The imperative for global decarbonization has placed renewable energy sources, particularly solar power, at the forefront of sustainable development strategies. Large-scale photovoltaic (PV) power plants are critical infrastructure in this transition. However, their operational efficiency is significantly compromised by environmental factors, with dust accumulation being a predominant issue, especially in arid and semi-arid regions. The deposition of dust, sand, and other particulate matter on solar panels creates an obstructive layer, reducing the amount of incident sunlight reaching the photovoltaic cells. This directly leads to a substantial decrease in power output. Furthermore, non-uniform dust accumulation can cause localized heating, known as hot spots, which accelerates the degradation of the solar panels and can lead to permanent damage and safety hazards. Therefore, the timely and accurate identification of dust accumulation levels is paramount for efficient plant maintenance, optimal energy yield, and the longevity of the PV assets.
Traditionally, the assessment of soiling on solar panels has relied heavily on manual inspection by site technicians. This method is subjective, labor-intensive, inconsistent, and non-scalable for large farms comprising thousands of modules. Automated methods have been explored, primarily focusing on correlating dust density with electrical output parameters like power loss or using optical techniques to measure light transmittance. While insightful, these approaches often involve installing additional sensors, which increases cost and complexity, and their models can lack generalizability across different solar panels and environmental conditions.
Recent advances in computer vision and deep learning offer a promising, cost-effective, and scalable alternative. By analyzing images of the solar panels, a model can learn the visual patterns—changes in color, texture, and contrast—associated with different levels of dust. This approach mirrors the human visual inspection process but with superior consistency, speed, and the potential for integration with drone or fixed-camera surveillance systems. The core challenge lies in developing a model that is both highly accurate and robust enough to handle the variability in lighting conditions, panel types, and dust distribution patterns encountered in real-world settings.
In addressing this challenge, we propose a novel deep learning framework based on an improved deep residual network for the automated recognition of dust accumulation degrees on solar panels. Our methodology systematically enhances a foundational architecture through structural refinements, advanced attention mechanisms, and a superior loss function, leading to significant gains in classification accuracy on a challenging real-world dataset.
1. Methodological Framework
Our model is built upon the ResNeXt-50 architecture, a powerful variant of residual networks that employs a split-transform-merge strategy using grouped convolutions. We introduce three key modifications to tailor this network for the specific task of dust recognition on solar panels.
1.1 Architectural Refinements: Decomposition and Sampling
The initial layer (often called Conv1) in standard architectures like ResNeXt uses a large 7×7 convolutional kernel. We apply the principle of convolution decomposition, replacing this single large kernel with a stack of three 3×3 convolutional layers. This offers two major advantages: it increases the network’s depth and non-linearity for richer feature extraction, and it dramatically reduces the number of parameters. The parameter count for the initial feature extractor is reduced by approximately 81%, lowering computational complexity without sacrificing representational power.
Furthermore, we address a subtle issue in the standard downsampling procedure within residual blocks. In the original design, the first 1×1 convolution in a block tasked with spatial reduction (e.g., halving the feature map size) performs the downsampling with a stride of 2. This operation can lead to significant information loss, as a large portion of the input activation map is skipped. We refine this by transferring the downsampling responsibility to the subsequent 3×3 convolution within the block. This simple adjustment allows the initial 1×1 convolution to process the full spatial information before reduction, mitigating information loss and leading to more robust feature maps. The comparison of these operations is illustrated below, where shaded areas represent information used in the convolution.
Let an input feature map have dimensions $H \times W$. A 1×1 convolution with stride $s=2$ uses only $\frac{1}{s^2} = \frac{1}{4}$ of the spatial locations directly. In contrast, our proposed method applies a 1×1 convolution with stride $s=1$ over the entire map, followed by a 3×3 convolution with stride $s=2$. This ensures all input data contributes to the channel-wise features before spatial pooling occurs, preserving more information critical for discerning subtle dust patterns on solar panels.
1.2 Incorporating Coordinate Attention (CA) Mechanism
Dust on solar panels is rarely uniformly distributed. It often accumulates along edges, mounting frames, or in specific patterns due to wind and gravity. A standard convolutional network may not consistently focus on these semantically important but spatially variable regions. To address this, we integrate a Coordinate Attention (CA) mechanism. Unlike channel-only attention modules (e.g., Squeeze-and-Excitation) or spatial attention modules that use large kernels, CA effectively captures long-range dependencies with precise positional information by decomposing the global pooling operation.
The CA mechanism operates in two sequential steps: Coordinate Information Embedding and Coordinate Attention Generation. For an input feature map $\mathbf{X}$ with $C$ channels, it first applies one-dimensional horizontal and vertical global pooling, generating two sets of direction-aware feature maps: $\mathbf{z}^h$ (size $C \times H \times 1$) and $\mathbf{z}^w$ (size $C \times 1 \times W$). These encode global receptive fields along each spatial dimension, capturing precise spatial structure—essential for locating dust streaks or patches on a solar panel. These pooled features are then concatenated and transformed via a shared 1×1 convolution and a non-linear activation $\delta$ (e.g., h-swish):
$$ \mathbf{f} = \delta\left(F_1\left(\left[\mathbf{z}^h, \mathbf{z}^w\right]\right)\right) $$
The resulting tensor $\mathbf{f}$ is split back into $\mathbf{f}^h$ and $\mathbf{f}^w$ along the spatial dimension. Separate 1×1 convolutional transforms $F_h$ and $F_w$ followed by sigmoid activations $\sigma$ produce the final attention weights:
$$
\mathbf{g}^h = \sigma(F_h(\mathbf{f}^h)), \quad \mathbf{g}^w = \sigma(F_w(\mathbf{f}^w))
$$
The output of the CA block is the recalibrated feature map $\mathbf{Y}$, where each element is computed as:
$$ y_c(i, j) = x_c(i, j) \times g^h_c(i) \times g^w_c(j) $$
This process allows the network to emphasize relevant features along both the height and width dimensions selectively. For our task, this means the model can learn to pay more attention to the bottom edges or corners of a solar panel where dust tends to settle, significantly improving its discriminative capability for non-uniform soiling patterns.
1.3 Supervised Contrastive Learning Loss
Classification tasks typically use the Cross-Entropy (CE) loss. However, CE treats each sample independently and may not optimally learn a feature space where samples from the same class (e.g., “heavily soiled” solar panels) are tightly clustered and well-separated from other classes. We employ Supervised Contrastive (SupCon) Loss to learn a more robust and generalizable embedding space.
The SupCon loss leverages label information to define positive and negative pairs explicitly. Within a mini-batch, all augmented versions of samples sharing the same class label are considered positives for each other, while samples from different classes are negatives. Formally, let $\mathbf{z}_i = \text{Proj}(\text{Enc}(\mathbf{x}_i))$ be the projected feature vector for an augmented sample $i$, normalized onto a unit hypersphere. Let $A(i)$ be the set of all indices in the batch except $i$, and $P(i) = \{p \in A(i) | \tilde{y}_p = \tilde{y}_i\}$ be the set of indices of all positives for anchor $i$. The SupCon loss is defined as:
$$
\mathcal{L}^{sup}_{out} = \sum_{i \in I} \mathcal{L}^{sup}_{out,i} = \sum_{i \in I} \frac{-1}{|P(i)|} \sum_{p \in P(i)} \log \frac{\exp(\mathbf{z}_i \cdot \mathbf{z}_p / \tau)}{\sum_{a \in A(i)} \exp(\mathbf{z}_i \cdot \mathbf{a}_a / \tau)}
$$
where $\tau$ is a temperature scaling parameter. This objective pulls embeddings of solar panels with the same soiling level closer together while pushing apart embeddings of panels with different soiling levels, leading to features with higher intra-class similarity and inter-class distinction. This results in a model with improved accuracy and better generalization on challenging or ambiguous samples.
2. Dataset Curation and Experimental Setup
2.1 Quantitative Basis for Dust Degree Classification
To create a reliably labeled dataset, we established a quantitative link between visual soiling and performance loss. The power output of a solar panel under soiling is influenced by reduced irradiance and increased temperature. We calculate the power loss rate $\alpha_p$ at rated temperature as follows:
$$ \alpha_p = 1 – \frac{\alpha}{\alpha_d} $$
Here, $\alpha$ is the actual efficiency at rated temperature under soiled conditions, and $\alpha_d$ is the expected clean efficiency at rated temperature, estimated by:
$$ \alpha_d = \frac{P_{\text{max}} – P_{\text{max}} \cdot (T_b – T_s) \cdot R}{A_i \cdot F_n} $$
where $P_{\text{max}}$ is the maximum power at standard test conditions, $T_b$ is the backside temperature, $T_s$ is the rated temperature, $R$ is the power temperature coefficient, $A_i$ is the panel area, and $F_n$ is the irradiance. By analyzing the distribution of $\alpha_p$ from field data, we defined four distinct soiling levels, providing a solid, non-subjective ground truth for our image dataset.
| Dust Accumulation Level | Power Loss Rate ($\alpha_p$) | Visual Description |
|---|---|---|
| Clean | $\alpha_p \leq 8\%$ | Surface appears mostly clear with minimal visible dust. |
| Light Dust | $8\% < \alpha_p \leq 12\%$ | Thin, semi-transparent layer of dust; panel color slightly muted. |
| Moderate Dust | $12\% < \alpha_p \leq 16\%$ | Clearly visible dust layer; significant texture and color shift. |
| Heavy Dust | $\alpha_p > 16\%$ | Thick, opaque dust coverage; original panel color largely obscured. |
2.2 Data Collection and Augmentation
We collected images from multiple PV power plants, capturing solar panels under various natural soiling conditions. After filtering out poor-quality images (blurry, overexposed, etc.), we constructed a balanced dataset of 800 original images (200 per class). To prevent overfitting and enhance model robustness, we applied an extensive data augmentation pipeline, including random resizing, cropping, rotation, translation, contrast adjustment, and Gaussian noise addition. This expanded the dataset to 22,400 images (5,600 per class). The dataset was split into training and testing sets with an 8:2 ratio.
2.3 Implementation and Training Details
The model was implemented using PyTorch. We used Stochastic Gradient Descent (SGD) with a momentum of 0.9 and weight decay of 5e-4. The batch size was set to 16. A learning rate warm-up strategy was employed for the initial epochs, followed by a cosine annealing schedule to gradually reduce the learning rate from 0.01 to a minimum of 0.0001 over 800 epochs. The temperature $\tau$ in the SupCon loss was set to 0.1.
| Component | Parameter / Setting |
|---|---|
| Backbone | Modified ResNeXt-50 |
| Attention Module | Coordinate Attention (after initial Conv1) |
| Loss Function | Supervised Contrastive Loss (SupCon) |
| Optimizer | SGD (Momentum=0.9, Weight Decay=5e-4) |
| Batch Size | 16 |
| Initial Learning Rate | 0.01 |
| Learning Rate Schedule | Warm-up + Cosine Annealing |
| Minimum Learning Rate | 0.0001 |
| Total Epochs | 800 |
3. Results and Analysis
We conducted a series of ablation studies and comparative experiments to validate the contribution of each proposed component to the final performance in recognizing dust on solar panels.
3.1 Ablation Study on Architectural Improvements
The impact of our structural modifications is demonstrated in the following tables. Replacing the 7×7 convolution with three 3×3 convolutions (Decomposed Conv1) not only reduced parameters but also improved accuracy.
| Model Variant (Conv1) | Kernel Size & Stack | Accuracy | Parameter Reduction |
|---|---|---|---|
| Original ResNeXt | 7×7 (Single) | 83.5% | Baseline |
| Proposed Decomposition | 3×3 (Three, cascaded) | 84.7% | ~81% |
Similarly, fine-tuning the downsampling location from the first 1×1 convolution to the subsequent 3×3 convolution consistently provided a performance boost by preserving more spatial information during the initial feature transformation.
| Model Variant (Downsampling) | Downsample Location in Block | Accuracy |
|---|---|---|
| Original ResNeXt | First 1×1 Conv (stride=2) | 83.5% |
| Proposed Fine-tuning | 3×3 Conv (stride=2) | 85.1% |
3.2 Effectiveness of the Coordinate Attention Mechanism
We experimented with integrating the CA module at different stages of the network. The placement after the initial Conv1 module yielded the best results, indicating the importance of guiding spatial focus from the early stages of processing images of solar panels.
| Model | Attention Module & Placement | Accuracy |
|---|---|---|
| Baseline (Original ResNeXt) | None | 83.5% |
| ResNeXt-CA (after Conv1) | Coordinate Attention | 85.8% |
| ResNeXt-CA (after Conv2) | Coordinate Attention | 85.1% |
| ResNeXt-SE | Squeeze-and-Excitation | 84.2% |
| ResNeXt-CBAM | Convolutional Block Attention Module | 85.1% |
The superior performance of CA over SE and CBAM can be attributed to its ability to encode precise spatial coordinate information, which is crucial for locating irregular dust deposits on the surface of solar panels. This allows the model to “look” at the right places, such as panel edges where dust accumulates, rather than dispersing attention across the entire uniform background.
3.3 Impact of Supervised Contrastive Loss
Replacing the standard Cross-Entropy loss with the SupCon loss provided a substantial improvement, demonstrating the benefit of learning a well-structured feature space for the different soiling classes of solar panels.
| Model | Loss Function | Accuracy |
|---|---|---|
| ResNeXt (Baseline) | Cross-Entropy | 83.5% |
| ResNeXt (with SupCon) | Supervised Contrastive | 85.6% |
3.4 Comprehensive Model Comparison
We evaluated our final integrated model—incorporating decomposed convolution, fine-tuned downsampling, Coordinate Attention, and SupCon loss—against several state-of-the-art and traditional models. Our proposed model achieves a top accuracy of 90.7%, significantly outperforming all others.
| Model | Type | Accuracy on Solar Panel Dust Dataset |
|---|---|---|
| Random Forest | Traditional ML | 66.7% |
| K-Nearest Neighbors (KNN) | Traditional ML | 71.3% |
| Support Vector Machine (SVM) | Traditional ML | 75.8% |
| MobileNetV3 | Lightweight CNN | 83.9% |
| ResNet50 | Standard Deep CNN | 81.8% |
| InceptionV3 | Deep CNN | 84.2% |
| Inception-ResNetV2 | Deep CNN | 85.9% |
| Original ResNeXt50 | Deep CNN | 83.5% |
| Our Proposed Improved ResNeXt50 | Enhanced Deep CNN | 90.7% |
The 7.2 percentage point improvement over the original ResNeXt50 and the clear lead over other powerful architectures like Inception-ResNetV2 validate the effectiveness of our holistic approach. The model demonstrates a strong capability to generalize across the challenging variations present in real-world images of solar panels.
3.5 Computational Cost Analysis
The performance gains come with a minimal increase in computational overhead. The added parameters from the CA module and the slightly more complex loss computation are negligible in the context of modern GPU-based inference, making the model highly suitable for practical deployment.
| Metric | Original ResNeXt50 | Our Proposed Model |
|---|---|---|
| Model Parameters | ~25.0 MB | ~25.5 MB |
| Inference Time (per image) | ~10.1 ms | ~10.3 ms |
| Training Time (approx.) | ~70.0 min | ~71.3 min |
4. Conclusion
This work presents a robust deep learning framework for the automated visual recognition of dust accumulation levels on solar panels. By systematically enhancing the ResNeXt50 architecture—through convolutional decomposition and downsampling refinement, the integration of a Coordinate Attention mechanism for precise spatial focus, and the application of a Supervised Contrastive loss for learning a discriminative feature space—we have developed a model that significantly outperforms existing approaches. Achieving an accuracy of 90.7% on a realistic, multi-source dataset, our method provides a reliable, scalable, and cost-effective solution for the intelligent operation and maintenance of photovoltaic power plants. This technology can facilitate condition-based cleaning schedules, optimize energy output, and prevent potential damage to solar panels, thereby contributing directly to the economic and operational efficiency of solar energy generation. Future work will focus on enhancing the model’s robustness under extreme weather conditions (e.g., fog, snow, rain shadows) and integrating it into an end-to-end system for autonomous inspection using drones or robotic cleaners.