<p>Forest inventory plays a key role in sustainable forest management by providing essential information on forest structure and composition. In this context, remote sensing data supplement field surveys, enabling the large-scale, consistent and cost-effective monitoring of forest inventories, as well as structural and compositional features. To improve this process, this study explores the fusion of two complementary remote sensing sources. Airborne laser scanning (ALS) and multi-temporal Sentinel-2 imagery. The goal is to improve the estimation of two important forest indicators: basal area and dominant tree species at the stand level. ALS point clouds offer detailed, three-dimensional representations of forest structures. These can be used to estimate attributes such as canopy height, crown geometry, biomass and basal area, through dedicated processing and modeling methods. However, their ability to discriminate between species depends on species-specific canopy architecture and growth strategies. The similar structural organization of different tree species at stand level can produce comparable LiDAR-derived metrics. This limits the performance of classifying tree species when relying solely on structural information. To address this issue, we integrated multispectral Sentinel-2 images to provide additional spectral information relevant for species discrimination, such as biochemical and phenological differences. A total of 129 structural and spectral metrics were extracted from both datasets and used as inputs to several machine learning models. The k-Nearest Neighbor (k-NN) algorithm was first applied to estimate the proportion of basal area per species, and the Synthetic Minority Oversampling Technique (SMOTE) was adopted to mitigate class imbalance. Dimensionality reduction through Principal Component Analysis (PCA) further improved model efficiency. Combining ALS measurements with Sentinel-2 time series greatly enhanced the detection of dominant tree species, according to classification trials using Random Forest (RF) and Support Vector Machines (SVM). Using Random Forest classification with SMOTE data augmentation, the fused configuration achieved an overall accuracy of 92% (<InlineEquation ID="IEq1"> <EquationSource Format="TEX">$\kappa = 0.87\%$</EquationSource> <EquationSource Format="MATHML"><math> <mrow> <mi>κ</mi> <mo>=</mo> <mn>0.87</mn> <mo>%</mo> </mrow> </math></EquationSource> </InlineEquation>), which is around 23 percentage points better than single-source setups (ALS only: 66%; Sentinel-2 only: 67%). This demonstrates the value of combining structural and spectral data. The multi-temporal Sentinel-2 configuration captured complementary phenological information during acquisitions in spring, summer and autumn. This contributed to species separability that would not have been possible from a single acquisition date, as evidenced by the spectral overlap observed during the summer acquisitions. Across all experiments, the Random Forest classifier consistently outperformed SVM. These results demonstrate that multi-sensor structural-spectral fusion is a robust, operationally feasible solution for improving forest species mapping and supporting sustainable forest inventory practices.</p>

Multi-temporal Sentinel-2 images and LiDAR data fusion in dominant tree species classification

期刊论文
发表日期：

2026年7月23日

查看全文

Douss Rim,
Imed Riadh Farah

摘要

Forest inventory plays a key role in sustainable forest management by providing essential information on forest structure and composition. In this context, remote sensing data supplement field surveys, enabling the large-scale, consistent and cost-effective monitoring of forest inventories, as well as structural and compositional features. To improve this process, this study explores the fusion of two complementary remote sensing sources. Airborne laser scanning (ALS) and multi-temporal Sentinel-2 imagery. The goal is to improve the estimation of two important forest indicators: basal area and dominant tree species at the stand level. ALS point clouds offer detailed, three-dimensional representations of forest structures. These can be used to estimate attributes such as canopy height, crown geometry, biomass and basal area, through dedicated processing and modeling methods. However, their ability to discriminate between species depends on species-specific canopy architecture and growth strategies. The similar structural organization of different tree species at stand level can produce comparable LiDAR-derived metrics. This limits the performance of classifying tree species when relying solely on structural information. To address this issue, we integrated multispectral Sentinel-2 images to provide additional spectral information relevant for species discrimination, such as biochemical and phenological differences. A total of 129 structural and spectral metrics were extracted from both datasets and used as inputs to several machine learning models. The k-Nearest Neighbor (k-NN) algorithm was first applied to estimate the proportion of basal area per species, and the Synthetic Minority Oversampling Technique (SMOTE) was adopted to mitigate class imbalance. Dimensionality reduction through Principal Component Analysis (PCA) further improved model efficiency. Combining ALS measurements with Sentinel-2 time series greatly enhanced the detection of dominant tree species, according to classification trials using Random Forest (RF) and Support Vector Machines (SVM). Using Random Forest classification with SMOTE data augmentation, the fused configuration achieved an overall accuracy of 92% ( $\kappa = 0.87\%$ $κ = 0.87 %$ ), which is around 23 percentage points better than single-source setups (ALS only: 66%; Sentinel-2 only: 67%). This demonstrates the value of combining structural and spectral data. The multi-temporal Sentinel-2 configuration captured complementary phenological information during acquisitions in spring, summer and autumn. This contributed to species separability that would not have been possible from a single acquisition date, as evidenced by the spectral overlap observed during the summer acquisitions. Across all experiments, the Random Forest classifier consistently outperformed SVM. These results demonstrate that multi-sensor structural-spectral fusion is a robust, operationally feasible solution for improving forest species mapping and supporting sustainable forest inventory practices.