Improving ANFIS Classification of Rice Growth Stages Using Airborne Remote Sensing: A Fuzzy C-Means Clustering Approach

Introduction: High-resolution crop phenology monitoring is vital for precision agricultural management, yet traditional Adaptive Neuro-Fuzzy Inference Systems (ANFIS) struggle with discrete rice growth stage classification due to an architectural mismatch. Because the native ANFIS framework optimizes continuous regression via Root Mean Squared Error (RMSE) minimization, standard initializations like rigid grid partitioning lead to severe boundary-rounding instabilities, rule explosion, and spatial overfitting when mapping complex vegetative growth phases.

Methods: This study develops a robust, data-driven agricultural mapping framework by integrating Fuzzy C-Means (FCM) clustering with the default hybrid ANFIS optimization engine (analytic Least-Squares Estimation and Gradient Descent backpropagation). Utilizing multi-band airborne raster imagery (Red, Green, and NIR bands), we systematically evaluated a 27-rule grid-partitioned baseline against an optimized 10-cluster FCM-ANFIS model to classify six distinct rice growth stages over an independent validation dataset.

Results: Validation tracking revealed a compelling divergence between continuous regression residuals and categorical classification accuracy. While the rigid grid model achieved a lower validation error (RMSE = 0.388), it failed to generalize across high-resolution canopy variations, yielding a low Overall Accuracy (OA) of 66.37% and a moderate Kappa coefficient of 0.5664 due to boundary over-parameterization. Conversely, the proposed FCM-ANFIS model successfully bypassed these limitations. By anchoring its 10 rules over natural spectral densities of the airborne crop canopy, it achieved a superior validation OA of 88.54% and a robust Kappa of 0.8417, despite a slightly higher regression residual (RMSE = 0.4083).

Conclusions: Slicing high-resolution airborne inputs into rigid geometric lattices degrades highly sensitive crop phenology boundaries. Aligning fuzzy rules with natural spatial clusters using FCM bridges the gap between fuzzy regression engines and discrete agricultural classification without requiring complex metaheuristic search alternatives.