Several studies related to soil type classification have been conducted. However, each of these studies uses different datasets. Only a small number of researchers share soil image datasets publicly. In addition, published datasets have an imbalance in the amount of data in each class which will result in poor model performance or over fit, especially deep learning. With data augmentation, new data variations can be formed that can handle the limited number of datasets. One of the modern augmentation models is DCGAN which is an extension of GAN. DCGAN is considered a good model in improving the stability of GAN training and the quality of image results. The resulting synthesized image is the result of mapping randomized latent vectors in an n-dimensional latent space. A meaningful image transformation is generated from the latent vector through arithmetic operations in the latent space dimension. The size of the latent space dimension is very important in enabling accurate reconstruction of the training data. To test the effect of latent space dimension on images, a quantitative evaluation using Fre'chet Inception Distance (FID) is used. The following results were obtained, for the best image quality in the alluvial soil category using latent space dimension 10 with FID score = 322.0. For clay soil category, the best image quality is generated using latent space dimension 100 with score FID = 332.84 and 512 with score FID = 322.08. In the black soil category, the best latent space dimension is 128 with FID score = 360.80. And for red soil, the best image quality is generated with the use of 512 latent spaces that have a FID score = 256.67.
Copyrights © 2023