gravity assist: first analysis and comparison with PRISMA data
- Machine-learning prediction of freezing temperature in Pb-contaminated soils remediated with biochar
BibTeX
@article{nejad2026machine,
title = {{Machine learning-based classification of Hyrcanian forests using PRISMA data}},
author = {Marjan Firoozy Nejad and Parviz Fatehi and Ali Asghar Darvishsefat and Elia Vangi},
journal = {European Journal of Remote Sensing},
year = {2026},
month = oct,
doi = {10.1080/22797254.2026.2748769},
url = {https://doi.org/10.1080/22797254.2026.2748769},
note = {CC BY},
}
RIS
TY - JOUR
TI - Machine learning-based classification of Hyrcanian forests using PRISMA data
AU - Marjan Firoozy Nejad
AU - Parviz Fatehi
AU - Ali Asghar Darvishsefat
AU - Elia Vangi
PY - 2026
DA - 2026/10/31
JO - European Journal of Remote Sensing
DO - 10.1080/22797254.2026.2748769
UR - https://doi.org/10.1080/22797254.2026.2748769
AB - Forest type mapping is crucial for sustainable forest management, but field-based methods are costly and time-consuming. PRISMA hyperspectral imageryis promising for forest type classification. This study evaluates PRISMA data for mapping mixed Hyrcanian forests in northern Iran using classical supervised machine learning (ML) algorithms. Four classifiers, i.e. Random Forests (RF), Support Vector Machine (SVM), Neural Network (NN) and Linear Discriminant Analysis (LDA), were employed to analyze five datasets: (1) original PRISMA bands; (2) Minimum Noise Fraction (MNF)-transformed bands; (3) original bands combined with vegetation indices; (4) bands most correlated with selected components and (5) principal component–based bands within near-infrared and shortwave infrared regions. Results showed that the RF classifier with MNF-transformed data achieved the highest accuracy based on 5-fold cross validation, with an overall accuracy of 88.65% ± 1.02% and a kappa coefficient of 0.83, outperforming the original spectral bands (OA = 72.16% ± 2.22%, KC = 0.59). Among the other classifiers, SVM and LDA achieved their highest accuracies using Dataset 3 (OA = 73.58% ± 2.98%, KC = 0.61 and OA = 72.57% ± 1.76%, KC = 0.61, respectively), while NN showed a lower performance (OA = 66.72% ± 4.09%, KC = 0.52). The classification performed well for pure beech and mixed forest types, highlighting the potential of PRISMA hyperspectral data, particularly after MNF preprocessing, for forest type mapping in Hyrcanian forests.
SN - 2279-7254
ER -