146 4.2. Model Architecture A Convolutional Variational Autoencoder (VAE-CNN) was used to learn latent representations from 9,600 subcrops (80 × 80 × 6) of the study area. The encoder includes three convolutional layers (32, 80, 128 filters) and a dense layer, producing a 3-dimensional latent space via parallel mean and log-variance layers. The decoder mirrors the encoder with transposed convolutions to reconstruct the input (see Table 1). Table 1 – Encoder Architecture Summary Layer Type Output Shape Filters / Units Parameters Input Input Layer 80 × 80 × 6 — 0 Conv1 Conv2D 40 × 40 × 32 32 1,760 Conv2 Conv2D 20 × 20 × 80 80 23,120 Conv3 Conv2D 10 × 10 × 128 128 92,288 Flatten Flatten 12,800 — 0 Dense Fully Connected 128 128 1,638,528 Latent Mean Dense (z_mean) 3 3 387 Latent Log-Var Dense (z_log_var) 3 3 387 Sampling Reparameterization 3 — 0 4.3. Loss Function The model is trained by minimizing a composite loss function consisting of: • Reconstruction loss, computed using binary cross-entropy between the input and reconstructed output • Kullback–Leibler (KL) divergence, which regularizes the latent space toward a standard normal distribution The total loss is defined as: :=: reconstruction +G:KL where β = regularization coefficient. 4.4. Training Procedure The model was trained for 25 epochs using a batch size of 512, the ADAM optimizer, a latent dimension of 3, and a β value of 1. This configuration was selected to ensure stable convergence while maintaining computational efficiency. 4.5. Model Evaluation Metrics During training, the total loss decreased from approximately 3959 in the first epoch to 2806 in the final epoch. The reconstruction loss showed a similar decreasing trend, indicating progressive improvement in the model’s ability to reconstruct the input samples. The KL divergence decreased sharply in the first epochs and then increased gradually, remaining at a low but non-zero value toward the end of training. The loss curves are presented in Figure 5.
RkJQdWJsaXNoZXIy MTM0Mzk2