Extreme precipitation events have recently increased due to global warming, leading to higher humidity and temperatures. Therefore, accurate, updated, and comprehensive precipitation estimations are essential to mitigate the severe losses caused by these extreme events. Satellite precipitation products provide high spatiotemporal resolution and global coverage, enabling improved detection and estimation of precipitation, especially in areas with limited ground observations. This research leverages some of the advanced satellite precipitation products and atmospheric dataset reanalysis available, enhancing the accuracy of precipitation estimates, particularly during heavy and extreme precipitation events. In this work, we use Convolutional Neural Networks (CNN) and optimization techniques to introduce a two-step methodology for improving satellite-based estimation and detection of heavy and extreme precipitation events. In the first step, feature centroids are integrated into a resampling method to increase the diversity of samples. The second step applies the Ant Colony Optimization (ACO) algorithm to refine categorical evaluation criteria, enhancing prediction accuracy. This study uses the ACO meta-heuristic algorithm to optimize a Multi-Layer Perceptron (MLP) model, using CNN-based precipitation estimates as input. The performance of the models and products in estimating and detecting heavy and extreme precipitation events is evaluated using various metrics, including Pearson correlation (r) and Kling-Gupta Efficiency (KGE). The results show that the Ant Colony-optimized MLP (ACO-MLP) outperforms the satellite precipitation products and CNN models in estimating the extreme and heavy precipitation with r = 0.89, 0.82, and KGE = 0.71, 0.65, over the continental area of Australia. The result of utilizing the proposed ACO-MLP and CNN models in Australia showed that the spatial pattern of the extreme precipitation event over the country's east coast on 23 February 2022 is effectively captured. This work contributes to improving water resources management and advancing climate studies, particularly in understanding and addressing extreme precipitation conditions.

中文翻译：

---

基于卫星的极端降水估计，使用卷积神经网络和蚁群优化的多层感知器


由于全球变暖，最近极端降水事件有所增加，导致湿度和温度升高。因此，准确、最新和全面的降水估计对于减轻这些极端事件造成的严重损失至关重要。卫星降水产品提供高时空分辨率和全球覆盖，能够改进降水的检测和估计，尤其是在地面观测有限的区域。这项研究利用了一些可用的高级卫星降水产品和大气数据集再分析，提高了降水估计的准确性，尤其是在强降水和极端降水事件期间。在这项工作中，我们使用卷积神经网络 （CNN） 和优化技术引入了一种两步方法，以改进基于卫星的强降水和极端降水事件的估计和检测。第一步，将要心集成到重采样方法中，以增加样本的多样性。第二步应用 Ant Colony Optimization （ACO） 算法来优化分类评估标准，从而提高预测准确性。本研究使用 ACO 元启发式算法来优化多层感知器 （MLP） 模型，使用基于 CNN 的降水估计作为输入。使用各种指标评估模型和产品在估计和检测强降水和极端降水事件方面的性能，包括皮尔逊相关 （r） 和 Kling-Gupta 效率 （KGE）。结果表明，蚁群优化 MLP （ACO-MLP） 在估计极端和强降水方面优于卫星降水产品和 CNN 模型，r = 0.89， 0.82， KGE = 0.71， 0。65 号，在澳大利亚大陆地区上空。在澳大利亚利用拟议的 ACO-MLP 和 CNN 模型的结果表明，有效捕获了 2022 年 2 月 23 日该国东海岸极端降水事件的空间模式。这项工作有助于改善水资源管理和推进气候研究，特别是在理解和解决极端降水条件方面。