| | Type of strategy | Characteristics and related methods | Classifier |
| | 1. Pixel-based | (i) Divide pixels into groups. It is usually straightforward to put into practice. One of the critical disadvantages of pixel-based classification is that it ignores information from neighbouring pixels to help identify the target pixel’s class more correctly. Consequently, pixels in a class with a lot of spectral heterogeneity can be divided into several classes [27, 28]. (ii) One of the main drawbacks of pixel-based strategy is that it ignores information from nearby pixels which may aid in accurately identifying the target pixel’s class. As a result, pixels in a class with substantial spectral heterogeneity may be labelled as separate classes [29]. (iii) There is the issue of mismatched pixels. Pixel-based approaches for extracting low-level characteristics are commonly utilized. Pixels in the overlapping region are misclassified due to class confusion, and the picture is classified based on spectral information [25, 26, 30] | Maximum Likelihood Classifier (MLC), Spectral Angle Mapper (SAM), Support Vector Machine (SVM), Mahalanobis, Maximum Entropy, Parallelepiped, Normal Bayes | | 2. Object-based | (A) Image pixels are collected into spectrally homogenous image objects using an image segmentation approach, and then the individual objects are classified [31]. (B) This is a hyperparameter used to estimate the relevance of variables. Rather than single pixels, the approach finds clusters of pixels that reflect presently existing things in a GIS database [32] | Random Forest (RF), K-Nearest Neighbour (KNN) | | 3. Rule-based | (i) A set of classification rules make up a rule-based classifier. Rules use the absence and presence of the term. In this classifier, we establish specific criteria for producing rules, and these rules are formed during training [33]. Because a record may not trigger any restrictions, simplified rules may no longer be comprehensive. (ii) A solution for mutually exclusive practices is as follows: A set of rules are organized logically. Voting techniques are used to deal with an unorganized rule set [34]. (iii) Early AI research uses master data, agreed-upon rules, and reasoning processes to extract meaningful information from large amounts of data [35] | DTC | | 4. Distance-based | Distance-based algorithms are nonparametric techniques that may be used for classification. These algorithms sort items into strategies depending on how unlike they are, as measured by distance functions [30]. Some of the most recent uses of distance-based algorithms are also explored. The distance between pixels in the highlight space is regulated [31, 36] | MDC | | 5. Neural-based | (A) Size and complexity: It is smaller and more complicated. It is incapable of complex pattern recognition [32]. (B) Information storage is replaceable, implying that new data may be replaced by old data [33]. (C) ANN offers numerous benefits, but it also has certain drawbacks, such as extended training times, high computational costs, and weight adjustment. (D) Artificial neural networks can give input for parallel processing, which implies that they can perform several tasks at once. (E) Artificial neural networks have been met with opposition [34]. It is a mathematical model that’s been applied neutrally [35]. (F) It has many linked processing components known as neurons to perform all activities. (G) Information held in the neurons is just the weighted connection of neurons [36, 37] | ANN |
|
|
