Get reliable rangeland science

A robust and accurate image classification model is essential for the development of a smartphone application to help farmers identify forages from weeds. This study focused on developing and comparing three models: Keras-based deep learning, Support Vector Machine (SVM), and Back Propagation Neural Network (BPNN). A total of 1500 images of alfalfa (Medicago sativa), sericea lespedeza (Lespedeza cuneata), and weeds were used. The Keras model was tested with varying image sizes, batch sizes, and epochs. The highest performance was achieved with an image size of 128, a batch size of 8, and 100 epochs, yielding accuracy, precision, recall, and F1 score s of 99.01%, but with the highest training time of 44.25 seconds. Alternatively, using a smaller image size of 32 and a batch size of 32 with 50 epochs resulted in a lower accuracy of 98.38%, but significantly reduced training time to 9.61 seconds. The SVM model, with a Radial Basis Function (RBF) kernel, had excellent performance metrics, achieving an accuracy, precision, recall, and F1 score of 99.02%, with an exceptionally low training time of 0.059 seconds and a testing time of 0.01 seconds. This indicates the SVM's efficiency and suitability for rapid classification tasks. The BPNN model, tested with an image size of 128 and a neuron structure of over 200 iterations, achieved an accuracy of 98.36%, with a training time of 2.17 seconds and a minimal testing time of 0.0017 seconds, also showing efficient computational performance. The SVM model is recommended for the smartphone application due to its high accuracy, precision, recall, and F1 score, with its minimal computational requirements, making