Data Filtering Method based on LSTM for Non-Water-Stressed Baseline Estimation in Real-time Crop Water Stress Monitoring

To determine crop water status and quantify crop’s stress, an indicator called crop water stress index (CWSI) using empirical approach was used. The experiment has been conducted with the purpose of estimating CWSI of a durian tree around durian orchard which is located at Chanthaburi, Thailand. Since real-time recorded dataset was affected by ever-changing environmental conditions, data points became scattered. As a result, by using the correlation between stem water potential and CWSI of the durian tree to evaluate accuracy of calculated CWSI, the correlation coefficient was averagely equal to -0.042. Therefore, this paper proposed a data cleansing framework that included time, irradiance, wind speed, air relative humidity as parameters and implemented long short-term memory (LSTM) to filter the dataset for estimating non-water-stressed baseline. The results show that this framework is able to produce correlation coefficient between stem water potential and CWSI to be averagely equal to -0.650.