Soil moisture prediction and forecast for planning against locust swarms
This is an app presenting a deep learning model developed by the Multiscale Hydrologic Processes and Intelligence (MHPI) group (led by Associate Professor Chaopeng Shen) at Penn State as a volunteer project for planning against the locust swarms which threaten food security for 13 million people. Soil moisture predictions can help determine where the locust swarms may lay their eggs, so that these areas can be targeted for control efforts. This project supports David Hughes, Professor of Entomology and Biology at Penn State in the effort to contain the locust. The MHPI group volunteered their time and effort to develop this model and app. Read an earlier report here. Read more about the eLocust3m app on PlantVillage website.
This app presents near-real-time prediction of soil moisture at 0.125-degree resolution using a time series deep learning model based on the long short-term memory (LSTM) architecture, trained on the Soil Moisture Active Passive (SMAP) satellite level-3 9km product. The adaptive data integration model, spiritually, wants to achieve what data assimilation does, and assimilates the most recently available SMAP data, filling in the gaps and making predictions based on our recently developed method (Fang et al., 2020).
Select a location on the map to see the time series for that location. For the small time series window, click and drag to zoom in on a time range, and right click to zoom back out. Select the box with arrow on the top right of the time series windo to view this plot larger in a new tab. The unit of SMAP is volumetric soil moisture [cm^3/cm^3]. Note that the predictions for today and the future are estimated based on forecasts from the Global Forecast System (GFS). The quality of soil mositure is influenced by the quality of the weather forecast. The further ahead you are from the most recent SMAP observation, the less accurate it will be.
Evaluation metrics
All metrics were evaluated between model forecast and SMAP L3 product during the test period, when GPM precipitation is available.
NSE (Nash-Sutcliffe model efficiency coefficient): Relationship between predicted (model) and measured (SMAP) data, similar concept to R^2 value (closer to one is better). NSE>0.5 is consider reasonable and >0.7 is consider good.
Bias: Mean difference between the model forecast and SMAP L3 product (closer to zero is better).
ubRMSE (unbiased root-mean-squared error): After removing long-term bias, how close the individual observations are to the line of best fit (closer to zero is better).
RMSE (root-mean-squared error): How close the individual observations are to the line of best fit (closer to zero is better).
Corr (correlation): Relationship between predicted (model) and measured (SMAP) data (closer to one is better)
Model details
The Soil Moisture Active Passive (SMAP) satellite senses the moisture in the top 3-5 cm of soil. Our model does not track SMAP exactly because SMAP data also has random measurement errors, as does any precipitation/forcing data. The accuracy level of SMAP varies due to many factors, including vegetation water content, snowpack, and mountainous terrains. The LSTM model can detect random errors in SMAP. Please check the related accuracy layers. When precipitation data was inaccurate, the LSTM model assimilates SMAP data and can ameliorate the error. The deep learning model seeks to find a balance between precipitation forcing and SMAP observations. However, there are some days in some locations where SMAP soil moisture observations are continuously low, while the model predicted an increase in soil moisture. In these cases, it is quite likely that the precipitation data is in error (which led to the predicted increase), while SMAP is correct. We are in the process of assessing such errors.
Inputs to the model include precipitation from Global Precipitation Mission (GPM, 1 day lag) and other weather variables from Global Land Data Assimilation System (GLDAS, 2 month lag) when available. Weather forecasts from the Global Forecast System (GFS) are used to fill the gaps in forcing data to the most recent day and in the future.
Overall model accuracy assessment for the study region
References
Fang, K. and CP. Shen, Near-real-time forecast of satellite-based soil moisture using long short-term memory with an adaptive data integration kernel, Journal of Hydrometeorology, JHM-D-19-0169.1, doi:10.1175/JHM-D-19-0169.1 (2020)
Fang, K*, M. Pan, and CP. Shen, The value of SMAP for long-term soil moisture estimation with the help of deep learning, Transactions on Geoscience and Remote Sensing, 57(4), 2221-2233, doi: 10.1109/TGRS.2018.2872131 (2018)
Shen, CP., A trans-disciplinary review of deep learning research and its relevance for water resources scientists, Water Resources Research. 54(11), 8558-8593, doi: 10.1029/2018WR022643 (2018)
Fang, K.*, CP. Shen, D. Kifer and X. Yang, Prolongation of SMAP to Spatio-temporally Seamless Coverage of Continental US Using a Deep Learning Neural Network, Geophysical Research Letters, doi: 10.1002/2017GL075619, preprint accessible at arXiv:1707.06611 (2017)
Fang, K.*, CP. Shen and D. Kifer, Evaluating aleatoric and epistemic uncertainties of time series deep learning models for soil moisture predictions, International Conference on Machine Learning (ICML) Workshop, Climate Change: What can AI do? Long Beach, CA, June 2019 (Spotlight talk, double-blind peer reviewed, non-archival).
Team
Contributors include Jiangtao Liu, Ashutosh Sharma, Wen-Ping Tsai, Kai Ma, Dapeng Feng, Kathryn Lawson and group lead Chaopeng Shen.
Webpage maintenance: Ashutosh Sharma
Links
More resources for hydrologic deep learning on Shen's homepage