AttentionUNet-OBIA: Hybrid Forest Mapping

• The paper presents a hybrid method combining AttentionUNet with OBIA to deliver both pixel-level discrimination and object-level interpretability for forest/non-forest classification.
• The methodology employs a UNet-style encoder-decoder with attention gates alongside mean-shift segmentation, enhancing feature focus and spatial coherence in high-resolution remote sensing data.
• The approach achieves state-of-the-art performance (OA 95.64%, IoU 0.9064) and outperforms traditional OBIA and other deep learning variants in identifying forest cover.

AttentionUNet-OBIA is a hybrid forest cover mapping methodology that integrates a deep learning model—AttentionUNet—with Object-Based Image Analysis (OBIA) for high-resolution multispectral remote sensing image analysis. Developed within the "ForCM" pipeline for Sentinel-2 imagery, it achieves state-of-the-art accuracy for forest/non-forest classification in the Amazon Rainforest, providing both pixel-wise discrimination and object-level interpretability with open-source tools (Haque et al., 29 Dec 2025).

1. Architecture and Attention Mechanism

The core of AttentionUNet-OBIA is a UNet-style encoder–decoder architecture augmented with attention gates (AG). The input consists of $512 \times 512 \times C$ images ($C=$ 3 or 4 bands). The model comprises four encoding stages, a bottleneck, and four decoding stages that symmetrically mirror the encoder. Each encoder level $\ell$ applies two consecutive $3\times3$ convolutions with ReLU activations (optionally batch-normalized), doubling feature channels at each downsampling (e.g., $64 \to 128 \to 256 \to 512$). Spatial resolution is reduced by $2 \times 2$ max-pooling (stride 2). The bottleneck contains two $3\times3$ convolutions at 1024 channels.

Decoding consists of $2\times2$ transposed convolution (up-convolution) to increase spatial resolution and halve channel count. At each decoder level $\ell$, an attention gate receives the upsampled decoder signal $g^\ell$ and the encoder skip feature $C=$0. The AG formula follows Oktay et al. (2018):

\begin{align*} f_x &= W_x\,x^\ell, \quad f_g = W_g\,g^\ell \ \Psi_\text{int} &= \operatorname{ReLU}(f_x + f_g + b) \ \alpha^\ell &= \sigma(\psi^T \Psi_\text{int} + b_\psi) \ x^{\ell\,\prime} &= \alpha^\ell \odot x^\ell \end{align*}

where $C=$1 are learned convolutions, $C=$2 the sigmoid, and $C=$3 element-wise multiplication. The AG output $C=$4 emphasizes salient features and suppresses irrelevant regions before skip connection concatenation. The final feature map passes through a $C=$5 convolution, followed by sigmoid activation, yielding a pixel-wise probability map $C=$6.

2. Data Preprocessing and Input Modalities

Input data are Sentinel-2 Level-2A images, pre-corrected for atmospheric effects using ESA Sen2Cor. Band selection includes both three-band (RGB) and four-band (RGB plus NIR, all at 10 m) sets. Input normalization policies are:

• Three-band images: divide by 255, cast to float32, yielding values in $C=$7.
• Four-band images: cast to float32, per-band divided by maximum reflectance ($C=$8), then rescaled to $C=$9.

No additional spectral indices such as NDVI are computed; only raw bands are provided to the model.

3. OBIA: Segmentation and Feature Extraction

Post-prediction, OBIA is performed in QGIS (v3.34.5, Orfeo Toolbox v8.1.2) using mean-shift segmentation, chosen for robust unsupervised object delineation. Mean-shift parameters are set to spatial radius $\ell$0, range radius $\ell$1, and minimum object size $\ell$2 pixels (trial-and-error selected). Each resulting image object $\ell$3 yields a feature vector $\ell$4, where $\ell$5 are mean band reflectances and $\ell$6 is the mean AttentionUNet pixelwise probability within $\ell$7. Optional features (not used in this work) include area, perimeter, compactness, and texture.

4. Fusion, Classification, and Post-processing

The classification stage fuses AttentionUNet-derived and OBIA-obtained features at object level. From the segmented object set $\ell$8, $\ell$9 are randomly sampled (visually-checked stratification) for manual ground-truth labeling. A linear-kernel Support Vector Machine (SVM, $3\times3$0) is trained to map $3\times3$1 to forest/non-forest labels. Label assignment is performed by evaluating SVM score $3\times3$2 with a threshold at 0 (or probability 0.5, if calibrated). Post-processing removes objects $3\times3$3 pixels by morphological opening, followed by boundary smoothing via a $3\times3$4 majority filter to reduce salt-and-pepper noise.

5. Training Protocol and Hyperparameters

Datasets are split as follows: for 3-band sets, V1: $3\times3$5 train/$3\times3$6 val/$3\times3$7 test; V2: $3\times3$8 train/$3\times3$9 val/$64 \to 128 \to 256 \to 512$0 test; V3: $64 \to 128 \to 256 \to 512$1 train/$64 \to 128 \to 256 \to 512$2 val/$64 \to 128 \to 256 \to 512$3 test; and for the 4-band set: $64 \to 128 \to 256 \to 512$4 train/$64 \to 128 \to 256 \to 512$5 val/$64 \to 128 \to 256 \to 512$6 test. Training applies binary cross-entropy loss with Adam optimizer ($64 \to 128 \to 256 \to 512$7, $64 \to 128 \to 256 \to 512$8), batch size $64 \to 128 \to 256 \to 512$9, initial learning rate $2 \times 2$0, with ReduceLROnPlateau scheduling (factor $2 \times 2$1, patience $2 \times 2$2), and no class weighting (class balance assumed). Data augmentation is limited to random horizontal/vertical flips performed on-the-fly. Training epochs: 20 for V1, 10 for V2, V3, and 4-band. Hardware employed: Intel i7-class CPU, 32 GB RAM, NVIDIA GeForce GTX TITAN X 12 GB GPU.

6. Evaluation Metrics and Comparative Results

Performance is assessed using mean Intersection over Union (IoU), overall accuracy (OA), precision, recall, and F1-score, computed on randomly selected test images. AttentionUNet-OBIA achieves:

Metric	Value
OA	95.64 %
IoU	0.9064
Precision	93.32 %
Recall	96.84 %
F1-score	0.9504

Comparative results show that AttentionUNet-OBIA surpasses traditional OBIA (OA 92.91 %, IoU 0.8992, F1 0.9365) and other DL-OBIA variants such as ResUNet-OBIA (OA 94.54 %, IoU 0.9101, F1 0.9525). Standalone AttentionUNet (no OBIA) attains OA 95.93 % and IoU 0.9168 on the 4-band test set. An example confusion matrix for 1000 test pixels:

	Pred Forest	Pred Non-Forest
True Forest	581	19
True Non-Forest	27	373

This evaluates to $2 \times 2$3.

7. Workflow Schematic

The pipeline is summarized as follows:

1. Load images and ground-truth masks.
2. Normalize bands to $2 \times 2$4 (float32).
3. Partition datasets into train/val/test.
4. Construct AttentionUNet:
   • Encoder (levels $2 \times 2$5): $2 \times 2$6
   • Bottleneck: $2 \times 2$7
   • Decoder ($2 \times 2$8): $2 \times 2$9 two $3\times3$0
   • Output: $3\times3$1 Conv (sigmoid).
5. Train with Adam optimizer and BCE loss, 10–20 epochs.
6. Inference: output $3\times3$2 probability map.
7. Segment objects with mean-shift (QGIS/OTB, $3\times3$3).
8. For each object $3\times3$4, compute $3\times3$5 (bands), $3\times3$6; assemble feature $3\times3$7.
9. Train linear SVM on labeled $3\times3$8.
10. Classify all $3\times3$9 as forest/non-forest.
11. Assign SVM label to all pixels within $2\times2$0 for final raster.
12. Post-process with small-object removal and majority filter.

A plausible implication is that the approach leverages spatial coherence, spectral consistency, and pixel-level DL inference for highly accurate, interpretable mapping, facilitated by accessible open-source software (Haque et al., 29 Dec 2025).