Climate change is leading to rapid environmental changes, including fluctuating precipitation and water levels, which raises the risk of flooding in coastal and riverine locations around the United States. This study focuses on Sacramento, California, a city significantly affected by these changes and recent severe flooding disasters. The ultimate goal is to understand the climate dynamics and create a more robust model to alert Sacramento and other communities to possible flooding and better prepare them for future climatic uncertainty. In this research, four classification machine learning models—Support Vector Machine (SVM), Random Forest (RF), Artificial Neural Network (ANN), and Long Short-Term Memory (LSTM)—are examined for their capacity to predict the occurrence of floods using historical precipitation temperature and soil moisture data. Our results demonstrate that the LSTM model, with an accuracy of 89.99%, may provide better reliable flood predictions, possibly due to its ability to process complicated temporal data. SVM, RF, and ANN showed accuracies of 81.25%, 83.75%, and 85%, respectively. The study explores the correlation between increasing precipitation incidents and severe climate variations, such as the El Niño and La Niña cycles, which could have increased flooding risks. Significant rainfall peaks occurred in 1998 and 2007, indicating that external atmospheric circumstances might have considerably impacted local weather patterns. While LSTM models show potential, there remains room to improve their accuracy and adaptability in extreme flood scenarios. Given these findings, future research could combine multiple environmental data sources and hybrid modeling approaches to enhance predictions.
Nonstationarity due to climate variation and anthropogenic disturbances has altered high flow regimes. However, the extent of change has not been evaluated for undisturbed versus disturbed watersheds. This article aimed to determine how partitioning watersheds into undisturbed and disturbed categories can improve the performance of probability distributions for flood analysis throughout the United States. We utilized peak flow information for 26 reference (undisturbed) and 78 nonreference (disturbed) watersheds with drainage areas ranging from 135 to 42,367 km2 and record lengths of 100 to 140 years. Results indicated that flood quantile estimates of the Log Pearson Type III (LP3) distribution were likely being overestimated for return periods of 2 to 10 years, while flood estimates of 50 years and higher might be underestimated. In contrast, the Generalized Extreme Value (GEV) distribution outperformed LP3 in estimating floods with return periods of 50 years or more. These findings enhance flood frequency analysis and forecasting under nonstationary conditions.
