Abstract:Ecological water demand (EWD) is the basis for regionally ecological water control and environment restoration. The Mara River Basin has a world famous ecosystem, and the vegetation EWD accounts for a large part of the total water demand in the Basin. Based on the ERA5 meteorological data, leaf index data (LAI) data from 1980 to 2020, and HSWD world soil data, we calculated the vegetation EWD temporal and spatial variation characteristics in the four seasons (short dry season, long rainy season, long dry season, and short rainy season) of Mara River Basin by using the Penman-Monteith method. Then, three different machine learning methods of support vector machine (SVM), random forest (RF), convolutional neural network (CNN) and seven environmental factors (temperature, precipitation, 10 m wind speed, LAI, solar radiation, relative humidity and terrain) were used to establish regression models. Thus, the vegetation EWD in different seasons from 2011 to 2020 was estimated by the machine learning regression models, and the estimation results were compared with the results calculated by the Penman-Monteith method in the fitting degree of time change series and the similarity in the space. The results showed that the EWD of vegetation in the Mara River Basin showed fluctuating trends in all seasons in the past 40 years. The vegetation EWD in different seasons from the most to the least was long rainy season, long dry season, short rainy season, and short dry season, the EWD in the long rainy season was about 1.5 times of that in the short dry season. The EWD in the upstream and downstream was larger than that in the midstream in all the seasons. LAI was the largest positive influence factor and wind speed was the largest negative influence factor. In terms of the accuracy of vegetation EWD estimation, RF performed the best, which mainly reflected in the minimum estimation error of the maximum, mean and minimum values, and the highest fitting degree of time change series. In the space, the best performance reflected in the most similar spatial distribution, and the smallest relative error. However, the estimation of SVM was relatively the worst. RF was the most suitable method for estimating vegetation EWD in the Mara River Basin. In this study, the three different machine learning methods were used to estimate the vegetation EWD in different seasons in the Mara River Basin, and the results can provide technical reference for the EWD estimation.