Aiming at the problem of low accuracy of traditional classification methods in identifying citrus planting spatial information, this paper proposes a county-level citrus planting spatial information extraction method combining multi-temporal Sentinel-2 and DeepLabV3+. Firstly, the optimized lightweight network MobileNetV2 is used as the backbone network, and the CBAM (Convolutional Block Attention Module) attention mechanism module is embedded to construct the improved lightweight DeepLabV3+; Then, the multi-temporal Sentinel-2 images are used to integrate the original band and spectral index to form the feature data set, and the optimal feature combination and phase of the model classification are determined through experimental comparative analysis; Finally, the image of the study area is segmented into a set of images to be predicted with overlap, and the optimal classification model is used to predict and splice to obtain the results of citrus orchard extraction. The results show that: 1) the extraction accuracy of the improved lightweight DeepLabV3+ is higher than that of DeepLabV3+ and Random Forest model. In the feature combination with the addition of red edge index RESI in the B2-B8A band, the OA can up to 91.1% , and the optimal extraction phase is November. 2) The extraction effect of the overlap prediction method is better than that of the direct prediction method, the edge error of the extracted citrus orchard map spots is basically eliminated, and the relative error between the extracted area and the statistical data of the whole region is kept within ±0.04%, which is of high applicability. This method can provide a reference for the automatic monitoring and planting planning of citrus orchards in the county area in southern China.