Technical Summary: Dual-Task Learning for Dead Tree Detection and Segmentation

Problem Statement

Accurate mapping of standing dead trees is essential for assessing forest health, monitoring biodiversity, and mitigating wildfire risks. However, existing methods face significant limitations when applied to high-resolution aerial imagery, particularly in dense boreal forests. Key challenges include:

• Dense Canopy Structures: Overlapping tree crowns make it difficult to separate individual instances.
• Spectral Overlaps: Spectral similarities between living and dead vegetation (especially in disturbed or seasonal ecosystems) hinder differentiation.
• Over-segmentation Errors: Traditional deep learning models often fail to balance detection accuracy with the generation of segmentation artifacts, leading to unreliable instance delineation.
• Data Limitations: Existing datasets often lack spectral richness (e.g., Near-Infrared bands) and geographic diversity, while pre-trained models struggle to generalize to high-resolution aerial data due to scale and perspective differences.

Methodology

The study proposes a hybrid postprocessing framework that integrates deep learning with watershed algorithms to refine tree segmentation. The core of the approach is the TreeMort-3T-UNet, a multi-task learning architecture designed for instance segmentation and centroid localization.

1. Architecture: TreeMort-3T-UNet

The model employs a ResNet-34 encoder pre-trained on ImageNet and fine-tuned on the FLAIR-INC dataset (comprising RGB-NIR imagery). This domain adaptation ensures robustness against spatial and temporal shifts. The architecture features a triple-output decoder:

• Segmentation Head: Predicts a binary mask distinguishing dead tree canopies from the background.
• Centroid Localization Head: Outputs a Gaussian heatmap where peaks correspond to the centers of individual dead trees.
• Hybrid SDT-Boundary Head: Generates a map encoding a Signed Distance Transform (SDT) fused with explicit boundary cues. Boundary pixels are assigned values near –1, while non-boundary pixels represent normalized distances up to +1.

Attention Mechanisms: Self-attention modules are integrated into the decoder to dynamically enhance relevant feature maps and suppress background noise, crucial for distinguishing subtle spectral variations in multispectral imagery.

2. Hybrid Loss Function

To jointly optimize the three outputs, a composite loss function is used:

• Segmentation Loss: A weighted combination of Binary Cross-Entropy (BCE) and Dice loss.
• Centroid Loss: Mean Squared Error (MSE) to penalize deviations between predicted and ground-truth Gaussian heatmaps.
• Hybrid SDT-Boundary Loss: A regression loss comprising Smooth L1 loss for non-boundary pixels and an up-weighted L1 loss for boundary pixels (targeting –1) to handle class imbalance.
• Total Loss: A weighted sum of the individual components.

3. Hybrid Postprocessing Pipeline

The final instance segmentation is achieved through a multi-stage refinement process that fuses the network outputs:

1. Initial Refinement: The segmentation mask is thresholded, and small spurious regions are removed via connected-component analysis.
2. Hybrid Filtering: The hybrid SDT-boundary map is thresholded to suppress regions lacking strong boundary signals, reducing over-segmentation.
3. Centroid Marker Extraction: The centroid heatmap is smoothed, and local maxima are detected to serve as seeds for individual trees.
4. Watershed Segmentation: The watershed algorithm is applied to the negative of the smoothed centroid map, using the refined binary mask as a constraint. This partitions the mask into distinct regions corresponding to individual trees.
5. Optional Vectorization: Labels can be converted to vector representations via ellipse fitting or contour extraction.

Key Contributions

The paper identifies three primary contributions:

1. Unified Architecture: A domain-adapted ResNet-34 based model that simultaneously performs instance segmentation and centroid localization, optimized for overlapping objects and multispectral variability.
2. Novel Loss Combination: The integration of Binary Cross-Entropy, Dice, Focal, and Mean Squared Error losses to balance pixel-level segmentation with precise centroid localization while mitigating class imbalance.
3. Multi-Step Refinement: A postprocessing pipeline incorporating adaptive thresholding, centroid-guided enhancement, morphological operations, and watershed segmentation to accurately delineate individual tree instances.

Experimental Results

The framework was evaluated on high-resolution aerial imagery from boreal forests in Finland (0.25 m/pixel resolution, RGB-NIR bands). The dataset included approximately 15,000 annotated dead tree polygons across 125 images.

Performance Metrics:

• Instance-Level Accuracy: The proposed method achieved a Mean Tree IoU of 0.3710, a significant improvement over the baseline U-Net (0.2622) and other state-of-the-art models (e.g., MaskFormer at 0.3503). This represents a 41.5% improvement in instance-level segmentation accuracy.
• Localization Precision: Centroid localization error was reduced to 3.70 pixels, a 57% reduction compared to the baseline U-Net (8.60 pixels).
• Recall and F1-Score: The method substantially improved recall (0.6687 vs. 0.4667 for U-Net) and F1-score (0.5895 vs. 0.4471), indicating a superior ability to recover missed tree instances.
• Statistical Significance: Paired tests confirmed that the improvements in both Pixel IoU and Tree IoU are statistically significant (p < 0.001).

Ablation Studies:
The study demonstrated that the watershed segmentation step was the most impactful component, significantly improving Tree IoU (from 0.3299 to 0.3689) and reducing centroid error. The combination of all postprocessing steps yielded the best overall performance.

Significance and Claims

The authors claim that this work bridges the gap between advanced remote sensing techniques and practical forest management needs. By balancing detection accuracy with the reduction of over-segmentation artifacts, the method enables the precise identification of individual dead trees, which is critical for:

• Wildfire Risk Assessment: Identifying fuel accumulations.
• Carbon Stock Estimation: Tracking emissions from decaying biomass.
• Precision Forestry: Targeting salvage loggings and reforestation strategies.

The framework is described as computationally efficient, supporting scalable applications such as wall-to-wall tree mortality mapping over large geographic regions. The study positions this work as an advancement in tools for large-scale ecological conservation and climate resilience planning, offering a robust solution for ecological monitoring in complex, densely vegetated environments.