Improved prediction of daily pan evaporation using Deep-LSTM model

Precise measurement or estimation of evaporation losses is extremely important for the development of water resource management strategies and its effective implementation, particularly in drought-prone areas for increasing agricultural productivity. Evaporation can either be measured directly using evaporimeters, or it can be estimated by means of empirical models with the help of climatic factors influencing evaporation process. In general, variations in climatic factors such as temperature, humidity, wind speed, sunshine and solar radiation influence and control the evaporation process to a great extent. Due to the highly nonlinear nature of evaporation phenomenon, it is invariably very difficult to model the evaporation process through climatic factors especially in diverse agro-climatic situations. The present investigation is carried out to examine the potential of deep neural network architecture with long short-term memory cell (Deep-LSTM) to estimate daily pan evaporation with minimum input features. Depending upon the availability of climatic data Deep-LSTM models with different input combinations are proposed to model daily evaporation losses in three agro-climatic zones of Chhattisgarh state in east-central India. The performance of the proposed Deep-LSTM models are compared with commonly used multilayer artificial neural network and empirical methods (Hargreaves and Blaney–Criddle). The results of the investigations in terms of various performance evaluation criteria reveal that the proposed Deep-LSTM structure is able to successfully model the daily evaporation losses with improved accuracy as compared to other models considered in this study.