Pooling operations are central to convolutional neural network design, yet conventional strategies, notably max pooling and average pooling, frequently discard sub-pixel textures and spectral cues critical for distinguishing visually similar plant diseases. This study presents a methodologically controlled comparison of five pooling mechanisms: Max Pooling, Average Pooling, Fast Fourier Transform Pooling, Discrete Wavelet Transform Pooling, and a proposed FFT–DWT Hybrid Pooling implemented within a single modified VGG-16 backbone (four convolutional layers; 256-channel feature depth) and evaluated on the PlantVillage tomato leaf dataset under identical experimental conditions, isolating each mechanism's contribution to classification performance and feature preservation. Within this comparison, the proposed FFT–DWT Hybrid Pooling achieved the highest classification accuracy (99.77%), outperforming all four alternative strategies while also better preserving global spectral features and localized multi-resolution textures. The findings indicate that hybrid pooling can simultaneously improve predictive performance and computational efficiency in VGG-16-based CNNs, and provide practical guidance for selecting pooling mechanisms when designing resource-constrained agricultural disease diagnosis systems.