Abstract: Soil moisture is one of the most crucial indicators to develop the intelligent irrigation in kiwifruit orchards. However, the complex kiwifruit trees with the canopy, varying coverage and substantial shading have posed some challenges on the accurate prediction of the soil moisture content. Current algorithms are still lacking on the canopy image to estimate the root soil water content (RSWC) of kiwifruit trees using canopy spectral information in the field of UAV remote sensing. In this study, a Compound Visual Convolutional Regression Network (CVCRNet) was proposed to combine two sizes of convolutional layers, in order to extract convolutional features from image data. The fully connected layers were used to reduce the dimensionality of convolutional features. Thus, multispectral images were directly analyzed for RSWC inversion. Since there was no pooling layer in the network, all data within multispectral images was fully utilized to enhance the accuracy of inversion. Multispectral images of the canopy and RSWC were collected at a depth of 40 cm during the swelling period (May-September) of Xuxiang kiwifruit trees. The canopy image was processed and normalized to directly served as the input, in order to eliminate the manual feature extraction or complex structural analysis of the fruit tree canopy, as well as the correlation of vegetation indices. Deep convolutional features were extracted from the Red-Green-Near Infrared (RGN) images of kiwifruit tree canopies, in order to train the remote sensing dataset of kiwifruit orchard. The RSWC gradient maps were obtained by cubic spline interpolation. A gradient map of kiwifruit tree distribution was then generated to reflect the actual situation of water control, where the RSWC gradient maps was overlapped with the original ones. As such, the field application of the CVCRNet inversion was realized in this case. Additionally, the performance of RSWC was compared on the vegetation indices and traditional numerical models. A Multilayer Perceptron (MLP) network was introduced to establish a dual-index estimation model using Renormalized Difference Vegetation Index (RDVI) and Green Normalized Difference Vegetation Index (GNDVI). The data training showed that the epoch69 weight was selected to optimize the loss and explained variance score of the training and testing sets during CVCRNet training. The Mean Squared Error (MSE) of the training set was 1.358, with an Explained Variance Score (EVS) of 0.710, while the MSE and EVS were 0.889 and 0.737, respectively, for the testing set. The results showed that the leaves were selected in the center of the canopy in the image using CVCRNet, and then the greater weight was assigned to their reflectance information, leading to the inversion superior to traditional vegetation indices. The coefficient of determination (R²) for the CVCRNet test set was 0.827, with a Root Mean Squared Error (RMSE) of 0.787%; R² was 0.743 and RMSE was 0.887% for all samples. The MLP test set yielded an R² of 0.759 and an RMSE of 0.983%; R² was 0.565, and RMSE was 2.516% for all samples. There was the significant lower CVCRNet inversion under bare ground, indicating only suitable for use during periods of high canopy coverage. The CVCRNet with images as the input was reduced the loss of multispectral image information in the complex distribution of kiwifruit orchard canopies. Canopy information extraction was enhanced to obtain the better soil moisture prediction. The soil data inversion was achieved in the complex canopy scenarios. The CNN networks can be expected for the canopy information inversion.