The paper introduces a flood detection system that is complete and makes use of the combination of wavelet transformation and deep learning, looking at different aspects of it, especially self-window modulation in this case, in the Cauvery Basin of Southern India. The research examines 4,971 daily samples from 2000 to 2013, consisting of 500 flood and 495 drought incidents. We put forward two models, wavelet–long short–term and wavelet–gated recurrent unit of with self-window modulation, which perform better than traditional methods by a significant margin. The wavelet-long-short term model that we have developed reaches 89.44% accuracy, 82.29% precision, 78.71% recall, and 84.67% F1-score, thereby also surpassing conventional long-short-term model by 7.1% accuracy. The method entails a combination of the discrete wavelet transform using the Daubechies-4 wavelet (energy: 1.37e+08, entropy: 21.43), feature extraction that points Flow_Mean14d (importance: 0.92) as the most informative and worthy feature to be selected, and selfwindow modulation technique asserting 48.2% input from window-1 in temporal study. A comparative study reveals that the integration of wavelets has improved the performance of the model by 6-9% on self-window modulation serent metrics, while self-window modulation contribution is an extra 3-5% boost. The system is capable of detecting floods in real–time with wavelet–long short–term and wavelet–gated recurrent units attaining AUC scores of 1.000 and 0.998, respectively, and at the same time providing large language model-generated safety advisories as support for operational decision-making.