The global transition towards sustainable energy has positioned solar power as a cornerstone technology. As the world’s largest market and manufacturing base for photovoltaic (PV) generation, efficient operation and maintenance of solar panel arrays are critical. However, solar panels, typically installed in open environments, are inevitably exposed to natural elements leading to the accumulation of dust and debris on their surfaces. This dust accumulation poses a significant threat to the efficiency and longevity of PV systems. It directly reduces light transmittance, diminishing the power output of the solar panels. Furthermore, uneven soiling can lead to hotspot formation, accelerating panel degradation, and potentially causing hardware failures that compromise the entire system’s reliability. Therefore, the timely and accurate identification of a solar panel’s dust accumulation state is paramount for ensuring optimal energy yield and facilitating predictive maintenance.
Traditional methods for assessing solar panel soiling rely heavily on manual inspection, which is labor-intensive, time-consuming, inconsistent, and costly. These methods are ill-suited for the scale and demand of modern, distributed solar farms, failing to meet the requirements for real-time monitoring and intelligent operation. Consequently, there is a pressing need for automated, reliable, and efficient computer vision-based solutions. While prior research has explored using deep learning for this task, existing approaches often struggle with balancing high accuracy, computational efficiency, and model complexity. Some models achieve reasonable accuracy but are too parameter-heavy for deployment on edge devices, while lighter models sacrifice performance. This work addresses these challenges by proposing a novel convolutional neural network (CNN) architecture based on an optimized Inception structure, specifically designed for robust and efficient solar panel dust state classification.
The detrimental effects of dust on a solar panel are well-documented. The layer of dust acts as a barrier, scattering and absorbing incident sunlight before it can reach the photovoltaic cells. The power loss is not linear and can be severe depending on the dust type, thickness, and environmental conditions. Beyond mere output reduction, partial shading from non-uniform dust layers can force certain cells within a panel to operate in reverse bias, dissipating power as heat and creating hotspots. These hotspots are a primary cause of long-term performance degradation and physical damage like cell cracking and encapsulant delamination. From a system perspective, undetected severe soiling across multiple strings can lead to underperformance alarms or even trigger protective shutdowns. Thus, automating the detection and classification of dust levels on a solar panel is not just an optimization task but a necessity for asset management and fault prevention.
Previous attempts to solve this problem using machine learning and computer vision have shown promise but also reveal clear limitations. Early approaches focused on image processing techniques, such as segmenting the solar panel from its background in specific color spaces (e.g., YCbCr) and then extracting hand-crafted features like texture descriptors (e.g., Local Binary Patterns) for classification. While computationally simpler, these methods are highly sensitive to lighting variations, panel model differences, and background clutter, leading to generalization issues and accuracy plateaus typically below 90%. The advent of deep learning, particularly CNNs, offered a paradigm shift. Researchers have employed architectures like ResNet50 for dust state recognition, achieving foundational results. However, classifying dust into many fine-grained states (e.g., 11 levels) with high accuracy remains challenging, with reported accuracies around 81% for complex networks. Improvements using architectures like ResNeXt with attention mechanisms have pushed accuracy higher (e.g., ~91%), but often at the cost of increased model complexity and parameter count, hindering deployment speed. Lightweight models like MobileNet offer fast inference but may not capture the subtle inter-class differences between light, moderate, and heavy dust on a solar panel, resulting in lower accuracy. The core challenge is to design a model that maintains the powerful feature extraction capabilities of deep networks while being parameter-efficient and fast enough for practical, real-world application on constrained hardware at solar power plants.
The proposed methodology centers on a custom deep learning model built upon the Inception paradigm, renowned for its efficiency in width and depth. The Inception module’s fundamental strength lies in its multi-path design, which processes input data through convolutional filters of varying sizes (e.g., 1×1, 3×3, 5×5) and a pooling operation simultaneously. These parallel streams are then concatenated, allowing the network to capture features at multiple scales within a single layer. This design leads to a wider, rather than just deeper, network that is computationally frugal due to the dimensionality reduction performed by 1×1 convolutions before more expensive operations. The proposed model incorporates several key innovations to adapt and enhance this foundation for the specific task of solar panel inspection.
The first major innovation is the redesign of the initial processing block, often called the Stem. The primary role of the Stem is to perform aggressive downsampling and preliminary feature extraction from the raw input image of the solar panel, reducing spatial dimensions to ease the computational load on subsequent, more complex modules. The standard Stem often uses small convolutional filters. In our optimized design, we strategically employ larger kernel sizes (5×5) in the initial convolutional layers. This modification serves a critical purpose: it immediately increases the effective receptive field of the network at its early stages. A larger receptive field allows the model to integrate contextual information from a broader area of the solar panel image from the very beginning. This is particularly beneficial for discerning dust patterns, as their classification may depend on the global distribution of dust relative to the panel’s grid lines and borders, not just local texture. The structure of the optimized Stem module is summarized in its computational flow:
Stem Module Computational Flow:
Input (299x299x3) → [5×5 Conv, stride 2] → [5×5 Conv] → [5×5 Conv] → … → Output (Reduced Spatial Dimensions, Increased Channels).
The second cornerstone of our architecture is the integration of a lightweight yet powerful attention mechanism, which we term the Swish-SE block. Attention mechanisms enable the model to dynamically highlight the most informative features and suppress less relevant ones. For a solar panel image, critical features might be the contrast between dusty and clean cells, the obscuration of the silver grid lines, or specific texture patterns. The Squeeze-and-Excitation (SE) mechanism is a channel-wise attention module that first squeezes global spatial information into a channel descriptor via global average pooling. It then excites this descriptor through a small multi-layer perceptron to produce a vector of scaling factors for each channel. These factors recalibrate the original feature maps, emphasizing important feature channels. We enhance this mechanism by replacing the standard ReLU activation in the excitation pathway with the Swish function. The Swish activation, defined as $$ f(x) = x \cdot \sigma(\beta x) $$, where $\sigma$ is the sigmoid function and $\beta$ is a learnable parameter, often provides smoother gradients and better performance in deeper networks. The Swish-SE block can be integrated into the Inception branches, allowing the network to focus computational resources on the most salient features for distinguishing dust states on the solar panel. The operation of the Swish-SE block is described by the following sequence for a feature map $\mathbf{U}$ with channels $C$:
1. Squeeze: Generate channel-wise statistics $\mathbf{z} \in \mathbb{R}^C$ via Global Average Pooling: $z_c = \frac{1}{H \times W} \sum_{i=1}^{H} \sum_{j=1}^{W} u_c(i, j)$.
2. Excitation: Learn a channel-wise modulation weight $\mathbf{s}$: $\mathbf{s} = \sigma(\mathbf{W}_2 \cdot \text{Swish}(\mathbf{W}_1 \mathbf{z}))$, where $\mathbf{W}_1 \in \mathbb{R}^{\frac{C}{r} \times C}$ and $\mathbf{W}_2 \in \mathbb{R}^{C \times \frac{C}{r}}$.
3. Scale: Recalibrate the original feature map: $\tilde{\mathbf{x}}_c = s_c \cdot \mathbf{u}_c$.
The complete network architecture is constructed by stacking the optimized Stem module, followed by a series of specialized Inception modules (Inception-A, Inception-B, Inception-C). These subsequent modules utilize factorized convolutions (e.g., replacing a 5×5 convolution with two 3×3 convs, or asymmetric convolutions like 1×7 followed by 7×1) to maintain representational power while controlling parameter growth. Swish-SE blocks are strategically placed within these modules. The final architecture culminates in a global average pooling layer and a fully connected layer with a softmax activation for the three-class prediction (Clean, Dusty, Heavily Soiled). The overall design ensures a deep and wide feature hierarchy capable of capturing both fine-grained textures and global contextual cues from the solar panel image.
A critical component of any deep learning solution is the dataset. For this study, a dedicated dataset of solar panel images under various soiling conditions was compiled. Images were sourced from operational solar power plants, captured via fixed monitoring cameras and mobile devices over several months to ensure diversity in lighting, angle, and background. After rigorous curation to remove blurry, redundant, or occluded images, the raw dataset contained 473 images. Each solar panel image was manually annotated into one of three states: No Dust (clean surface with visible gridlines), Dusty (visible dust layer partially obscuring gridlines), and Heavily Soiled (gridlines largely or completely obscured by a thick dust layer). To combat overfitting and improve model generalization, extensive data augmentation was applied, including random scaling, brightness/contrast adjustment, rotation, and cropping. This process expanded the dataset to 3,015 images, with 1,005 images per class. The dataset was then partitioned into training (80%), validation (10%), and testing (10%) sets, ensuring no data leakage between splits.
The training process was carefully designed to ensure stable convergence and optimal performance. All input solar panel images were resized to 299×299 pixels to match the network’s input expectations. The model was trained for 60 epochs with a batch size of 32. Batch Normalization was employed after each convolutional layer to accelerate training and improve stability. The Adaptive Moment Estimation (Adam) optimizer was selected for its adaptive learning rate properties. The Adam update rules for parameters $\theta$ at time step $t$ are as follows:
$$ m_t = \beta_1 m_{t-1} + (1 – \beta_1) g_t $$
$$ v_t = \beta_2 v_{t-1} + (1 – \beta_2) g_t^2 $$
$$ \hat{m}_t = \frac{m_t}{1 – \beta_1^t} $$
$$ \hat{v}_t = \frac{v_t}{1 – \beta_2^t} $$
$$ \theta_{t+1} = \theta_t – \frac{\eta \cdot \hat{m}_t}{\sqrt{\hat{v}_t} + \epsilon} $$
where $g_t$ is the gradient, $m_t$ and $v_t$ are estimates of the first and second moments, $\beta_1=0.9$, $\beta_2=0.999$, $\epsilon=10^{-8}$, and $\eta$ is the learning rate.
A sophisticated learning rate schedule was implemented to further enhance training. For the first 20 epochs, a Warmup strategy was used, linearly increasing the learning rate from a small value ($\eta_{start}=5\times10^{-4}$) to the base learning rate ($\eta_0=0.001$). This prevents instability in the early phase of training. For the remaining 40 epochs, a polynomial decay schedule was applied, gradually reducing the learning rate to allow for fine-tuning. The combined schedule is defined as:
$$ \eta_{\text{warm}}(e) = \eta_{start} + (\eta_0 – \eta_{start}) \times \frac{e}{E_{\text{warm}}} \quad \text{for} \quad e \leq E_{\text{warm}} $$
$$ \eta_{\text{poly}}(e) = \eta_0 \times \left(1 – \frac{e – E_{\text{warm}}}{E_{\text{total}} – E_{\text{warm}}}\right)^{p} \quad \text{for} \quad e > E_{\text{warm}} $$
where $e$ is the current epoch, $E_{\text{warm}}=20$, $E_{\text{total}}=60$, and $p=0.9$.
Given the multi-class nature of the problem, the Categorical Cross-Entropy loss function was used to measure the discrepancy between the predicted probability distribution $\mathbf{P}$ and the ground-truth one-hot encoded label $\mathbf{y}$. For a batch of $N$ samples, the loss $L$ is computed as:
$$ L = -\frac{1}{N} \sum_{i=1}^{N} \sum_{c=1}^{C} y_{i,c} \log(p_{i,c}) $$
where $C=3$ is the number of dust state classes, $y_{i,c}$ is 1 if sample $i$ belongs to class $c$, and $p_{i,c}$ is the predicted probability for that class.
The model was implemented using the PyTorch framework and trained on a system with an NVIDIA GeForce RTX 3060 Ti GPU. To rigorously evaluate the proposed model’s performance, it was compared against two well-established and relevant CNN architectures: Inception-V3 and MobileNet-V2. Inception-V3 represents a powerful and accurate family of models, while MobileNet-V2 is a benchmark for lightweight, efficient networks. The comparison focused on key metrics: classification accuracy, precision, recall, and F1-score on the held-out test set. Additionally, practical deployment metrics such as the number of trainable parameters, training time per epoch, and average inference time per solar panel image were recorded.
The experimental results demonstrate the clear superiority of the proposed optimized Inception architecture for the solar panel dust recognition task. As shown in Table 1, our model achieved a test accuracy of 97.05%, significantly outperforming both MobileNet-V2 (91.14%) and Inception-V3 (95.41%). It also attained the highest precision (97.38%) and F1-score (98.67%), indicating a robust balance between correctly identifying positive cases and minimizing false positives across all three dust states.
| Model | Accuracy (%) | Precision (%) | Recall (%) | F1-Score (%) |
|---|---|---|---|---|
| MobileNet-V2 | 91.14 | 91.14 | 100 | 95.36 |
| Inception-V3 | 95.41 | 95.08 | 100 | 97.48 |
| Proposed Model | 97.05 | 97.38 | 100 | 98.67 |
The practical efficiency metrics, presented in Table 2, are equally compelling. The proposed model successfully achieves a “sweet spot” between performance and efficiency. With only 3.59 million parameters, it is over 5.7 times more parameter-efficient than the standard Inception-V3 (20.78M) while being significantly more accurate than the ultra-lightweight MobileNet-V2 (2.23M). This parameter efficiency translates directly into faster inference. The proposed model can process a solar panel image in approximately 9.03 milliseconds, which is about 27% faster than Inception-V3 (11.43 ms) and maintains a practical speed close to MobileNet-V2 (5.66 ms) while offering far superior accuracy.
| Model | Parameters (Millions) | Training Time (s/epoch) | Inference Time (ms/image) |
|---|---|---|---|
| MobileNet-V2 | 2.23 | 10.89 | 5.66 |
| Inception-V3 | 20.78 | 16.47 | 11.43 |
| Proposed Model | 3.59 | 15.68 | 9.03 |
To validate the contribution of each architectural innovation, an ablation study was conducted. The results, summarized in Table 3, clearly show the incremental benefits. A baseline model using a standard Inception structure without the optimized Stem or Swish-SE achieved an accuracy of 89.83%. Incorporating the optimized Stem module alone improved accuracy to 92.79%, confirming that the larger initial receptive field aids in capturing better global features from the solar panel. Adding the Swish-SE attention mechanism to the baseline (without the optimized Stem) yielded a substantial boost to 95.41%, highlighting the importance of channel-wise feature recalibration. Finally, the full proposed model, combining both the optimized Stem and the Swish-SE mechanism, achieved the best performance at 97.05% accuracy. This demonstrates that the two components work synergistically: the Stem provides better initial feature maps, and the Swish-SE mechanism allows the network to intelligently focus on the most discriminative aspects of those features for dust state classification.
| Model Configuration | Accuracy (%) | Precision (%) | F1-Score (%) |
|---|---|---|---|
| Baseline (Original Inception) | 89.83 | 89.83 | 94.64 |
| Baseline + Optimized Stem | 92.79 | 92.46 | 96.08 |
| Baseline + Swish-SE | 95.41 | 95.41 | 97.65 |
| Full Proposed Model (Baseline + Optimized Stem + Swish-SE) | 97.05 | 97.38 | 98.67 |
The convergence behavior of the proposed model was also examined. The learning curves showed a smooth and steady decrease in training loss, with the validation accuracy plateauing at a high level after approximately 40-50 epochs, indicating effective training without severe overfitting. This stability can be attributed to the combination of Batch Normalization, the carefully designed learning rate schedule, and the regularization implicit in the architecture’s design.
In conclusion, this work presents a novel, efficient, and highly accurate deep learning model for the automatic recognition of dust accumulation states on solar panels. By innovatively optimizing the Inception architecture—specifically through a Stem module with an enlarged receptive field and the integration of a Swish-activated Squeeze-and-Excitation attention mechanism—the proposed model achieves state-of-the-art performance. It strikes an optimal balance, delivering superior classification accuracy (97.05%) compared to strong benchmarks while maintaining a low parameter footprint (3.59M) and fast inference speed suitable for real-time or near-real-time applications. This makes it a compelling solution for integration into the intelligent operation and maintenance (O&M) platforms of modern solar power plants. By enabling automated, precise, and frequent soiling assessment, this technology can help optimize cleaning schedules, maximize energy yield, prevent damage, and reduce O&M costs, contributing to the overall efficiency and reliability of solar energy generation. Future work will focus on expanding the dataset to include more granular dust levels, varying panel types, and extreme weather conditions, as well as exploring model quantization and pruning for deployment on ultra-low-power edge devices directly at the solar farm.