Current methods for detecting apple watercore are expensive and potentially damaging to the fruit. To determine whether different batches of apples are suitable for long-term storage or long-distance transportation, and to classify the apples according to quality level to enhance the economic benefits of the apple industry, it is essential to conduct non-destructive testing for watercore. This study proposes an innovative detection method based on optical parameter inversion and the MobileNetV3 model. Initially, a three-layer plate model of apples was constructed using the Monte Carlo method to simulate the movement of photons inside the apple, generating a simulated brightness map of photons on the apple’s surface. This map was then used to train the MobileNetV3 network with dilated convolution, resulting in a pre-trained model. Through transfer learning, this model was applied to measured spectral data to detect the presence of watercore. Comparative experiments were conducted to determine the optimal transfer strategy for the frozen layers, achieving model accuracy rates of 99.13%, 97.60%, and 95.32% for two, three, and four classifications, respectively. Furthermore, the model parameters were low at 7.52 M. Test results of this study confirmed the effectiveness and lightweight characteristics of the method that combines optical property parameter inversion, the DC-MobileNetV3 model, and transfer learning for detecting apple watercore. This model provides technical support to detect watercore and other internal diseases in apples.