Abstract: 【Objective】Based on deep learning and hyperspectral imaging technology， a non-destructive detection model for pre-harvest citrus sugar content （SC） adapted to orchard environments was established， providing theoretical reference and technical support for citrus maturity assessment and intelligent harvesting in orchards.【Method】The region selection method and the threshold segmentation method were employed to extract the region of interest （ROI） from citrus hyperspectral images respectively， to obtain the effective spectral reflectance area of citrus fruit and calculate the raw spectra. Six spectral preprocessing methods， including multiplicative scatter correction （MSC）， standard normal variate （SNV）， Savitzky-Golay smoothing （SG）， wavelet transform （WT）， first derivative （FD）， and the SGMSC combination， were applied to correct the raw spectra. Four feature band extraction algorithms， namely competitive adaptive reweighted sampling （CARS）， random frog （RF）， synergy interval partial least squares （SIPLS）， and correlation coefficient （CORR）， were used to select feature bands highly correlated with citrus SC， and combined with spectral preprocessing. By integrating feature augmentation and residual blocks into a convolutional neural network （CNN）， the CNN-FA-RB model was established. Meanwhile， seven models， including partial least squares regression （PLSR）， support vector regression （SVR）， feed-forward neural network （FFNN）， CNN， AlexNet， ResNet18， and SENet18， were constructed for comparison.【Result】The ROI extracted by the threshold segmentation method yielded superior modeling performance， with the coefficient of determination （R²） improved by an average of 19.33% compared with the region selection method. The SG+CARS combination was identified as the optimal preprocessing strategy， extracting 25 feature bands covering a broad range （399.2-963.0 nm） and capturing sugar-related absorption peaks （C-H/O-H bond vibrations）. The PLSR model test showed an R² of 0.709 and a root mean square error （RMSE） of 0.894， demonstrating effective smoothing of high-frequency noise and selection of feature bands strongly correlated with SC. CNN-FA-RB model achieved the highest SC detection accuracy with a relatively fast training speed. Under orchard conditions， its R² was 0.816， RMSE was 0.751， mean absolute error （MAE） was 0.576， mean absolute percentage error （MAPE） was 5.616%， and the training speed was 13.158 epoch/s. Compared with the CNN model， R² increased by 8.66%， RMSE decreased by 13.97%， MAE decreased by 15.79%， MAPE decreased by 14.32%， and the training speed improved by 3.71 times.【Conclusion】The constructed CNN-FA-RB model can overcome outdoor environmental interference and achieve accurate SC prediction. The integrated scheme of threshold segmentation + SG + CARS + CNN-FA-RB can effectively realize non-destructive detection of pre-harvest citrus SC， providing technical support for citrus maturity assessment， robotic harvesting， orchard monitoring， and other application scenarios.