In order to address the challenge of determining optimal resolutions for capturing images of different features using UAVs, the DJI M600Pro UAV was employed to acquire visible light images of cotton fields during the bud stage. By combining ground survey data and utilizing three supervised classification algorithms: artificial neural networks (ANN), support vector machines (SVM), and random forest (RF), field feature identification was conducted. The analysis encompassed varying resolutions (1.00cm, 2.50cm, 5.00cm, 7.50cm, 10.00cm) to evaluate the accuracy of feature recognition. Additionally, algorithm execution times were considered, with the aim of identifying the best resolution and optimal algorithm for cotton field feature recognition at the field scale in Southern Xinjiang, considering resolution, accuracy, and processing time. Experimental results indicated that at a spatial resolution of 1.00cm, SVM exhibited the highest accuracy in feature recognition, achieving an overall accuracy of 99.857% and a Kappa coefficient of 0.997. As spatial resolution was decreased, both overall accuracy and Kappa coefficient demonstrated a decreasing trend. At resolutions of 2.50cm and 5.00cm, when utilizing the RF algorithm, the shortest execution times were observed. Land, cotton, and drip irrigation lines displayed favorable recognition accuracy, with overall accuracy and Kappa coefficients surpassing 99.137% and 0.983, respectively. With resolutions exceeding 5.00cm, both overall accuracy and Kappa coefficient declined, notably impacting the mapping accuracy of drip irrigation lines (producer's accuracy, PA) and user accuracy (user's accuracy, UA). Images with resolutions lower than 5.00cm effectively identified characteristic features of bud-stage cotton fields, offering guidance for the identification of field feature types and their distribution patterns.