The effectiveness of three MobileNet variations—MobileNetV1, MobileNetV2, and MobileNetV3—in correctly classifying dusty and immaculate Photovoltaic (PV) surfaces is investigated. To maintain PV panels' efficiency and maximize energy production, precise detection of dust accumulation is crucial. The demand for automated solutions arises from the inefficiency and high labor costs of conventional inspection techniques. A dataset consisting of 400 images, with an equal number of clean and dusty PV surfaces, was used to ensure a fair representation of both groups. Prior to being divided into training and validation sets, the images underwent preprocessing and normalization. Subsequently, each variant of MobileNet underwent training and evaluation using this dataset. Performance indicators such as training accuracy, validation accuracy, F1-score, and loss values were assessed. MobileNetV1 demonstrated superior performance, with a training accuracy of 88.53%, validation accuracy of 91.25%, and an F1-score of 0.9114. MobileNetV3 exhibited the lowest performance, achieving a training accuracy of 59.90%, a validation accuracy of 61.87%, and an F1-score of 0.6115. The study's findings establish that MobileNetV1 is the optimal model for accurately identifying dusty and clean PV surfaces. The research illustrates the viability of using Deep Learning (DL) algorithms in PV maintenance, and choosing the most suitable algorithm for doing the task.