Daily multistep soil moisture forecasting by combining linear and nonlinear causality and attention-based encoder-decoder model

End-to-end (E2E) models, including the attention-based encoder-decoder (AED) models, have achieved promising performance on the automatic speech recognition (ASR) task. However, the supervised training process of the E2E model needs a large amount of speech-text paired data. In contrast, self-supervised pre-training can pre-train the model on the unlabeled data and then fine-tune it on the limited labeled data to realize better performance. Most of the previous self-supervised pre-training methods focus on learning hidden representations from speech but ignore how to utilize the unpaired text. As a result, previous works often pre-train an acoustic encoder and then fine-tune it as a classification based ASR model, such as Connectionist Temporal Classification (CTC) based model, rather than an AED model. In this paper, we propose a self-supervised pre-training method for the AED model (SP-AED). The SP-AED method contains acoustic pre-training for the encoder, linguistic pre-training for the decoder, and an adaptive combination fine-tuning for the whole system. We first design a linguistic pre-training method for decoder by utilizing the text-only data. The decoder will be pre-trained as a noise-condition language model to learn the prior distribution of the text. Then, we pre-train the AED encoder with the wav2vec2.0 method with some modifications. Finally, we combine the pre-trained encoder and decoder and fine-tune them on the limited labeled data. We design an adaptive combination method during fine-tuning by modifying the decoder's input and output to prevent catastrophic forgetting. Experiments prove that compared with the random initialized models, the SP-AED pre-trained models can realize up to 17% relative improvement. And with similar model size or computational cost, we can get comparable results to other classification-based models on both English and Chinese corpus.

Recently, attention based sequence-to-sequence model become a research hotspot in speech recognition. The attention model has the problem of slow convergence and poor robustness. In this paper, a model that jointed a bottleneck feature extraction network and attention model is proposed. The model is composed of a Deep Belief Network as bottleneck feature extraction network and an attention-based encoder-decoder model. DBN can store the priori information from Hidden Markov Model so that increasing convergence speed of and enhancing both robustness and discrimination of features. Attention model utilizes the temporal information of feature sequence to calculate the posterior probability of phoneme. Then the number of stack recurrent neural network layers in attention model is reduced in order to decrease the calculation of gradient. Experiments in the TIMIT corpus showed that the phoneme error rate is 17.80% in test set, the average training iteration decreased 52%, and the number of training iterations decreased from 139 to 89. The word error rate of WSJ eval92 is 12.9% without any external language model.

Abstract. In this second part of the two-part paper, the data driven modeling (DDM) experiment, presented and explained in the first part, is implemented. Inputs for the five case studies (half-hourly actual evapotranspiration, daily peat soil moisture, daily till soil moisture, and two daily rainfall-runoff datasets) are identified, either based on previous studies or using the mutual information content. Twelve groups (realizations) were randomly generated from each dataset by randomly sampling without replacement from the original dataset. Neural networks (ANNs), genetic programming (GP), evolutionary polynomial regression (EPR), Support vector machines (SVM), M5 model trees (M5), K-nearest neighbors (K-nn), and multiple linear regression (MLR) techniques are implemented and applied to each of the 12 realizations of each case study. The predictive accuracy and uncertainties of the various techniques are assessed using multiple average overall error measures, scatter plots, frequency distribution of model residuals, and the deterioration rate of prediction performance during the testing phase. Gamma test is used as a guide to assist in selecting the appropriate modeling technique. Unlike two nonlinear soil moisture case studies, the results of the experiment conducted in this research study show that ANNs were a sub-optimal choice for the actual evapotranspiration and the two rainfall-runoff case studies. GP is the most successful technique due to its ability to adapt the model complexity to the modeled data. EPR performance could be close to GP with datasets that are more linear than nonlinear. SVM is sensitive to the kernel choice and if appropriately selected, the performance of SVM can improve. M5 performs very well with linear and semi linear data, which cover wide range of hydrological situations. In highly nonlinear case studies, ANNs, K-nn, and GP could be more successful than other modeling techniques. K-nn is also successful in linear situations, and it should not be ignored as a potential modeling technique for hydrological applications.

<strong> Description:</strong> A <strong>seamless global daily</strong> (<strong>SGD</strong>) AMSR2 soil moisture long-term (2013-2019) dataset is generated through the proposed model. This daily products include <strong>2553</strong> global soil moisture NetCDF4 files, starting from Jan 01, 2013 to Dec 31, 2019 (about <strong>20GB</strong> memory after uncompressing this zip file). To further validate the effectiveness of these productions, three verification ways are employed as follow: 1) In-situ validation; 2) Time-series validation; And 3) simulated missing regions validation. More validation results can be viewed at <strong>SGD-SM</strong>. An example Python code of extracting this dataset is also available at <strong>https://github.com/qzhang95/SGD-SM</strong>. Official LPRM AMSR2 Descending L3 soil moisture products indeed only have 28 daily files in May 2013 (missing data files in date May 11, May 12, and May 13). This soil moisture dataset is comprised of netCDF4 (*.nc) files. Therefore, users need to install <strong>netCDF4</strong> toolkit before reading the data: <pre><code class="language-python">pip install netCDF4 pip install numpy</code></pre> It should be noted that the original and reconstructed soil moisture data are both recorded in one NC file. User can read the original data, reconstructed data, and mask data as follows: <pre><code class="language-python">Data = nc.Dataset(NC_file_position) Ori_data = Data.variables['original_sm_c1'] Rec_data = Data.variables['reconstructed_sm_c1'] Ori = Ori_data[0:720, 0:1440] Rec = Rec_data[0:720, 0:1440] Mask_ori = np.ma.getmask(Ori)</code></pre>

Altered-LiNGAM (ALiNGAM) for solving nonlinear causal models when data is nonlinear and noisy

Code generation from a graphical user interface via attention-based encoder–decoder model

The Variable Infiltration Capacity (VIC) land surface macroscale hydrology model was used to reconstruct 60 years (1950-2009) of daily soil moisture values for three soil layers (0–20 cm, 20–100 cm, and 0–100 cm) over the three Canadian Prairie Provinces with a total area of 1,964,000 km2. VIC was applied over a grid of 4,393 points with a resolution of 0.25° × 0.25°, and was driven by observed daily maximum and minimum air temperature and precipitation from 1,167 meteorological stations. The model was first calibrated using observed hydrographs from seven catchments with drainage areas varying from 3,750 to 7,870 km2. Special attention was given to modelling of rainfall-runoff processes over the non-contributing drainage area of the Prairies. VIC was then validated over these seven catchments at different periods and over an additional five catchments with drainage areas ranging from 36,500 to 131,000 km2. An estimation procedure to determine model parameters was developed and applied to catchments where hydrographs are not available for the standard calibration process. In situ soil moisture measurements from six Alberta sites were also used for model validation. VIC performed well over both calibration and validation catchments. The results clearly demonstrate that incorporating non-contributing drainage areas into runoff calculations could substantially improve the ability of VIC to simulate surface and sub-surface runoff in regions where poor drainage network development is a dominant feature of drainage basins. The VIC reconstructed 60–year average of the soil moisture in the top 1 m shows some expected climatological features of the Prairies. For example, the reconstructed soil moisture climatology portrays the dry Palliser Triangle region and the Prairie Dry Belt in the southern Prairies. The VIC simulated soil moisture was used to calculate the daily Soil Moisture Anomaly Percentage Index (SMAPI) for the three soil layers; SMAPI can be used as an index of agricultural drought severity taking into account climatology. The value of the calculated SMAPI in quantifying and documenting prairie drought events is demonstrated through an intensive examination of the April 2002 drought case.

An in-situ data based model to downscale radiometric satellite soil moisture products in the Upper Hunter Region of NSW, Australia

In large river basins where in situ data were limited or absent, satellite-based soil moisture estimates can be used to supplement ground measurements for land and water resource management solutions. Consistent soil moisture estimation can aid in monitoring droughts, forecasting floods, monitoring crop productivity, and assisting weather forecasting. Satellite-based soil moisture estimates are readily available at the global scale but are provided at spatial scales that are relatively coarse for many hydrological modeling and decision-making purposes. Soil moisture data are obtained from NASA’s soil moisture active passive (SMAP) mission radiometer as an interpolated product at 9 km gridded resolution. This study implements a soil moisture downscaling algorithm that was developed based on the relationship between daily temperature change and average soil moisture under varying vegetation conditions. It applies a look-up table using global land data assimilation system (GLDAS) soil moisture and surface temperature data, and advanced very high resolution radiometer (AVHRR) and moderate resolution imaging spectroradiometer (MODIS) normalized difference vegetation index (NDVI) and land surface temperature (LST). MODIS LST and NDVI are used to obtain downscaled soil moisture estimates. These estimates are then used to enhance the spatial resolution of soil moisture estimates from SMAP 9 km to 1 km. Soil moisture estimates at 1 km resolution are able to provide detailed information on the spatial distribution and pattern over the regions being analyzed. Higher resolution soil moisture data are needed for practical applications and modelling in large watersheds with limited in situ data, like in the Lower Mekong River Basin (LMB) in Southeast Asia. The 1 km soil moisture estimates can be applied directly to improve flood prediction and assessment as well as drought monitoring and agricultural productivity predictions for large river basins.

Accurate estimates of daily rainfall are essential for understanding and modeling the physical processes involved in the interaction between the land surface and the atmosphere. In this study, daily satellite soil moisture observations from the Advanced Microwave Scanning Radiometer–Earth Observing System (AMSR–E) generated by implementing the standard National Aeronautics and Space Administration (NASA) algorithm are employed for estimating rainfall, firstly, through the use of recently developed approach, SM2RAIN and, secondly, the nonlinear autoregressive network with exogenous inputs (NARX) neural modelling at five climate stations in the Karkheh river basin (KRB), located in south-west Iran. In the SM2RAIN method, the period 1 January 2003 to 31 December 2005 is used for the calibration of algorithm and the remaining 9 months from 1 January 2006 to 30 September 2006 is used for the validation of the rainfall estimates. In the NARX model, the full study period is split into training (1 January 2003 to 31 September 2005) and testing (1 September 2005 to 30 September 2006) stages. For the prediction of the rainfall as the desired target (output), relative soil moisture changes from AMSR–E and measured air temperature time series are chosen as exogenous (external) inputs in NARX. The quality of the estimated rainfall data is evaluated by comparing it with observed rainfall data at the five rain gauges in terms of the coefficient of determination R2, the RMSE and the statistical bias. For the SM2RAIN method, R2 ranges between 0.32 and 0.79 for all stations, whereas for the NARX- model the values are generally slightly lower. Moreover, the values of the bias for each station indicate that although SM2RAIN is likely to underestimate large rainfall intensities, due to the known effect of soil moisture saturation, its biases are somewhat lower than those of NARX. Moreover, Precipitation Estimation from Remotely Sensed Information using Artificial Neural Networks–Climate Data Record (PERSIANN–CDR) is employed to evaluate its potential for predicting the ground-based observed station rainfall, but it is found to work poorly. In conclusion, the results of the present study show that with the use of AMSR–E soil moisture products in the physically based SM2RAIN algorithm as well as in the NARX neural network, rainfall for poorly gauged regions can be predicted satisfactorily.

Eddy covariance sites are ideally suited for the study of extreme events on ecosystems as they allow the exchange of trace gases and energy fluxes between ecosystems and the lower atmosphere to be directly measured on a continuous basis. However, standardized definitions of hydroclimatic extremes are needed to render studies of extreme events comparable across sites. This requires longer datasets than are available from on-site measurements in order to capture the full range of climatic variability. We present a dataset of drought indices based on precipitation (Standardized Precipitation Index, SPI), atmospheric water balance (Standardized Precipitation Evapotranspiration Index, SPEI), and soil moisture (Standardized Soil Moisture Index, SSMI) for 101 ecosystem sites from the Integrated Carbon Observation System (ICOS) with daily temporal resolution from 1950 to 2021. Additionally, we provide simulated soil moisture and evapotranspiration for each site from the Mesoscale Hydrological Model (mHM). These could be utilised for gap-filling or long-term research, among other applications. We validate our data set with measurements from ICOS and discuss potential research avenues.

&amp;lt;p&amp;gt;Soil moisture is an important part of the vegetation cycle and a controlling factor for agriculture. Withstanding the role of agricultural productivity in economic development of a nation, it is imperative that water resources planners and managers are able to assess and forecast agricultural drought. As agricultural drought is related to declining soil moisture, this paper studies the dynamics of soil moisture based drought in the transboundary Koshi river basin in the Himalayan region. By applying the J2000 hydrological model, the daily soil moisture is derived for the whole basin for a 28-year time frame (1980-2007). The soil moisture deficit index (SMDI) is calculated based on a fully distributed spatial representation by considering the derivation from the long term soil moisture on a weekly time scale. In order to analyze the variation of soil moisture drought spatially, the river basin is subdivided into three distinct geographical areas, i.e. Northern Tibet, High and Middle Mountains, and Southern Plain. Further, temporally the SMDI is calculated for four distinct seasons based on wetness and dryness patterns observed in the study area, i.e. monsoon, post-monsoon, winter and pre-monsoon. A multi-site and multi-variable (streamflow at one station and evapotranspiration at three stations) approach was used for the calibration and validation of the J2000 model. Results show that the J2000 model is able to simulate the hydrological cycle of the basin with high accuracy. The model properly represents the winter drought of 2005 and 2006 was the most severe drought in the 28-year time period. Results also show considerable increases in the frequency of pre-monsoon and post-monsoon soil moisture drought in recent years. Severe droughts have had a high frequency in recent years, which is also reflected by an increase of areas that were impacted. In summary, our results show that severity and occurrence of agricultural drought has increased in the Koshi river basin in the last three decades, especially in the winter and pre-monsoon. This will have serious implications for agricultural productivity and for water resources management of the basin.&amp;lt;/p&amp;gt;

Multi-decadal soil moisture trends and their relationships to precipitation (P) and evapotranspiration (E) were explored in Mongolia. We modified a simple soil-moisture model intended for use in a wide range of practical applications for monitoring of pasture drought across the country, by incorporating soil freezing and snow melt. Daily soil moisture (Wm) was simulated using the model during 1961–2006 at nine stations distributed in the three different vegetation zones: forest steppe (FS), steppe (ST), and desert steppe (DS). The model performed reasonably well in simulating seasonal and interannual variations in observed soil moisture (Wo). The Wm was controlled by the subtle balance between P and evapotranspiration (E), which had in turn impacted on the vegetation during the growing season (May–September). All vegetation zones showed a decreasing trend in Wm during 1961–2006 due to decreased P and increased potential evapotranspiration, although the drying trend was only significant (p <0.05) in the FS. Composite analyses revealed that Wm anomalies were most manifested in summer due to large P and E anomalies, which were maintained throughout the freezing of winter into spring. In addition, autocorrelation analysis of Wm for the FS zone showed that during the autumn and winter, the temporal scale was 6.0–7.0 months larger than during spring and summer (1.8–3.0 months). These findings indicate that Wm acts as an efficient memory of P anomalies via freezing of the soil and as an initial moisture condition for vegetation activity in the subsequent year.

We use the Variable Infiltration Capacity (VIC) land surface macroscale hydrology model driven by observed maximum and minimum air temperatures and precipitation to map daily soil moisture values over China for the period 1 January 1971 to 31 July 2005. The model is applied over a grid of 10 458 points with a resolution of 30 km × 30 km. The model is first calibrated using observed hydrographs from 35 catchments with drainage areas varying from 190 to 351 530 km2. The model is then validated over these 35 catchments at different periods, and over an additional eight catchments with drainage areas ranging from 1230 to 10 010 km2. An estimation procedure to determine model parameters is developed and applied to catchments where hydrographs are not available for the standard calibration process. In situ soil moisture measurements from 28 sites around the country are also used for model validation. VIC performs well over both calibration and validation catchments especially in humid and semi‐humid regions. The 35‐year soil moisture climatology for the top 1 m from VIC is consistent with known soil moisture conditions in China.

Since drought and excessive rainfall can have significant socioeconomic impacts, it is important to have accurate high-resolution gridded datasets that can help improve analysis and forecasting of these conditions. One such widely used dataset is the Parameter-Elevation Regressions on Independent Slopes Model (PRISM). PRISM uses a digital elevation model (DEM) to obtain gridded elevation analyses and then uses a regression analysis along with approximately 15 000 surface precipitation measurements to produce a 4-km resolution daily precipitation product over the conterminous United States. The U.S. Climate Reference Network (USCRN) consists of 114 stations that take highly accurate meteorological measurements across all regions of the United States. A comparison between the USCRN and PRISM was performed using data from 2006 to 2018. There were good comparisons between the two datasets across nearly all seasons and regions; most mean daily differences were &lt;1 mm, with most absolute daily differences ~5 mm. The most general characteristics were for a net dry bias in the PRISM data in the Southwest and a net moist bias in the southern United States. Verifying the PRISM dataset provides us with confidence it can be used with estimates of evapotranspiration, high-resolution gridded soil properties, and vegetation datasets to produce a daily gridded soil moisture product for operational use in the analyses and prediction of drought and excessive soil moisture conditions.