YOLO-DCAT: Enhanced Gravity Wave Detector

• The paper introduces YOLO-DCAT, which integrates Multi Dilated Residual Convolution and Simplified Spatial and Channel Attention into the YOLOv5 backbone for improved gravity wave localization.
• It employs dilated convolutions with paired rates (2,3) and parallel attention mechanisms to capture multi-scale gravity wave features while addressing noise from clouds and city lights.
• Empirical results demonstrate significant gains, with mAP improving from 41.8% to 55.3% and IoU from 31.62% to 48.74%, indicating enhanced accuracy and training stability.

YOLO-DCAT is an enhanced version of YOLOv5 for gravity wave localization in satellite imagery, introduced in the context of VIIRS Day/Night Band nighttime granules under new-moon conditions. It is designed for a detection regime characterized by high object variability, lower spatial resolution, dominant unwanted objects such as clouds and city lights, instrumental noise, and a single image channel. The model incorporates Multi Dilated Residual Convolution (MDRC) and Simplified Spatial and Channel Attention (SSCA) into the YOLOv5-small (v6.0) backbone in order to improve localization of atmospheric gravity waves whose scale, shape, and pattern extent vary substantially across scenes (Mostafa et al., 29 Jul 2025).

1. Architectural position within the YOLOv5 family

The base model is YOLOv5-small (v6.0), selected for efficiency on limited data and compute. YOLO-DCAT modifies the CSPDarknet backbone by replacing every standard convolution layer with a two-stage block composed of an MDRC unit followed immediately by an SSCA module. By contrast, the neck and head are unchanged: PANet remains responsible for multi-scale feature fusion, and the detection head continues to predict class, objectness, and bounding boxes (Mostafa et al., 29 Jul 2025).

The high-level data flow is reported as:

$$\text{Input} \rightarrow \text{Focus} \rightarrow [\text{Conv} \rightarrow \text{MDRC} \rightarrow \text{SSCA}]_1 \rightarrow [\text{C3} \rightarrow \text{MDRC} \rightarrow \text{SSCA}]_2 \rightarrow \ldots \rightarrow \text{SPPF} \rightarrow \text{PANet} \rightarrow \text{Detection Head}$$

The corresponding architectural interpretation is narrow and specific. YOLO-DCAT does not redefine the YOLOv5 detection pipeline as a whole; instead, it concentrates its intervention in the backbone, where feature extraction is most directly exposed to the scale variability and low-contrast structure of gravity-wave signatures. A plausible implication is that the method is intended as a targeted backbone enhancement rather than a full detector redesign.

2. Multi Dilated Residual Convolution

MDRC is introduced to expand the receptive field without spatial down-sampling, with the explicit goal of capturing gravity waves at multiple scales. Its structure consists of two parallel $3\times 3$ dilated convolutions with dilation rates $d_1=2$ and $d_2=3$, concatenation of their outputs, a $1\times 1$ convolution followed by BatchNorm and ReLU to reduce channels back to $C$, and a residual skip connection back to the input (Mostafa et al., 29 Jul 2025).

For input $x \in \mathbb{R}^{H \times W \times C}$, the formulation is given as:

$$y_1 = x * W_1^{(d_1)}, \qquad y_2 = x * W_2^{(d_2)}$$

where $*$ denotes convolution and $$W_i^{(d_i)} \in \mathbb{R}^{3 \times 3 \times C \times C}$$ with dilation $3\times 3$0.

$3\times 3$1

$3\times 3$2

$3\times 3$3

The reported functional rationale is explicit: $3\times 3$4 captures fine-grained wave crests, $3\times 3$5 captures broader wavefronts, and the residual skip preserves low-level detail and eases gradient flow. In the gravity-wave setting, this establishes MDRC as the multi-scale context extractor within the modified backbone.

3. Simplified Spatial and Channel Attention

SSCA is designed to highlight important spatial locations and feature channels while suppressing noise from clouds, city lights, and other background structure. Its spatial pathway applies channel-wise average pooling and max pooling to form two $3\times 3$6 maps, concatenates them, and passes the result through a $3\times 3$7 convolution followed by a sigmoid. Its channel pathway applies spatial average pooling followed by a two-layer projection with reduction ratio $3\times 3$8 and sigmoid gating. The final output is formed by broadcasted element-wise multiplication of the input feature map with both attention masks (Mostafa et al., 29 Jul 2025).

The equations are reported as:

$3\times 3$9

$d_1=2$0

$d_1=2$1

$d_1=2$2

with $d_1=2$3, $d_1=2$4, and $d_1=2$5.

A central design distinction is that SSCA computes spatial and channel attention in parallel, unlike CBAM’s sequential attention. The stated rationale is reduced overhead and tighter coupling of spatial-channel information. Within YOLO-DCAT, SSCA therefore functions as the selective amplification stage that follows MDRC’s multi-scale feature generation.

4. Data regime, preprocessing, and optimization

The dataset consists of VIIRS Day/Night Band nighttime granules under new-moon conditions. The images are single-channel $d_1=2$6 px at approximately $d_1=2$7 GSD with radiance of approximately $d_1=2$8. The annotation set contains 600 expert-annotated gravity-wave bounding boxes plus a “noise” class for clouds and city lights. The split is 70% train, 20% validation, and 10% test (Mostafa et al., 29 Jul 2025).

The preprocessing pipeline is specified in four steps: subtract the minimum pixel value to set a baseline of zero; scale by the median to reduce brightness bias; normalize to mean $d_1=2$9 to center the distribution; and transform the histogram to uniform in a range-preserving manner to equalize subtle patterns. Data augmentation is inherited from YOLOv5 and includes Mosaic augmentation, MixUp, random horizontal and vertical flip, and hue/saturation jitter.

Optimization uses SGD with momentum, the default YOLOv5 configuration, together with cosine annealing, auto-anchor box fitting, mixed-precision FP16 training, and a loss defined as BCE for objectness and classification plus CIoU for bounding-box regression. The reported hyperparameters are batch size 16, initial learning rate 0.01, weight decay $d_2=3$0, and 300 training epochs.

These design choices indicate that YOLO-DCAT was developed under a constrained-data regime in which preprocessing and inductive bias are expected to be unusually important. This suggests that the architecture is tuned not merely for generic object detection, but for a physically and instrumentally specific observation domain.

5. Empirical performance and comparative behavior

The reported evaluation metrics are Precision (%), Recall (%), mAP@50 (%), mAP@50–95 (%), and IoU (%). Against the baseline YOLOv5 configuration, YOLO-DCAT is reported to improve mean Average Precision by over 14% and Intersection over Union by approximately 17% in the abstract, while the tabulated summary gives the baseline and full-model values more precisely as mAP50 $d_2=3$1 and IoU $d_2=3$2, corresponding to gains of $d_2=3$3 points and $d_2=3$4 points respectively (Mostafa et al., 29 Jul 2025).

Model variant	mAP50 (%)	IoU (%)
Baseline YOLOv5	41.80	31.62
+MDRC only	49.90	32.76
+SSCA only	50.90	38.64
YOLO-DCAT (MDRC+SSCA)	55.30	48.74

The baseline YOLOv5 is further reported at Precision 56.20 and Recall 39.70. Relative to other attention modules, CBAM without MDRC reaches mAP50 50.40 and IoU 36.92, whereas SSCA without MDRC reaches 50.90 and 38.64. With MDRC added, MDRC+CBAM reaches mAP50 52.80 and IoU 44.48, compared with YOLO-DCAT at 55.30 and 48.74. Transformer and ViT variants are reported to struggle on single-channel data, with mAP50 less than or equal to 46.4%.

Statistical robustness is summarized over five runs: the baseline yields $d_2=3$5 mAP50 and $d_2=3$6 IoU, whereas YOLO-DCAT yields $d_2=3$7 mAP50 and $d_2=3$8 IoU. The lower standard deviations suggest more stable performance in repeated training under the reported experimental conditions.

A common misconception in attention-augmented detection is that stronger modules or more globally expressive blocks automatically dominate simpler designs. The reported comparison argues against that generalization in this setting: on single-channel VIIRS gravity-wave data, the specifically tailored MDRC+SSCA combination outperforms the listed Transformer-, ViT-, and CBAM-based alternatives.

6. Ablation logic and placement sensitivity

The ablation studies emphasize that the efficacy of YOLO-DCAT depends not only on module design but also on placement. When MDRC or SSCA is inserted in C3 layers, performance drops below baseline, with mAP50 approximately 34% and 37%. When inserted in standard convolution layers, the best results are obtained: mAP50 49.90 for MDRC and 50.90 for SSCA (Mostafa et al., 29 Jul 2025).

The dilation-rate study reports a similarly constrained optimum. Single dilation choices give mAP50 values of 45.80 for $d_2=3$9, 44.60 for $1\times 1$0, and 39.90 for $1\times 1$1. Paired rates $1\times 1$2 give 49.90, whereas $1\times 1$3 and $1\times 1$4 degrade to approximately 41–42%. The stated conclusion is that dilated convolutions must be added at the correct depth and that the multi-scale pair $1\times 1$5 optimally balances local and global context.

Attention placement follows the same pattern: only insertion in standard convolution layers yields benefit, while C3 insertion hurts performance. This makes the paper’s empirical position unusually specific. The relevant lesson is not merely that multi-scale convolutions and attention help, but that they help under a particular interaction between feature hierarchy and module placement. A plausible implication is that the backbone stage at which low-level texture and mid-level structure are fused is especially important for this localization task.

7. Scope, limitations, and projected extensions

The current design is tailored to single-channel VIIRS data, and multi-band data would require re-tuning attention pooling. The model also reflects a model size versus detection speed trade-off: the small YOLO variant was used, and scaling to larger YOLO variants such as v12 or to DETR is identified as a possible route to higher accuracy at higher cost (Mostafa et al., 29 Jul 2025).

The future enhancements listed for the framework are: extending SSCA to hierarchical attention such as global-local modeling; integrating residual-attention blocks for explicit denoising of clouds and city lights; applying the method to other atmospheric and oceanic phenomena including ocean eddies, mesospheric bores, cloud-property retrieval, and mesoscale hurricanes; and exploring transformer-CNN hybrids or lightweight DETR variants for end-to-end detection.

These limitations define the current boundaries of the method. YOLO-DCAT is not presented as a universal detector for all remote-sensing modalities; rather, it is a specialized adaptation of YOLOv5 for a demanding single-channel localization problem. Its reported contribution lies in the tight coupling of multi-dilated convolutions and parallel spatial-channel attention within the backbone, with the stated outcome of improved localization accuracy and more precise downstream support for climate research and modeling.