Deep Learning Performance Comparison Using Multispectral Images and Vegetation Index for Farmland Classification

This study aims to develop an efficient farmland management system through large-area farmland image mapping and deep learning farmland classification. The first step was to photograph the kimchi cabbage farmland using a drone equipped with a multispectral camera, resulting in 14,668 images in an area of about 1.6 km². To preprocess the image data efficiently, an algorithm was used to remove unnecessary images based on each image's GPS location and altitude, reducing the total number of images to 8930. This preprocessing step improved the image mapping speed by about 8.3 times compared to the original data image mapping speed. To achieve efficient large-scale farmland classification, the input dataset was constructed based on multispectral images, and deep learning results were compared. A total of eight input data sets were constructed using five wavelength bands and vegetation index data obtained through a multispectral camera, and farmland classification was performed using deep learning. The accuracy of farmland classification was analyzed using Mean IoU (intersection over union), and the case including red, green, blue, red edge, and near IR showed the highest accuracy value of 0.789.