close column table with tabularx in LaTex

The problem is that when one sums up all your manually defined columns widths, they are larger than the available \textwidth. As you are using a tabularx environment, you could instead use X columns and latex will automatically calculate a suitable width:

\documentclass{article}
\usepackage{tabularx}

\usepackage{paralist}

\makeatother
\begin{document}
\begin{table}[!b]
\centering


 
\caption{ Comparison between ST-Kriging, Bayesian inference, and ANNs.}
\label{tab:2}
\renewcommand{\arraystretch}{1.5}
\scriptsize

\begin{tabularx}{\textwidth}{|X|X|X|X|}

\hline
\textbf{} & \textbf{Bayesian Inferences} & \textbf{ST-Kriging} & \textbf{ANNs}  \\ \hline
\textbf{Computational Complexity}   & NP-hard \cite{ref-satria2020spatial} & $O(N^2)$ & $O(i\times o\times n + n\times o)$ or $O(n \times o \times (i+1))$ for training a single epoch. \cite{ref-taylor1995freeway} \\ \hline
\textbf{Performance Evaluation }  & Provides a posterior probability distribution with confidence interval.  & 
Ensure linear unbiased predictors. & Epoch with the lowest sum of squared error.\\ \hline
\textbf{Weaknesses}   & Very computationally intensive due to choosing the proper prior distribution. 
& 

\begin{compactitem}
\item Missing value causes error in unmatched dimensions.
\item Can not handle large datasets.
\item Require normal distribution.
 \end{compactitem} 

 &
 Require intensive data training, and this might lead to an overfitting problem. \\ \hline
\textbf{Strengths }  
& 
\begin{compactitem}

\item Handle large and small data. 
\item Handle missing values.
\item Prior knowledge about uncertain input is not required.

\end{compactitem}
 &
\begin{itemize} 
\item Handle small data.
\item Computational efficiency.
\end{itemize}
&

\begin{compactitem}
\item Handle big data and small data.
\item Accommodate missing values without a separate estimation step [108]
\item Computational efficiency due to the parallelity feature.
\item Prior knowledge about uncertain input is not required.
\end{compactitem}
\\ \hline
\textbf{Overcoming the Limitation}   
&
Use uninformative prior to reduce the computational time, however, it can affect the prediction accuracy negatively.
&
Remove observations that include missing values.  
& 
\begin{compactitem}

\item Decrease the number of layers of the network.

\item  Use iterative methods to stop the training process such as gradient descent.
\end{compactitem}
    \\ \hline
\end{tabularx}


\end{table}

\end{document}

If you use tabularx one of the column should be X which adapts its width to a remaining available width.

Your table is quite wordy and due to number of columns there doesn't seem to be enough space for long texts. You could consider rotating a table and use a pageheight as an available space for a width of the table.

While correcting your table, I have also made a few adjustments:

• indentation removed in lists to increase horizontal space
• the caption moved upwards to add a little space between title and the table
• increased vertical spacing in the first row by inserting a strut
• space between columns slightly decreased
• added rules from booktabs
• reduced margins on the page with the table.

\documentclass{article}
\usepackage{tabularx}
\usepackage{calc}
\usepackage{paralist}
%% Added packages
\usepackage{booktabs}   % For improved rules
\usepackage{caption}    % For extra space between table and caption
\usepackage{rotating}   % For sidewaystable
\usepackage{ragged2e}   % Improves typesetting in
\usepackage{microtype}  %   in narrow environments
\usepackage{geometry}
\captionsetup[table]{position=top,skip=3pt,font=normalsize}
%% END Added packages

\setdefaultleftmargin{0em}{}{}{}{}{}


\begin{document}

\newgeometry{top=1in,bottom=1in}
\begin{sidewaystable}
  \RaggedLeft
  \setlength\tabcolsep{4pt}
  \newcommand\xstrut[1][1]{%
    \rule[-0.35\normalbaselineskip*\real{#1}]{0pt}{#1\normalbaselineskip}}
  \newcommand\xthead[1]{\normalsize\textbf{#1}}
  \renewcommand{\arraystretch}{1.25}
  \centering
  \caption{Comparison between ST-Kriging, Bayesian inference, and ANNs.}
  \label{tab:2}
  \footnotesize
  \begin{tabularx}{0.9\textwidth} {p{3cm}XXX}
    \specialrule{\heavyrulewidth}{0pt}{0pt}
    \xstrut[2.2]
    & \xthead{Bayesian Inferences}
    & \xthead{ST-Kriging}
    & \xthead{ANNs} \\
    \midrule[\heavyrulewidth]
    \textbf{Computational Complexity}
    & NP-hard \cite{ref-satria2020spatial}
    & $O(N^2)$
    & $O(i\times o\times n + n\times o)$ or $O(n \times o \times (i+1))$\newline for training a single epoch. \cite{ref-taylor1995freeway} \\
    \midrule
    \textbf{Performance Evaluation}
    & Provides a posterior probability distribution with confidence interval.
    & Ensure linear unbiased predictors.
    & Epoch with the lowest sum of squared error. \\
    \midrule
    \textbf{Weaknesses}
    & Very computationally intensive due to choosing the proper prior distribution.
    & \begin{compactitem}
        \item Missing value causes error in unmatched dimensions.
        \item Can not handle large datasets.
        \item Require normal distribution.
    \end{compactitem}
    & Require intensive data training, and this might lead to an overfitting problem. \\
    \midrule
    \textbf{Strengths }
    &
    \begin{compactitem}
      \item Handle large and small data.
      \item Handle missing values.
      \item Prior knowledge about uncertain input is not required.
    \end{compactitem}
    &
    \begin{itemize}
      \item Handle small data.
      \item Computational efficiency.
    \end{itemize}
    &
    \begin{compactitem}
      \item Handle big data and small data.
      \item Accommodate missing values without a separate estimation step [108]
      \item Computational efficiency due to the parallelity feature.
      \item Prior knowledge about uncertain input is not required.
    \end{compactitem} \\
    \midrule
    \textbf{Overcoming the Limitation}
    &
    Use uninformative prior to reduce the computational time, however, it can affect the prediction accuracy negatively.
    &
    Remove observations that include missing values.
    &
    \begin{compactitem}
      \item Decrease the number of layers of the network.
      \item  Use iterative methods to stop the training process such as gradient descent.
    \end{compactitem} \\
    \bottomrule
  \end{tabularx}
\end{sidewaystable}
\restoregeometry

\end{document}