The global energy crisis has intensified the search for efficient and sustainable renewable energy sources. Among these, photovoltaic (PV) technology, which directly converts solar energy into electricity, has become a focal point of attention and development. As the application of solar panel technology expands and industry demand surges, the analysis and detection of PV system performance have become critically important. The solar panel, as the core component of the system, has its output power constrained by numerous factors. Ash accumulation is an inevitable problem for solar panels during long-term operation. Due to variable weather conditions, the degree of ash accumulation on solar panels also constantly changes. Therefore, accurately identifying the ash accumulation degree on solar panels is essential for comprehensively understanding the operational status and performance of PV systems.
Current research on identifying ash accumulation on solar panels primarily utilizes image processing and computer vision techniques. Methods have been developed based on image algorithms, sensor technologies combined with computer vision, satellite image spectral analysis, and image segmentation using deep learning models. While these approaches show promise, their accuracy in detecting the degree of ash accumulation still requires significant improvement and offers substantial room for optimization. A common challenge identified across various image recognition applications, such as coal gangue detection, fruit recognition, and road navigation, is the susceptibility to lighting conditions. The presence of varying illumination introduces noise and reduces the reliability of visible light image analysis. This highlights the necessity for a comparative study of ash detection methods on solar panels to improve recognition accuracy under diverse environmental conditions.
To address these limitations, this research introduces a novel methodological approach. We first establish a quantifiable link between the physical state of the solar panel and its digital image. By numerically processing the grayscale histogram of solar panel images and introducing the concept of Average Grayscale Value (AGV), and combining this with experimental data from a dedicated ash accumulation visualization testbed, we verify a definitive correlation between the AGV and the ash density on the solar panel surface. Building upon this foundational relationship, we then explore the efficacy of multi-spectral image analysis. Using a dual-spectral image fusion experimental setup, we capture both visible light and infrared images of solar panels under natural soiling conditions. Five distinct image fusion algorithms are applied to combine the information from these two spectral bands. The performance in identifying the degree of ash accumulation is then evaluated and compared across seven image types: the original visible light images, the original infrared images, and the five types of fused images. The results demonstrate that infrared image analysis provides the most accurate and stable identification of ash accumulation degree on solar panels, with the least influence from variable solar irradiance, and reflects the most significant trend change in soiling level. This conclusion provides a crucial reference for the accurate monitoring and maintenance of photovoltaic systems.
1. Theoretical Foundation: Average Grayscale Value and Ash Density
The core of the proposed method lies in establishing a quantifiable metric derived from images that reliably corresponds to the physical amount of ash on a solar panel. Traditional visual inspection or simple image comparison is subjective. We propose using the Average Grayscale Value (AGV) of a processed solar panel image as this metric. The process to obtain the AGV involves several image processing steps.
First, a color image of the solar panel is converted to a grayscale image. This simplification reduces computational load and emphasizes structural features. The weighted average method is commonly used for this conversion, calculating the grayscale intensity Gray from the Red (R), Green (G), and Blue (B) channels of each pixel:
$$Gray = 0.3 \times R + 0.59 \times G + 0.11 \times B$$
Second, perspective transformation is applied to the grayscale image. When images are captured from an angle rather than directly overhead, perspective distortion occurs, making accurate panel-to-panel comparison difficult. A general perspective transformation maps coordinates (x, y) from the original image plane to coordinates (X’, Y’) on a transformed plane. This is represented by a transformation matrix:
$$
\begin{bmatrix} X \\ Y \\ Z \end{bmatrix} = \begin{bmatrix} a_{11} & a_{12} & a_{13} \\ a_{21} & a_{22} & a_{23} \\ a_{31} & a_{32} & a_{33} \end{bmatrix} \begin{bmatrix} x \\ y \\ 1 \end{bmatrix}
$$
where the resulting 2D coordinates are given by:
$$
X’ = \frac{X}{Z} = \frac{a_{11}x + a_{12}y + a_{13}}{a_{31}x + a_{32}y + a_{33}}, \quad Y’ = \frac{Y}{Z} = \frac{a_{21}x + a_{22}y + a_{23}}{a_{31}x + a_{32}y + a_{33}}
$$
By selecting four known reference points on the solar panel in both the original and desired “head-on” view, the transformation matrix coefficients can be solved, correcting the distortion and producing a standardized image of the panel surface.
Finally, the Average Grayscale Value is calculated from the corrected grayscale image. The grayscale histogram describes the frequency distribution of gray levels (from 0 for black to 255 for white) in the image. The frequency P(k) of gray level k is:
$$P(k) = \frac{n_k}{n}$$
where \(n_k\) is the number of pixels with gray level \(k\) and \(n\) is the total number of pixels. The AGV, denoted as \(M_{GV}\), is the weighted average of all gray levels:
$$M_{GV} = \frac{\sum_{i=0}^{255} i \times N_i}{n}$$
where \(i\) is the gray level and \(N_i\) is the pixel count at that level. Visually, an ash-covered solar panel typically appears lighter than a clean one, leading to a higher concentration of pixels in brighter gray levels and consequently a higher AGV. This establishes the theoretical basis for using AGV as an indicator of ash accumulation on solar panels.
2. Experimental Setup and Methodology
2.1 Visualization Testbed for Ash Density-AGV Correlation
To empirically validate the correlation between ash density and AGV, a dedicated visualization testbed was constructed. The testbed comprised multiple monocrystalline silicon solar panels placed in a natural outdoor environment to accumulate dust. Over a defined period, ash collection was performed at regular intervals. Before each collection, a high-resolution digital camera was used to capture an image of the soiled solar panels. The ash collection process involved meticulously wiping the panel surface, washing the collected ash into a container, drying it, and weighing it using a precision electronic balance. The ash density (\(\rho_{ash}\)) was then calculated by dividing the collected ash mass by the surface area of the solar panel:
$$\rho_{ash} = \frac{m_{ash}}{A_{panel}}$$
A total of 51 such data points (image and corresponding ash density) were collected. Each image underwent grayscale conversion and perspective transformation, and its AGV was computed. The resulting dataset allowed for the analysis of the relationship between \(\rho_{ash}\) and \(M_{GV}\).
2.2 Dual-Spectral Image Fusion Testbed for Method Comparison
To compare the effectiveness of different image types (visible, infrared, fused) for ash degree identification, a dual-spectral image fusion testbed was established. This setup featured three adjacent solar panels under identical environmental conditions. The panel on the left was manually cleaned daily to serve as a clean reference. The middle panel was cleaned every 15 days, and the right panel was cleaned every 30 days, allowing them to represent varying degrees of natural ash accumulation.
A key component was a thermal imaging dual-spectral network camera, capable of capturing both standard visible light images and long-wave infrared (thermal) images. This camera was installed facing the solar panels and protected in a weather-proof enclosure. Image capture was automated: visible light images were taken at noon daily, while infrared images were captured at night (e.g., 11 PM) to minimize the influence of direct solar heating on the thermal signature. All images were transmitted remotely for analysis. This setup generated a time-series dataset of co-registered visible and infrared images of the three panels with known cleaning schedules.
2.3 Image Fusion Algorithms
To create a comprehensive image dataset for comparison, five pixel-level image fusion algorithms were applied to pair the visible light image (V) and infrared image (IR) captured on the same day. Let \(V(m,n)\) and \(IR(m,n)\) represent the pixel intensity at location \((m, n)\) in the visible and infrared images, respectively, and \(F(m,n)\) represent the pixel in the fused image. The fusion methods are defined as follows:
- Pixel-wise Maximum: Selects the brighter pixel from the two sources.
$$F_{max}(m,n) = \max(V(m,n), IR(m,n))$$ - Pixel-wise Minimum: Selects the darker pixel from the two sources.
$$F_{min}(m,n) = \min(V(m,n), IR(m,n))$$ - Pixel-wise Average: Takes the simple mean of the two pixels.
$$F_{avg}(m,n) = \frac{V(m,n) + IR(m,n)}{2}$$ - Regional Energy Maximum: A more sophisticated method that considers a local neighborhood (e.g., 3×3 window). It calculates a “regional energy” for the corresponding window in each source image and selects the pixel from the image with higher local energy. The regional energy \(RE_A(i,j)\) for a window centered at \((i,j)\) in image A is:
$$RE_A(i,j) = \sum_{p,q \in N} w(p,q) \cdot L_{A,N}(i+p, j+q)$$
where \(w(p,q)\) is a weighting matrix (often a Gaussian kernel), \(N\) defines the neighborhood, and \(L_{A,N}\) is the pixel intensity. The fused pixel is chosen from the source image with the higher \(RE\) at that location. - Regional Energy Minimum: Similar to the above, but selects the pixel from the source image with the lower regional energy.
Applying these five methods to the paired visible and infrared images yielded five distinct sets of fused images. Together with the original visible light and infrared image sets, they formed a comprehensive dataset of seven image types for performance evaluation.
3. Results and Analysis
3.1 Correlation between Average Grayscale Value and Ash Density
The data from the visualization testbed confirmed a strong and clear relationship between the calculated Average Grayscale Value of the solar panel image and the physically measured ash density. As ash accumulated on the solar panels, the AGV consistently increased. This trend is intuitively logical: ash particles scatter and reflect light, causing the panel surface to appear lighter in the grayscale image, shifting the histogram towards higher intensity values and raising the average. The experimental data provided the empirical validation needed to proceed with using AGV as a reliable, non-contact proxy measurement for the degree of ash accumulation on solar panels. This established AGV as the primary metric for evaluating the different image analysis methods in the subsequent phase.
3.2 Performance Comparison of Seven Image Types
Using the dual-spectral testbed dataset, the AGV was calculated for the clean reference panel, the 15-day soiled panel, and the 30-day soiled panel across all seven image types over the experimental period. The AGV for the constantly cleaned panel remained relatively stable across all methods, serving as a control.
The key analysis involves the difference in AGV between a soiled panel and the clean reference panel (\(\Delta AGV = AGV_{soiled} – AGV_{clean}\)). This differential metric cancels out common background variations and directly highlights the change due to ash. For the rightmost panel (cleaned every 30 days, experiencing continuous soiling), \(\Delta AGV\) was calculated for each day and for each of the seven image types.
The trend of \(\Delta AGV\) over time for each image type was then fitted with a linear regression model. The coefficient of determination (\(R^2\)) of the linear fit was used as the primary metric to evaluate performance. A higher \(R^2\) value indicates that the AGV differential from that image type follows a more consistent and linear trend with time (and thus with increasing ash accumulation), implying greater accuracy and robustness for identification. The slope of the fitted line indicates the sensitivity of the method to ash accumulation; a steeper slope suggests the method can detect smaller changes in ash degree more clearly.
The performance results for the seven image types are summarized in the table below:
| Image Type | Coefficient of Determination (R²) |
|---|---|
| Visible Light Image | 0.914 |
| Infrared Image | 0.964 |
| Pixel-wise Maximum Fusion | 0.935 |
| Pixel-wise Minimum Fusion | 0.932 |
| Pixel-wise Average Fusion | 0.928 |
| Regional Energy Maximum Fusion | 0.937 |
| Regional Energy Minimum Fusion | 0.935 |
The analysis leads to several critical findings:
- Superiority of Infrared Imaging: The infrared images achieved the highest \(R^2\) value of 0.964. This signifies that the AGV differential derived from thermal images had the strongest linear correlation with the progression of ash accumulation. Furthermore, the fitted line for infrared data had the steepest slope, meaning it exhibited the greatest sensitivity and displayed the most pronounced change in the calculated ash degree over time.
- Limitation of Visible Light Images: While visible light images performed reasonably well (\(R^2 = 0.914\)), they were consistently outperformed by infrared. This is attributed to the inherent sensitivity of visible light to ambient illumination conditions (sun angle, cloud cover, time of day), which introduces noise and variance in the AGV that is unrelated to ash.
- Performance of Fused Images: The five image fusion methods yielded results that were generally better than visible light alone but did not surpass the performance of the standalone infrared images. The most complex method, Regional Energy Maximum fusion, yielded the best result among fusions (\(R^2 = 0.937\)), yet it still fell short of the infrared baseline. This suggests that for the specific task of ash degree quantification on solar panels, the simpler thermal image contains more directly useful and consistent information than any of the tested combinations with visible light.
- Robustness of Infrared: The superior performance of infrared images can be explained by their fundamental operating principle. Thermal cameras detect long-wave infrared radiation emitted by objects based on their temperature and emissivity. Ash accumulation on a solar panel changes the surface thermal properties (emissivity, heat capacity) and affects the operating temperature of the panel cells themselves. This thermal signature appears to be a more stable and direct indicator of surface contamination than reflected visible light, which is highly dependent on external irradiance.
4. Conclusion
This study presented a systematic approach for identifying the degree of ash accumulation on solar panels based on digital image analysis. The foundational work established a quantifiable and verified correspondence between the physical ash density on a solar panel and the Average Grayscale Value (AGV) calculated from a geometrically corrected grayscale image of the panel. This established AGV as a valid non-contact metric for ash degree.
Building on this, a comprehensive comparative evaluation was conducted using a dual-spectral imaging testbed. The performance of seven different image types—original visible light, original infrared, and five varieties of fused images—was assessed for their accuracy and sensitivity in tracking ash accumulation over time. The key conclusion is unequivocal: infrared image analysis provides the most effective method for identifying the ash accumulation degree on solar panels. It achieved the highest accuracy (as indicated by the highest \(R^2\) value of 0.964), was the least affected by variable solar irradiance that plagues visible-light methods, and demonstrated the greatest sensitivity to changes in ash level.
The findings have significant practical implications. For photovoltaic plant operation and maintenance (O&M), integrating thermal imaging cameras for periodic inspection can provide a more reliable, objective, and quantifiable assessment of soiling levels across solar panels than visual inspection or visible-light camera systems. This can optimize cleaning schedules, prioritize cleaning efforts for the most affected areas, and ultimately improve the energy yield and economic return of PV installations. The methodology and conclusions offer a solid theoretical foundation and a practical tool for advanced ash monitoring in the solar energy industry.