Among all the natural disasters, drought has the most catastrophic encroachment on the surrounding and environment. Gulbarga, one of the semi-arid districts of Karnataka state, India receives about 700 mm of average annual rainfall and is drought inclined. In this study, the forecasting of drought for the district has been carried out for a lead time of 1 month and 6 months. The multi-temporal Standardized Precipitation Index (SPI) has been used as the drought quantifying parameter due to the fact that it is calculated on the basis of one simplest parameter, i.e., rainfall and additionally due to its ease of use. The fine resolution daily gridded precipitation data (0.25º × 0.25º) procured from Indian Meteorological Department (IMD) of 21 grid locations within the study area have been used for the analysis. Forecasting of drought plays a significant role in drought preparedness and mitigation plans. With the advent of machine learning (ML) techniques over the past few decades, forecasting of any hydrologic event has become easier and more accurate. However, the use of these techniques for drought forecasting is still obscure. In this study, Artificial Neural Network (ANN) and Support Vector Regression (SVR) techniques have been employed to examine their accuracy in drought forecasting over shorter and longer lead times. Furthermore, two hybrid approaches have been formulated by coupling a data transformation method with each of the aforementioned ML approaches. At the outset, pre-processing of input data (i.e., SPI) has been carried out using Wavelet Packet Transform (WPT) and then used as inputs to ANN and SVR models to induce hybrid WP-ANN and WP-SVR models. The performance of the hybrid models has been evaluated based on the statistical indices such as R² (co-efficient of determination), RMSE (Root Mean Square Error), and MAE (Mean Absolute Error). The results showed that the hybrid techniques have better forecast performance than the standalone machine learning approaches. Hybrid WP-ANN model performed relatively better than WP-SVR model for most of the grid locations. Also, the forecasting results deteriorated as the lead time increased from 1 to 6 months.

中文翻译：

---

干旱小波的混合小波包机器学习方法

在所有自然灾害中，干旱对周围环境和环境造成的破坏性最大。印度卡尔纳塔克邦半干旱地区之一的古尔巴加（Gulbarga）年平均降雨量约为700毫米，并且倾向于干旱。在这项研究中，该地区的干旱预报已经提前了1个月和6个月。由于多时间标准化降水指数（SPI）是根据一个最简单的参数（即降雨量）计算得出的，并且由于其易用性而被用作干旱量化参数。从印度气象部门（IMD）获得的研究区域内21个网格位置的高分辨率日网格降水数据（0.25º×0.25º）已用于分析。干旱预报在干旱准备和减灾计划中发挥着重要作用。在过去的几十年中，随着机器学习（ML）技术的出现，对任何水文事件的预测变得更加容易和准确。但是，将这些技术用于干旱预报仍然不清楚。在这项研究中，人工神经网络（ANN）和支持向量回归（SVR）技术已被用来检验它们在较短和较长交货时间内的干旱预报准确性。此外，通过将数据转换方法与上述每种ML方法耦合在一起，已制定了两种混合方法。首先，对输入数据进行预处理（即，SPI）已使用小波包变换（WPT）进行，然后用作ANN和SVR模型的输入，以引入WP-ANN和WP-SVR混合模型。混合模型的性能已根据统计指标进行了评估，例如R ²（确定系数），RMSE（均方根误差）和MAE（平均绝对误差）。结果表明，混合技术比独立的机器学习方法具有更好的预测性能。对于大多数网格位置，混合WP-ANN模型的性能相对优于WP-SVR模型。同样，随着交付周期从1个月增加到6个月，预测结果也会恶化。