Pixel-based agricultural drought forecasting based on deep learning approach: Considering the linear trend and residual feature of vegetation temperature condition index

• ARIMA-LSTM model captured linear trend and residual feature of time series VTCI. • VTCI was a strong indicator of drought conditions and drought category. • The ARIMA model’s parameters were affected by topography and cropland type. • The one-step-ahead drought forecasting was more suitable for application. Drought is one of the most recurring natural hazards worldwide, causing serious damages on crop yield and economic development, so accurate agricultural drought forecasting in advance is essential for reducing disaster risk and enhancing agricultural management. Traditional statistical models, represented by autoregressive integrated moving average (ARIMA) model, demonstrate huge advantages in drought index based-forecasting, but overlook the complex underlying features inherent within the time series remotely sensed variables, such as residual feature. Long short-term memory (LSTM) is proven as an effective deep learning method for capturing nonlinear relationships, and is suitable for dealing with residual feature of time series drought index. With the study aim to explore the potential of a drought forecasting model constituted by a statistical model and a deep learning model, an advanced ARIMA-LSTM drought forecasting model was constructed by combining the linear expressiveness ability of the ARIMA model and the advantage in capturing the nonlinear relationship of the LSTM model. The proposed model considered both the linear trend and residual feature in the time series vegetation temperature condition index (VTCI) for improving the accuracy of drought forecasting in the Sichuan Basin, PR China. The results show that ARIMA-LSTM (RMSE = 0.06) achieved better forecast accuracy compared to the ARIMA model (RMSE = 0.10) without considering residual feature. In addition, the ARIMA-LSTM model more accurately expressed the severe drought condition in the east parallel ridge and valley and the wet condition in the central hilly area of the basin, presenting a more detailed spatial distribution that was closer to the drought monitoring results. From the quantitative analysis of forecasting errors, the absolute error range of the ARIMA-LSTM model was narrowed from [-0.6, 0.2] to [-0.2, 0.1] after the introduction of the residual feature. In summary, the ARIMA-LSTM drought forecasting model considering the linear trend and residual feature of time series VTCI accurately identified the drought condition in advance and provided a novel insight for drought forecasting at the pixel level to guarantee food security and limit the risk of drought.