<p>In this paper, fuzzy entropy is evaluated as a robust image feature under noisy conditions. The membership function (MF) parameters are varied in the context of linear widening and narrowing of the width of linguistic values (or fuzzy values) to maximize fuzzy entropy using a graphical solution instead of Evolutionary Computation (EC) or Particle Swarm Optimization (PSO), which are time-consuming and may suffer from convergence issues. The stability of fuzzy and Shannon entropy is analyzed under Additive White Gaussian Noise (AWGN), added to four standard grayscale images (Boat, Onion, Barbara, Baboon) across a wide SNR range (0 dB to –45 dB). Result shows that fuzzy entropy maintains its stability under noise down to –27 dB, significantly outperforming Shannon entropy, which begins to degrade beyond –18 dB. Furthermore, to evaluate classification performance under noise, four fuzzy features are extracted from each image and used to train three machine learning (ML) classifiers: Linear Discriminant Analysis (LDA), Naive Bayes (NB), and Decision Tree (DT), on a rice leaf disease dataset consisting of 300 images across three classes. Their performance is compared with two deep learning (DL) models, CNN and LSTM, trained directly on images, varying the SNR of images. While the DL models achieve higher peak accuracy, the ML classifiers using fuzzy entropy features demonstrate more stable performance under increasing noise. This study highlights the potential of fuzzy entropy as a low-complexity, noise-resilient alternative for image classifications, and identifies a trade-off between stability and accuracy in ML vs. DL approaches.</p>