Traditional hyperspectral single point detection methods of obtaining chlorophyll content of longan leaves are inefficiency, low accuracy and time consuming. Combined with the state-of-the-art deep learning technology, a distribution model of chlorophyll content for longan leaves based on convolution neural networks (CNN) and deep neural networks (DNNs) was proposed. Firstly, the spectral noise was reduced by Savitzky-Golay filter. The initial features extraction was carried out by using a principle component analysis (PCA) to identity a number of potential characteristic wavelengths (483nm, 518nm, 625nm, 631nm, 642nm and 675nm) according to the weight coefficient distribution curve of the first three principle component images (PC1, PC2 and PC3) under the full wavelengths. For the characteristic spectral images and the principal component images, the texture based on the gray level co-occurrence matrix was extracted from those images, and the structure information of those images was also extracted based on CNN simultaneously. Among the 300 samples, there were total of 1800 spectral images and 900 principal component images, in which a sample corresponding to six characteristic spectrum images and three PCA images for a sample. Gray-level co-occurrence matrix (GLCM) was utilized to extract texture features. The hyperspectral wavelength feature data, texture data, images structure data and the combined data were utilized to develop particle swarm optimization-support vector regression (PSO-SVR) and independent component analysis-deep neural networks (ICA-DNNs), respectively. The particle swarm optimization (PSO) was introduced to intelligently optimize the parameters (γ and c) in the SVR model to find the optimum. Some main conclusions ware obtained: the performance of PSO-SVR model based on characteristic spectrum was the best, and the coefficient of determination (R2) of calibration set and validation set of the entire growth state were up to 0.8220 and 0.8152, respectively. Multi-source data fusion performance of ICA-DNNs was the best, and the precision of ICA-DNNs model was improved based on feature spectrum, texture feature and images structural feature, the R2 of calibration set and validation set were 08358 and 0.8210, respectively. Compared with traditional methods of SVR, DNNs was more robust for lager data set. The longan leaf textures and CNN characteristics were less relevant to longan chlorophyll content distribution. Chlorophyll content distribution region for tender, pale and dark green longan leaves were: mesophyll of the leaf-root, part of mesophyll of lateral veins and the whole mesophyll. Finally, hyperspectral technology could obtain accurate chlorophyll content of longan leaves rapidly, quantitatively and non-destructively. The research result can provide a theoretical basis for nutrition surveillance of longan growth and longan disease such as leaf spot, brown spot and leaf blight.