Abstract: The spectral detection model for avocado stem-end rot is highly susceptible to variations in growth environments across different production regions, leading to insufficient model generalization ability. To improve the detection accuracy of avocado stem-end rot and enhance the model’s adaptability to samples from different regions, a spectral detection model was developed using visible-near-infrared spectroscopy, combined with deep learning and transfer learning methods. Experimental data were collected from avocados produced in two distinct regions of Peru, comprising a total of 1994 samples, including both healthy and stem-end rot samples. Multiple data preprocessing and feature selection methods were evaluated to optimize the model's performance. The results demonstrated that the model accuracy was significantly improved using standard normal variate transformation and subwindow permutation analysis to preprocess the spectral data. Building on this, transfer learning was introduced to fine-tune the pretrained model, increasing its recognition accuracy in the target region from 0.4455 to 0.7268. The research findings validate the effectiveness of combining deep learning with fine-tuning in improving cross-domain recognition capabilities, providing reliable technical support and practical guidance for region-adaptive detection of avocado diseases.