In our research, we explored a novel methodology for the efficient and precise estimation of crucial forest structural parameters. These parameters, such as average diameter at breast height (DBH), stand basal area, Lorey’s height, and above-ground biomass, are fundamental indicators for assessing forest ecosystem health, functionality, and growth status. To address the need for accurate, high-resolution, and non-destructive monitoring, we integrated modern UAV drone technology with a sophisticated deep neural network (DNN) regression model. Furthermore, we employed the SHAP (Shapley Additive Explanations) framework to demystify the “black-box” nature of the deep learning model, thereby enhancing the interpretability and trustworthiness of our predictions.
The foundation of our study lies in data acquired by UAV drones. We utilized a commercial UAV equipped with a high-resolution RGB camera to capture aerial imagery over a managed forest plantation. The flight mission was meticulously planned to ensure high overlap between consecutive images, a critical requirement for generating detailed three-dimensional reconstructions.
This imagery served as the primary data source from which we derived two distinct but complementary datasets: Digital Orthophoto Mosaics (DOM) and three-dimensional point clouds generated through Structure-from-Motion (SfM) and Multi-View Stereo (MVS) algorithms, often referred to as Digital Aerial Photogrammetry (DAP) point clouds.
From the UAV-derived DOM, we extracted a suite of two-dimensional spectral features. Since we were working with standard RGB imagery, we computed a comprehensive set of visible-band vegetation indices (VIs). These indices, such as the Visible Difference Vegetation Index (VDVI), the Green-Red Ratio Index (GRRI), and the Excess Green Index (ExG), along with the raw digital number (DN) values for the Red, Green, and Blue channels, provide information on vegetation vigor, health, and coverage. The formulas for two exemplary indices are:
$$ \text{VDVI} = \frac{2 \times G – R – B}{2 \times G + R + B} $$
$$ \text{ExG} = 2 \times G – R – B $$
Concurrently, from the dense 3D point cloud, we extracted a wide array of structural metrics. These metrics describe the vertical distribution and density of points, which correlate directly with forest canopy structure. We calculated height percentiles (e.g., H_50 for median height, H_95 for dominant height), moments of the height distribution (mean, standard deviation, skewness, kurtosis), and canopy cover metrics based on point density across different height strata. A summary of key point cloud variables is presented in the table below.
| Variable Category | Example Metrics | Description |
|---|---|---|
| Height Percentiles | H_10, H_25, H_50, H_75, H_95, H_99 | The height value below which a given percentage of points fall. |
| Statistical Moments | H_mean, H_std, H_skewness, H_kurtosis | Describe the central tendency, dispersion, and shape of the height distribution. |
| Canopy Complexity | H_CV (Coefficient of Variation), H_IQ (Interquartile Range) | Measures of vertical heterogeneity and canopy layer depth. |
| Point Density | D_1, D_2, … D_10 | Proportion of points in successive vertical strata (e.g., 10 equal slices). |
For model training and validation, we established ground truth from a network of field plots. Within the area covered by the UAV drones, we surveyed numerous sample plots, measuring the diameter at breast height and tree height for every individual tree. From these individual tree measurements, we calculated the plot-level forest structural parameters of interest:
Average Diameter (AD): $$ D_g = \sqrt{\frac{1}{N}\sum_{i=1}^{N} D_i^2} $$
Lorey’s Mean Height (HL): $$ H_L = \frac{\sum_{i=1}^{N} (H_i \cdot g_i)}{\sum_{i=1}^{N} g_i} $$
Basal Area (BA): $$ BA = \sum_{i=1}^{N} g_i = \sum_{i=1}^{N} \frac{\pi}{4} D_i^2 $$
Above-Ground Biomass (AGB): We calculated individual tree biomass using species-specific allometric equations and summed for the plot total. A generic form is: $$ M_i = a \cdot D_i^b \cdot H_i^c $$, where $a$, $b$, and $c$ are model parameters.
The core of our analytical approach was the design and training of a Deep Neural Network (DNN) model. We configured a multi-layer perceptron regression model with several hidden layers. The network architecture was designed to learn complex, non-linear relationships between the extracted UAV drone features (both spectral and structural) and the target forest parameters. The model’s hyperparameters, including the number of layers, neurons per layer, activation functions, dropout rates, optimizer, and learning rate, were systematically tuned using a combination of grid search and validation set performance. The optimal configurations for the different parameter estimations are summarized below.
| Target Parameter | Data Source Used | Optimal Optimizer | Learning Rate |
|---|---|---|---|
| Average Diameter (AD) | DOM + DAP Point Cloud | Rmsprop | 0.005 |
| Basal Area (BA) | DOM + DAP Point Cloud | Adam | 0.0005 |
| Lorey’s Height (HL) | DOM + DAP Point Cloud | Adam | 0.005 |
| Above-Ground Biomass (AGB) | DOM + DAP Point Cloud | Rmsprop | 0.001 |
To comprehensively evaluate the contribution of different data sources from UAV drones, we trained and tested three distinct DNN model configurations for each forest parameter: one using only DOM spectral features, one using only DAP point cloud structural features, and one using the combined feature set.
The prediction results clearly demonstrated the superiority of the integrated approach. For all four forest structural parameters, the DNN model trained on the combined feature set (DOM + DAP point cloud) achieved the highest estimation accuracy. The performance metrics, averaged over multiple validation runs, are telling. For instance, the estimation of Above-Ground Biomass (AGB) benefited greatly from data fusion, with the combined model achieving a significantly higher coefficient of determination ($R^2$) and lower error compared to models using a single data source. This pattern held true for basal area and Lorey’s height as well. The estimation of average diameter proved to be the most challenging, yet the combined UAV drone data source still yielded the best performance.
While the DNN model provided excellent predictive accuracy, understanding *why* it made certain predictions is crucial for scientific insight and operational trust. This is where we applied the SHAP explainability framework. SHAP assigns each input feature an importance value (the SHAP value) for a specific prediction, representing how much that feature pushed the model’s output away from a baseline expectation.
The global SHAP analysis, which aggregates contributions across all predictions, revealed consistent patterns. For every forest parameter model (AD, BA, HL, AGB) built from the combined UAV drone data, the point cloud-derived “Coefficient of Variation of Height” (H_CV) emerged as the most important feature. This metric, describing the relative variability of canopy heights within a plot, was a powerful predictor of stand structure and productivity. Furthermore, point cloud height metrics (percentiles, mean, std) consistently ranked higher in importance than the spectral vegetation indices from the DOM. This highlights the critical value of the 3D structural information captured by UAV drone photogrammetry, which directly relates to the biophysical parameters we aim to estimate.
The SHAP framework also allows for local explanation. For any individual plot prediction, we can generate a force plot that visually decomposes the prediction. It shows how features like a high H_95 value (pushing the predicted biomass up) and a low greenness index (pushing it down) combine to arrive at the final estimated value. This level of transparency is invaluable for diagnosing model behavior and identifying potential outliers or systematic errors in specific forest conditions.
The success of this research underscores several key points. Firstly, UAV drones are a remarkably potent tool for forestry inventory, providing cost-effective, high-resolution spectral and, crucially, 3D structural data over complex terrain. The DAP point clouds generated from overlapping UAV imagery offer a viable and rich alternative to more expensive LiDAR systems for many applications. Secondly, deep learning models, particularly DNNs, are exceptionally well-suited for harnessing the complex, high-dimensional feature spaces derived from these UAV drone datasets. They can model the intricate, non-linear relationships between canopy reflectance, structure, and ground-measured forest attributes more effectively than many traditional machine learning algorithms.
However, the “black-box” critique of deep learning is valid. Our integration of the SHAP interpretability framework directly addresses this concern. By using SHAP, we can move beyond simply reporting prediction accuracy to explaining which features from the UAV drone data are driving those predictions. We found that structural metrics from the point cloud were generally more influential than spectral indices, providing clear empirical evidence for the added value of 3D data in forest parameter estimation. This interpretability fosters greater confidence in the model and provides actionable insights for forest ecologists and managers—for example, highlighting which structural attributes are most sensitive indicators of biomass or stand density.
In conclusion, our study demonstrates a robust and interpretable framework for forest monitoring. By synergistically combining the data acquisition capabilities of UAV drones, the predictive power of deep neural networks, and the explanatory clarity of SHAP analysis, we can achieve accurate and understandable estimates of key forest structural parameters. This methodology offers a significant advancement over traditional field-based surveys and simpler remote sensing approaches. It provides a scalable, efficient, and insightful solution for forest resource inventory, carbon stock assessment, and sustainable ecosystem management. Future work will focus on testing the transferability of these models across different forest types and ecological zones, further optimizing the UAV drone data acquisition protocols, and exploring even more advanced deep learning architectures to push the boundaries of precision in forestry remote sensing.