Aiming at the problem of accurate management of orchard under the background of complex environment such as bare soil, shelter, fruit tree shadow and weeds, the image data of apple orchard was obtained by UAV equipped with multispectral camera, and then the fruit tree pixels were extracted and the global fruit tree row navigation line was extracted. The obtained multispectral image data were preprocessed to obtain digital orthophoto map (DOM) and digital surface model (DSM) image. The normalized difference greenness index (NDGI) and ratio vegetation index (RVI) distribution maps that were easy to distinguish apple trees from weeds were selected and calculated, and the NDGI and RVI images were fused with DSM image; the excess green (EXG) index and normalized difference canopy shadow index (NDCSI) were comprehensively used to eliminate the pixels such as soil, shelter and shadow in the fusion image by threshold segmentation method, so as to reduce the interference of non-vegetation mixed pixels on the classification and recognition of fruit trees. Support vector machine (SVM), random forest (RF) and maximum likelihood (MLC) method were used to extract the apple trees in the fused image and ordinary orthophoto respectively, calculate the confusion matrix, and compare and evaluate the recognition accuracy. The experimental results showed that the MLC method had the best recognition effect on fruit trees in the fused image, and its user accuracy, mapping accuracy, overall classification accuracy and Kappa coefficient were 88.57%, 93.93%, 93.00% and 0.8824, respectively; compared with ordinary orthophoto images, the final fusion image constructed effectively improved the recognition accuracy of the three methods. The fused image improved the user accuracy of RF method the most, which was 27.12 percentage points; the mapping accuracy of SVM method was improved the most, which was 9.03 percentage points; the overall classification accuracy of the three methods was improved by 13.00 percentage points; the Kappa coefficient of SVM method was improved the most, which was 22.55%, and the improvement of the other two methods was also more than 20%. Finally, after denoising, binarization and morphological transformation of the apple tree pixel extraction result image, the fruit tree row feature points were extracted by the region of interest division method, and the fruit tree row navigation line was obtained by fitting each row feature points by the least square method. The average angular deviation of this method was 0.5975°, and the overall average time after ten tests was 0.4023s. The research result can provide a basis for the identification and extraction of fruit tree pixels and fruit tree row navigation line in complex environments.