ResUNet-OBIA: Deep Learning & OBIA Integration

• The paper introduces a hybrid framework that combines ResUNet-based segmentation with OBIA, achieving improved accuracy and boundary precision in diverse imaging tasks.
• ResUNet-OBIA employs an encoder-decoder architecture with residual blocks and skip connections to enhance feature representation and multi-scale context aggregation.
• By integrating pixel-level predictions with object-level analysis, the framework enables efficient object extraction and robust classification in both remote sensing and medical imaging.

ResUNet-OBIA is a hybrid methodology integrating deep learning-based residual U-Net (ResUNet) semantic segmentation with object-based image analysis (OBIA) to enable precise, efficient, and scalable object extraction from diverse remote sensing and medical imaging modalities. The framework leverages the pattern representation capacity of deep convolutional neural networks for pixel-level prediction and the boundary-awareness of OBIA to deliver object-wise maps optimized for downstream analysis, mapping, or classification tasks.

1. Architectural Foundations of ResUNet-OBIA

The core of ResUNet-OBIA is a ResUNet segmentation backbone, typically structured as an encoder–decoder convolutional neural network with skip connections and within-block residual links. Its main architectural traits include:

• Encoder–decoder with skip connections: The encoder applies cascaded convolution–batch normalization–activation layers, down-sampling spatial resolution and increasing feature channels per stage. Feature maps are passed via skip connections to the decoder, enabling spatial information recovery at matching resolutions.
• Residual-convolutional blocks: Each block follows the form $y = F(x) + x$, where $F(x)$ denotes a sequence such as convolution–activation–convolution; $x$ is the block input. This facilitates gradient flow and stable training.
• Atrous/dilated convolutions and multi-scale context (application-specific): In models such as ResUNet-a, each Residual-Atrous block includes four parallel 3×3 convolutions at varying dilation rates for multiscale receptive fields. A pyramid scene parsing (PSP) pooling layer may aggregate context at multiple scales and fuse it into the main decoder path (Waldner et al., 2019).
• Conditioned inference: Some ResUNet-OBIA variants employ interdependent prediction, where the output of one head (e.g., distance-to-boundary) conditions subsequent heads (e.g., boundary probability, field/region label) to mutually reinforce spatial and semantic consistency (Waldner et al., 2019).

Loss functions are typically task-specific (cross-entropy for masks, L2 for regression, Tanimoto/Dice for overlap), combined into a weighted sum to guide optimization.

2. Integration with Object-Based Image Analysis (OBIA)

The OBIA module transforms ResUNet's dense pixel-wise outputs into discrete objects suitable for geospatial or clinical analysis. Implementation details vary by application but follow this general sequence:

• Segmentation: OBIA partitions the image into candidate objects using clustering (e.g., Mean Shift) or watershed-based region growing. Parameters such as spatial/radiometric radius and minimum object size are empirically optimized (Haque et al., 29 Dec 2025).
• Feature aggregation: For each segment (object), aggregated statistics of the ResUNet probability heatmap (mean, weighted mean) are combined with spectral and textural features from the original imagery.
• Object-level classification: A downstream classifier, often a support vector machine (SVM) or gradient-boosted tree model, is trained on these per-object features to assign final classes or scores (e.g., forest/non-forest, thyroid pathology type) (Sabouri et al., 2024, Haque et al., 29 Dec 2025). This step enables boundary sharpening and object-level interpretability.

The OBIA post-processing may include morphological operations, thresholding, and vectorization to yield closed, non-overlapping objects amenable to GIS or radiomic analysis.

3. Domain-Specific Methodological Implementations

Satellite Imagery and Land Cover

In field boundary extraction from Sentinel-2, ResUNet-OBIA produces per-pixel probabilities for field interior, boundary, and normalized distance to boundary. The post-processing pipeline converts these into closed crop field polygons using either:

• Cutoff + symmetric difference: Threshold field and boundary maps, combine via XOR, extract connected components.
• Seeded watershed: Identify seeds as local maxima of the distance map; using predicted masks as foreground/background, apply the watershed transform to produce objects matching vector field boundaries (Waldner et al., 2019).

Mean Shift OBIA is used for forest cover mapping: segments are created, DL-generated probabilities are aggregated, and each object is classified via SVM. ResUNet-OBIA achieves superior accuracy relative to traditional OBIA: for Amazon forest mapping, overall accuracy climbs from 92.91 % (OBIA alone) to 94.54 % (ResUNet-OBIA) and F₁ from 0.9365 to 0.9525 (Haque et al., 29 Dec 2025).

Medical Imaging

In thyroid scintigraphy, the pipeline applies ResUNet for thyroid gland auto-segmentation (DSC: 0.84 ± 0.03 for MNG, 0.71 ± 0.06 for TH, 0.86 ± 0.02 for DG). OBIA defines each ResUNet mask as a region, extracts 93 PyRadiomics features per ROI, followed by reduction (Spearman’s ρ masking, recursive feature elimination with XGBoost), and ultimately multi-class XGBoost classification. This automated approach yields classification accuracy 0.74 ± 0.05 and ROC AUC 0.90 ± 0.02, closely matching manual contouring, and substantiated by TOST equivalence on key metrics (Sabouri et al., 2024).

4. Training Procedures and Dataset Preparation

Training ResUNet-OBIA networks involves domain- and modality-specific steps:

• Input data: Sentinel-2 L2A multispectral composites (bands B, G, R, NIR), medical images (e.g., 128×128 thyroid scintigrams), consistently normalized and resampled to standard pixel size (Waldner et al., 2019, Haque et al., 29 Dec 2025, Sabouri et al., 2024).
• Ground truth: Manually digitized polygons (fields, organs), rasterized for mask generation; per-pixel class and auxiliary labels (distance, boundary).
• Augmentation: Strategies include random flips, brightness jitter; rotation omitted in agricultural applications to retain pivot symmetry (Waldner et al., 2019).
• Hyperparameter optimization: Empirical optimization (e.g., AdamW/Adam, batch size, learning rate scheduling, threshold selection) is performed iteratively, often with cross-validation (e.g., leave-one-center-out, LOCOCV, or 5-fold grid search) (Sabouri et al., 2024).

5. Quantitative Performance and Comparative Evaluation

The following summarizes reported performance on benchmark tasks:

Task	OBIA Only	ResUNet-OBIA	AttentionUNet-OBIA
IoU (Amazon, mean)	0.8992	0.9101	0.9064
Overall Accuracy (%)	92.91	94.54	95.64
Precision (%)	92.43	93.69	93.32
Recall (%)	95.12	96.98	96.84
F₁ Score	0.9365	0.9525	0.9504

In field boundary extraction from composite Sentinel-2 (South Africa), pixel-level OA ≈ 91.7 %, MCC ≈ 0.82, object-wise hit rate ≈ 99.4 %. Generalization is robust to temporal, spatial, and sensor variations: accuracy is marginally reduced for single-date or lower-resolution input (e.g., MCC from 0.82 to ~0.78, hit rate ~99 %) (Waldner et al., 2019).

For thyroid imaging across nine folds, ResUNet-OBIA achieves accuracy 0.74 ± 0.05, ROC AUC 0.90 ± 0.02. Results show only ~2 percentage points difference versus manual ROI, with statistical equivalence in multiple metrics (Sabouri et al., 2024).

6. Advantages, Limitations, and Application Scope

Advantages:

• Fuses deep learning’s feature representation with OBIA’s contour fidelity, marrying pixel-level confidence with object-level interpretability and boundary precision.
• Demonstrates consistent gains over OBIA alone (e.g., +1.6 % OA in forest cover tasks), and cost-effective open-source tooling (QGIS, OrfeoToolbox, PyRadiomics, XGBoost) (Haque et al., 29 Dec 2025, Sabouri et al., 2024).
• Modular and domain-agnostic; permits “plug-and-play” of various deep learning backbones into OBIA pipelines.

Limitations:

• OBIA sensitivity to segmentation parameters; substantial trial-and-error required, especially for Mean Shift or watershed thresholding (Haque et al., 29 Dec 2025, Waldner et al., 2019).
• Pipeline is not fully end-to-end; object classification is decoupled from segmentation, limiting joint optimization of deep features for object-wise loss (Haque et al., 29 Dec 2025).
• Dependent on quality of input imagery and segmentation; under/over-segmentation degrades performance.

A plausible implication is that robust consensus can be achieved across temporal, spatial, and sensor domains by averaging predictions across multiple well-chosen images, reducing output variance and improving large-scale generalization (Waldner et al., 2019).

7. Impact and Future Perspectives

ResUNet-OBIA is established as a state-of-the-art approach for extracting discrete objects with precise boundaries from complex imagery, scalable from agricultural field mapping and forest cover estimation to clinical region-of-interest identification and radiomic pipeline automation. By streamlining manual labor and reducing expert-driven variability, it supports operational workflows in environmental monitoring, land management, and clinical diagnostics (Waldner et al., 2019, Sabouri et al., 2024, Haque et al., 29 Dec 2025).

Future developments may target end-to-end parameter learning, improved object segmentation algorithms supplanting Mean Shift, and integration of semi- or weakly-supervised learning for large-scale, label-efficient modeling. ResUNet-OBIA’s demonstrated sensor, domain, and spatiotemporal generalization supports its adoption as a canonical methodology for object-centric analysis in remote sensing and medical imaging.