Evaluation of Three Feature Dimension Reduction Techniques for Machine Learning-Based Crop Yield Prediction Models

Machine learning (ML) has been widely used worldwide to develop crop yield forecasting models. However, it is still challenging to identify the most critical features from a dataset. Although either feature selection (FS) or feature extraction (FX) techniques have been employed, no research compares their performances and, more importantly, the benefits of combining both methods. Therefore, this paper proposes a framework that uses non-feature reduction (All-F) as a baseline to investigate the performance of FS, FX, and a combination of both (FSX). The case study employs the vegetation condition index (VCI)/temperature condition index (TCI) to develop 21 rice yield forecasting models for eight sub-regions in Vietnam based on ML methods, namely linear, support vector machine (SVM), decision tree (Tree), artificial neural network (ANN), and Ensemble. The results reveal that FSX takes full advantage of the FS and FX, leading FSX-based models to perform the best in 18 out of 21 models, while 2 (1) for FS-based (FX-based) models. These FXS-, FS-, and FX-based models improve All-F-based models at an average level of 21% and up to 60% in terms of RMSE. Furthermore, 21 of the best models are developed based on Ensemble (13 models), Tree (6 models), linear (1 model), and ANN (1 model). These findings highlight the significant role of FS, FX, and specially FSX coupled with a wide range of ML algorithms (especially Ensemble) for enhancing the accuracy of predicting crop yield.

Enhancing Crop Yield Prediction Utilizing Machine Learning on Satellite-Based Vegetation Health Indices

Accurate crop yield forecasting is essential in the food industry's decision-making process, where vegetation condition index (VCI) and thermal condition index (TCI) coupled with machine learning (ML) algorithms play crucial roles. The drawback, however, is that a one-fits-all prediction model is often employed over an entire region without considering subregional VCI and TCI's spatial variability resulting from environmental and climatic factors. Furthermore, when using nonlinear ML, redundant VCI/TCI data present additional challenges that adversely affect the models' output. This study proposes a framework that (i) employs higher-order spatial independent component analysis (sICA), and (ii), exploits a combination of the principal component analysis (PCA) and ML (i.e., PCA-ML combination) to deal with the two challenges in order to enhance crop yield prediction accuracy. The proposed framework consolidates common VCI/TCI spatial variability into their respective subregions, using Vietnam as an example. Compared to the one-fits-all approach, subregional rice yield forecasting models over Vietnam improved by an average level of 20% up to 60%. PCA-ML combination outperformed ML-only by an average of 18.5% up to 45%. The framework generates rice yield predictions 1 to 2 months ahead of the harvest with an average of 5% error, displaying its reliability.

The 2019-2020 Rise in Lake Victoria Monitored from Space: Exploiting the State-of-the-Art GRACE-FO and the Newly Released ERA-5 Reanalysis Products

During the period 2019-2020, Lake Victoria water levels rose at an alarming rate that has caused various problems in the region. The influence of this phenomena on surface and subsurface water resources has not yet been investigated, largely due to lack of enough in situ measurements compounded by the spatial coverage of the lake's basin, incomplete/inconsistent hydrometeorological data, and unavailable governmental data. Within the framework of joint data assimilation into a land surface model from multi-mission satellite remote sensing, this study employs the state-of-art Gravity Recovery and Climate Experiment follow-on (GRACE-FO) time-variable terrestrial water storage (TWS), newly released ERA-5 reanalysis, and satellite radar altimetry products to understand the cause of the rise of Lake Victoria on the one hand, and the associated impacts of the rise on the total water storage compartments (surface and groundwater) triggered by the extreme climatic event on the other hand. In addition, the study investigates the impacts of large-scale ocean-atmosphere indices on the water storage changes. The results indicate a considerable increase in water storage over the past two years, with multiple subsequent positive trends mainly induced by the Indian Ocean Dipole (IOD). Significant storage increase is also quantified in various water components such as surface water and water discharge, where the results show the lake's water level rose by ∼1.4 m, leading to approximately 1750 gigatonne volume increase. Multiple positive trends are observed in the past two years in the lake's water storage increase with two major events in April-May 2019 and December 2019-January 2020, with the rainfall occurring during the short rainy season of September to November (SON) having had a dominant effect on the lake's rise.

Altimetry-derived surface water data assimilation over the Nile Basin

Global hydrological models facilitate studying of water resources and their variations over time. The accuracies of these models are enhanced when combined with ever-increasing satellite remotely sensed data. Traditionally, these combinations are done via data assimilation approach, which permits the use of improved hydrological outputs to study regions with limited in-situ measurements such as the Nile Basin. This study aims at using the state-of-art satellite radar altimetry data to enhance a land-based hydrological model for studying water storage changes over the Nile Basin. Altimetry-derived surface water storage, for the first time, is assimilated into the model using the ensemble Kalman filter (EnKF) for the period of 2003 to 2016. Multiple datasets from ground measurements, as well as space observations, are used to evaluate the performance of the assimilated satellite altimetry data. Results indicate that the assimilation successfully improves model outputs, especially the surface water component. The process increases the correlation between surface water storage changes and water level variations from satellite radar altimetry by 0.44 and reduces the surface water discharge root-mean-square errors (RMSE) by approximately 33%.

Improved remotely sensed satellite products for studying Lake Victoria's water storage changes

Lake Victoria (LV), the world's second largest freshwater lake, supports a livelihood of more than 42 million people and modulates the regional climate. Studying its changes resulting from impacts of climate variation/change and anthropogenic is, therefore, vital for its sustainable use. Owing to its shear size, however, it is a daunting task to undertake such study relying solely on in-situ measurements, which are sparse, either missing, inconsistent or restricted by governmental red tapes. Remotely sensed products provide a valuable alternative but come with a penalty of being mostly incoherent with each other as they originate from different sources, have different underlying assumptions and models. This study pioneers a procedure that uses a Simple Weighting approach to merge LV's multi-mission satellite precipitation and evaporation data from various sources and then improves them through a Postprocessing Filtering (PF) scheme to provide coherent datasets of precipitation (p), evaporation (e), water storage changes (Δs), and discharge (q) that accounts for its water budget closure. Principal component analysis (PCA) is then applied to the merged-improved products to analyze LV's spatio-temporal changes resulting from impacts of climate variation/change. Compared to the original unmerged data (0.62 and 0.37 average correlation for two samples), the merged-improved products are largely in agreement (0.91 average correlation). Furthermore, smaller imbalances between the merged-improved products are obtained with precipitation (37%) and water storage changes (35%) being the largest contributors to LV's water budget. This data improvement scheme could be applicable to any inland lake of a size similar to LV.

Understanding the global hydrological droughts of 2003-2016 and their relationships with teleconnections

Droughts often evolve gradually and cover large areas, and therefore, affect many people and activities. This motivates developing techniques to integrate different satellite observations, to cover large areas, and understand spatial and temporal variability of droughts. In this study, we apply probabilistic techniques to generate satellite derived meteorological, hydrological, and hydro-meteorological drought indices for the world's 156 major river basins covering 2003-2016. The data includes Terrestrial Water Storage (TWS) estimates from the Gravity Recovery And Climate Experiment (GRACE) mission, along with soil moisture, precipitation, and evapotranspiration reanalysis. Different drought characteristics of trends, occurrences, areal-extent, and frequencies corresponding to 3-, 6-, 12-, and 24-month timescales are extracted from these indices. Drought evolution within selected basins of Africa, America, and Asia is interpreted. Canonical Correlation Analysis (CCA) is then applied to find the relationship between global hydro-meteorological droughts and satellite derived Sea Surface Temperature (SST) changes. This relationship is then used to extract regions, where droughts and teleconnections are strongly interrelated. Our numerical results indicate that the 3- to 6-month hydrological droughts occur more frequently than the other timescales. Longer memory of water storage changes (than water fluxes) has found to be the reason of detecting extended hydrological droughts in regions such as the Middle East and Northern Africa. Through CCA, we show that the El Niño Southern Oscillation (ENSO) has major impact on the magnitude and evolution of hydrological droughts in regions such as the northern parts of Asia and most parts of the Australian continent between 2006 and 2011, as well as droughts in the Amazon basin, South Asia, and North Africa between 2010 and 2012. The Indian ocean Dipole (IOD) and North Atlantic Oscillation (NAO) are found to have regional influence on the evolution of hydrological droughts.

The application of multi-mission satellite data assimilation for studying water storage changes over South America

Constant monitoring of total water storage (TWS; surface, groundwater, and soil moisture) is essential for water management and policy decisions, especially due to the impacts of climate change and anthropogenic factors. Moreover, for most countries in Africa, Asia, and South America that depend on soil moisture and groundwater for agricultural productivity, monitoring of climate change and anthropogenic impacts on TWS becomes crucial. Hydrological models are widely being used to monitor water storage changes in various regions around the world. Such models, however, comes with uncertainties mainly due to data limitations that warrant enhancement from remotely sensed satellite products. In this study over South America, remotely sensed TWS from the Gravity Recovery And Climate Experiment (GRACE) satellite mission is used to constrain the World-Wide Water Resources Assessment (W3RA) model estimates in order to improve their reliabilities. To this end, GRACE-derived TWS and soil moisture observations from the Advanced Microwave Scanning Radiometer - Earth Observing System (AMSR-E) and Soil Moisture and Ocean Salinity (SMOS) are assimilated into W3RA using the Ensemble Square-Root Filter (EnSRF) in order to separately analyze groundwater and soil moisture changes for the period 2002-2013. Following the assimilation analysis, Tropical Rainfall Measuring Mission (TRMM)'s rainfall data over 15 major basins of South America and El Niño/Southern Oscillation (ENSO) data are employed to demonstrate the advantages gained by the model from the assimilation of GRACE TWS and satellite soil moisture products in studying climatically induced TWS changes. From the results, it can be seen that assimilating these observations improves the performance of W3RA hydrological model. Significant improvements are also achieved as seen from increased correlations between TWS products and both precipitation and ENSO over a majority of basins. The improved knowledge of sub-surface water storages, especially groundwater and soil moisture variations, can be largely helpful for agricultural productivity over South America.

Assessing data assimilation frameworks for using multi-mission satellite products in a hydrological context

With a growing number of available datasets especially from satellite remote sensing, there is a great opportunity to improve our knowledge of the state of the hydrological processes via data assimilation. Observations can be assimilated into numerical models using dynamics and data-driven approaches. The present study aims to assess these assimilation frameworks for integrating different sets of satellite measurements in a hydrological context. To this end, we implement a traditional data assimilation system based on the Square Root Analysis (SQRA) filtering scheme and the newly developed data-driven Kalman-Takens technique to update the water components of a hydrological model with the Gravity Recovery And Climate Experiment (GRACE) terrestrial water storage (TWS), and soil moisture products from the Advanced Microwave Scanning Radiometer - Earth Observing System (AMSR-E) and Soil Moisture and Ocean Salinity (SMOS) in a 5-day temporal scale. While SQRA relies on a physical model for forecasting, the Kalman-Takens only requires a trajectory of the system based on past data. We are particularly interested in testing both methods for assimilating different combination of the satellite data. In most of the cases, simultaneous assimilation of the satellite data by either standard SQRA or Kalman-Takens achieves the largest improvements in the hydrological state, in terms of the agreement with independent in-situ measurements. Furthermore, the Kalman-Takens approach performs comparably well to dynamical method at a fraction of the computational cost.

Understanding the association between climate variability and the Nile's water level fluctuations and water storage changes during 1992-2016

With the construction of the largest dam in Africa, the Grand Ethiopian Renaissance Dam (GERD) along the Blue Nile, the Nile is back in the news. This, combined with Bujagali Dam on the White Nile are expected to bring ramification to the downstream countries. A comprehensive analysis of the Nile's waters (surface, soil moisture and groundwater) is, therefore, essential to inform its management. Owing to its shear size, however, obtaining in-situ data from "boots on the ground" is practically impossible, paving way to the use of satellite remotely sensed and models' products. The present study employs multi-mission satellites and surface models' products to provide, for the first time, a comprehensive analysis of the changes in Nile's stored waters' compartments; surface, soil moisture and groundwater, and their association to climate variability (El Niño Southern Oscillation (ENSO) and Indian Ocean Dipole (IOD)) over the period 1992-2016. In this regard, remotely sensed altimetry data from TOPEX/Poseidon (T/P), Jason-1, and Jason-2 satellites along with the Gravity Recovery And Climate Experiment (GRACE) mission, and the Tropical Rainfall Measuring Mission Project (TRMM) rainfall products are applied to analyze the compartmental changes over the Nile River Basin (NRB). This is achieved through the creation of 62 virtual gauge stations distributed throughout the Nile River that generate water levels, which are used to compute surface water storage changes. Using GRACE total water storage (TWS), soil moisture data from multi-models based on the Triple Collocation Analysis (TCA) method, and altimetry derived surface water storage, Nile basin's groundwater variations are estimated. The impacts of climate variability on the compartmental changes are examined using TRMM precipitation and large-scale ocean-atmosphere ENSO and IOD indices. The results indicate a strong correlation between the river level variations and precipitation changes in the central part of the basin (0.77 on average) in comparison to the northern (0.64 on average) and southern parts (0.72 on average). Larger water storages and rainfall variations are observed in the Upper Nile in contrast to the Lower Nile. A negative groundwater trend is also found over the Lower Nile, which could be attributed to a significantly lower amount of rainfall in the last decade and extensive irrigation over the region.

A study of Bangladesh's sub-surface water storages using satellite products and data assimilation scheme

Climate change can significantly influence terrestrial water changes around the world particularly in places that have been proven to be more vulnerable such as Bangladesh. In the past few decades, climate impacts, together with those of excessive human water use have changed the country's water availability structure. In this study, we use multi-mission remotely sensed measurements along with a hydrological model to separately analyze groundwater and soil moisture variations for the period 2003-2013, and their interactions with rainfall in Bangladesh. To improve the model's estimates of water storages, terrestrial water storage (TWS) data obtained from the Gravity Recovery And Climate Experiment (GRACE) satellite mission are assimilated into the World-Wide Water Resources Assessment (W3RA) model using the ensemble-based sequential technique of the Square Root Analysis (SQRA) filter. We investigate the capability of the data assimilation approach to use a non-regional hydrological model for a regional case study. Based on these estimates, we investigate relationships between the model derived sub-surface water storage changes and remotely sensed precipitations, as well as altimetry-derived river level variations in Bangladesh by applying the empirical mode decomposition (EMD) method. A larger correlation is found between river level heights and rainfalls (78% on average) in comparison to groundwater storage variations and rainfalls (57% on average). The results indicate a significant decline in groundwater storage (∼32% reduction) for Bangladesh between 2003 and 2013, which is equivalent to an average rate of 8.73 ± 2.45mm/year.