Abstract: Parcel-based classification of crops is paramount to quantify changes in ecological systems and improve management strategies in precision agriculture. Specifically, the obtained location and boundary of farmland together with crop types can contribute to the specific payment of planting subsidies and resource survey. Multi-source high-spatial and temporal resolution satellite images can provide an effective way to realize parcel-based crop mapping. However, some deficiencies still remain in the parcel extraction of farmland and construction of spatiotemporal features. In this present study, a novel model was constructed to implement a parcel-based classification of crops using the spatiotemporal collaborated satellite data with high-spatial and temporal resolution. Four steps were included in a parcel-based crop mapping: 1) A D-LinkNet deep learning model was selected to extract the parcels from the 0.6m high-spatial-resolution Google Earth images; 2) Time series data set was constructed for each parcel using multi-source observations from Landsat 8 and Sentinel-2 satellite, where the tiles with high cloud cover were removed from the images; 3) A weighted Double-Logistic fitting was utilized to reconstruct the parcel-based Normalized Difference Vegetation Index (NDVI) time series for the extraction of phenological parameters, such as the duration of the growth cycle, the time of growth starting and ending, thereby calculating spectral indexes from Landsat8 and Sentinel-2 multispectral data; 4) A Mean Decrease Accuracy (MDA) indicator was used to estimate the feature importance. A field experiment was also conducted to collect the data of crop types for the training of random forest classification model in a parcel-based crop mapping. The Fusui County in Guangxi Zhuang Autonomous Region of China was taken as the study area. There was a relatively complex planting structure in the study area, because it was cloudy and rainy with the rainfall days of about 130-220 d, as well as the diverse and complex topography with a high level and fragmentation. The dominated crops included sugarcane, paddy rice, banana, and orange. The results showed that the farmland parcels were well segmented in the whole images, while the crop distributions of resultant parcels were also well extracted by a D-LinkNet deep learning model, with an edge accuracy of 84.54% and a produce accuracy of 83.06%, compared with the conventional multi-scale segmentation. Phenological features were achieved in the reconstructed NDVI time series of sugarcane, paddy rice, and banana. The NDVI of sugarcane and paddy rice first increased and then decreased significantly. The growth season of sugarcanes started from March to the following March. In addition, the growth season of paddy rice lasted for about 3-4 months, in which there was the most intense change in the NDVI time series. There was a relatively steady state in the reconstructed NDVI time series of evergreen eucalyptus and orange in the whole year. The eucalyptus with high vegetation cover showed high NDVI values during the observation period. The MDA indicator demonstrated that the images captured in summer and autumn were better for the crop classification in the study area. A best performance of classification was achieved to combine the phenological and spectral red-edge features in Sentinel-2 images. The overall accuracy reached 88%, and the accuracy of sugarcane reached over 95% in the study areas. The crop mapping indicated that sugarcane was spatially distributed around the whole study area, including plain and mountainous areas. The planting area of sugarcane accounted for nearly 70%, orange for 18.6%, and paddy rice for 7.12% of farmland. Furthermore, the paddy rice was mostly distributed near the settlement places. Consequently, the Landsat 8 and Sentinel-2 multi-source observations can be expected to successfully extract the phenological features in the parcel-based crop mapping. The finding can provide a series of practical schemes to acquire parcel-based crop distribution.