Event Aware Flood Mapping for Agricultural Landscapes: A Robustness Oriented Comparison of Deep Learning and Machine Learning in the Arkansas 2025 Flood Event

Accurate flood mapping is essential for rapid impact assessment in agricultural landscapes, where inundation can interrupt production, damage rural infrastructure, and complicate recovery. Recent research has shown that deep learning can support flood delineation from remotely sensed data, but the literature still reports recurring weaknesses in robustness assessment, uncertainty interpretation, and transparent benchmarking against simpler machine learning baselines. In this study, we developed an event aware flood mapping framework for the Arkansas 2025 flood event by integrating terrain, antecedent water conditions, rainfall dynamics, and hydrologic proximity within a spatial learning workflow. We evaluated progressively richer predictor configurations using a U Net based convolutional neural network and assessed robustness through repeated seed experiments under fixed train, validation, and test splits. A Random Forest benchmark was introduced to provide a transparent classical machine learning reference. Mean probability, uncertainty, and agreement maps were also generated to characterize spatial confidence patterns. The results showed that adding rainfall substantially improved performance relative to the more complex advanced deep learning configuration, increasing mean best validation IoU by 7.3%, mean test IoU by 15.8%, and mean test F1 by 12.4%, while also reducing the variability of test IoU, test F1, and threshold selection. The rainfall based U Net achieved a best validation IoU of 0.278 ± 0.015, a test IoU of 0.264 ± 0.014, and a test F1 of 0.417 ± 0.018. The more complex configuration that added distance to river and Focal Tversky optimization did not deliver a consistent gain, with a best validation IoU of 0.259 ± 0.011, a test IoU of 0.228 ± 0.033, and a test F1 of 0.371 ± 0.044. The Random Forest benchmark was unexpectedly strong, reaching a validation IoU of 0.608 at the selected threshold and a test IoU of 0.351 with a test F1 of 0.519, equivalent to gains of 33.0% in test IoU and 24.5% in test F1 relative to the best deep learning configuration. Feature importance analysis identified elevation, slope, rainfall peak, and event total rainfall as dominant contributors. Spatial uncertainty was concentrated along transitional flood margins, while agreement maps highlighted coherent high confidence cores. Cropland overlay analysis further demonstrated that the framework can support first pass agricultural exposure assessment. Rather than promoting model complexity for its own sake, this study shows that event aware predictors, robustness oriented evaluation, and explicit benchmark comparison are central to credible geospatial flood intelligence in agricultural environments.