Classifying each pixel in a digital image into distinct land use/land cover classifications is the aim of image classification. There are two categories based on how the interpreter and computer interact throughout the classification process. Supervised and unsupervised classification procedures are the two primary categories that are utilized to produce categorized output. When an analyst is well-versed in the field, they tend to choose supervised classification out of the two main image classification techniques. Analysts choose representative samples for every land cover class in supervised classification. These “training sites” are then applied to the full image by the software. The spectral signature specified in the training set is used in supervised classification. In supervised classification, a variety of classification methods are employed, including minimum-distance and maximum likelihood.
The concept behind supervised classification is that a user can choose a subset of pixels in an image that represent particular classes, and then instruct the image processing program to use these training sites as a guide when classifying all of the other pixels in the picture. Training locations are chosen according to the user’s level of expertise. Additionally, the user establishes the criteria for how similar other pixels must be in order to be grouped together. These limits are frequently established using the training area’s spectral properties, plus or minus a predetermined increment. Additionally, the user specifies how many classes the image belongs to.To provide final output analysis and classified maps, many analysts combine supervised and unsupervised classification procedures.
- 1. Training data
- 2. Feature selection
- 3. Selection of appropriate classification algorithm
- 4. Post classification smoothening
- 5. Accuracy assessment
1. Training Data:
Training fields are areas of known identification that are delineated on the digital picture. Typically, this is done by utilizing line and column numbers within the digital image’s coordinate system to indicate the corner points of a rectangular or polygonal area. Naturally, the analyst needs to be aware of the appropriate class for each sector. Typically, the analyst starts by compiling maps and aerial photos of the region that has to be categorized. For every informative category, specific training areas are determined based on the essential training area criteria. Finding a collection of pixels that faithfully captures the spectral variance found in each information region is the goal.
Important Features of the Training Area
a) Shape: Training area shapes are irrelevant as long as they don’t make it difficult to accurately delineate and position region outlines on digital photographs. Rectangular or polygonal areas are typically the easiest to define because they require fewer vertices to be provided.
b) Location: The placement of many training regions throughout the image should represent each informational category, therefore location is crucial. Training areas should be placed where their outlines may be easily and accurately transferred from maps and aerial photos to the digital image. The training data must not be grouped in areas of the image that are preferred because this may not accurately reflect the conditions seen across the image. This is because the training data are meant to depict variation within the image.
c) Number: The ideal number of training areas is determined by the variety of the categories to be mapped, the number of training areas that may be defined, and the resources available. To guarantee that the spectral characteristics of every category are represented, a minimum of five to ten training areas should be assigned to each information category or spectral subclass.
d) Positioning: Training areas ought to be positioned within the image in a way that allows for precise and easy location in relation to distinguishing elements like water or the border between distinguishing features. They ought to be dispersed across the picture in order to serve as the foundation for depicting the scene’s diversity.
e) Uniformity: A good training area’s homogeneity and uniformity are arguably its most crucial characteristics. For each spectral band to be used, the data in each training area should show a unimodal frequency distribution.
Considering Signatures:
To ascertain whether the signature data accurately depicts the pixels that need to be categorized for each class, several tests must be carried out. Signatures produced during supervised or unsupervised training can be assessed. There are several techniques for assessing the signatures:
- Graphical method
- Signature saperability
- Divergence
- Transformed divergence
i) Graphical Method: For each pair of bands, create and view scatter plots and elliptical diagrams of the data file values.
ii) Signature Saperability: This statistical metric quantifies the separation of two Any band combination used in categorization can have its saperability determined. The Euclidian distance is calculated using this method as follows:
Where,
- D=spectral distance
- n=number of band
- i=a particular band
- di=data file value of pixel d in band i
- ei=data file value of pixel e in band i
iii) Divergence:
Where
i,and j=the two signatures being compared
Ci=covariance matrix of signature i
µi=mean vector of signature i
tr=trace function
T=transposition function
iv) Transformed Divergence:
Where,
- i,and j=the two signatures being compared
- Ci=covariance matrix of signature i
- µi=mean vector of signature i
- tr=trace function
- T=transposition function
Divergence values can be on a scale of 0 to 2,000. Generally speaking, the classes can be divided if the result is higher than 1,900. The gap between 1,700 and 1,900 is decent. The separation is weak below 1,700.
Choosing the right classification algorithm
It is possible to assign an unknown pixel to one of several classes using a variety of supervised classification algorithms. The type of input data and the intended result determine which classifier or decision rule is best. In the training phase of the supervised classification, the parametric classification algorithm assumes that the observed measurements Xe for reach class in each spectral band are Gaussian, or regularly distributed. Such assumptions are not made by non-parametric classification algorithms. The greatest likelihood, minimal distance, and parallelepiped decision criteria are some of the most widely used classification methods.
A) Classification Algorithm using Parallelepipes:
This popular decision rule is based on basic “and/or” Boolean logic. The categorization is carried out using training data in n spectral bands. The multispectral imagery’s brightness values are utilized to create an n-dimensional mean vector.
With mck representing the mean value of the training data acquired for class c in band k out of m available classes, Mc = (µck1, µc2, µc3, … µcn). Out of m possible classes, Sck is the standard deviation of the training data for class c of band k.
One computationally effective technique for categorizing data from remote sensors is the parallelepiped algorithm. Regretfully, an unknown candidate pixel may meet the requirements of multiple classes due to the overlap of some parallelepipeds. In these situations, it is often placed in the first class for which it satisfies all requirements. Using a minimum distance to means decision rule to assign this pixel, which can be assigned to multiple classes, to only one class is a more elegant method.
B) Minimum Distance to Mean Classification Algorithm:
This decision rule is widely used and has a straightforward computation. When applied correctly, it can produce classification accuracy on par with other techniques that need more computing power, such the maximum likelihood approach. It requires the user to supply the mean vectors for every class in every band mck from the training data, just like the parallelepiped algorithm does. A program must determine the distance to each mean vector, mck, from each unknown pixel (BVijk) in order to carry out a minimum distance classification.
The formula: Dist = SQRT (BVijk -mck ) 2 + (BVijl -mcl) 2 is used to calculate the Euclidean distance from a point to the mean of Class-1 measured in band, where mck and mcl stand for the mean vectors for class c measured in bands k and l.
Even though a pixel is closest to the category mean, many minimum-distance algorithms analysts set a threshold or distance from the class means beyond which it will not be classified as belonging to that category.
The reasoning behind calculating the distance between only two locations in n space can be extended when more than two bands are assessed in a classification by applying the equation n
Dab = Ʃ( ai – bi)2
I=1
C) Maximum Likelihood Classification Algorithm
Every pixel with pattern measurements or features X is assigned to class c using the greatest probability decision rule, whose units are most likely to have produced feature vector x. It makes the assumption that the training data statistics are Gaussian, or regularly distributed, for every class in every band. Stated differently, it is not optimal to use training data that has trimodal or bimodal histograms in a single band. In these situations, the individual modes most likely correspond to distinct classes that ought to be trained on separately and designated as such. The normal distribution condition would then be satisfied by the unimodal, Gaussian training class statistics that would result from this.
Statistics such as the mean measurement vector, divergence, standard deviation, and probability are used in maximum likelihood classification. Each class’s mean (Mc) and the class c covariance matrix (Vc) for bands k through l. The unknown measurement vector X is subject to the following decision rule.
Determine that X belongs to class C if and only if
where i = 1, 2, 3, …, m possible classes, and pc³ pi
and
PC is equal to {-0.5loge[det(Vc)]}. If (X-Mc)TVc-1 (X-Mc) = 0.5
where the covariance matrix Vc’s determinant is det(Vc). Therefore, the maximum probability decision rule calculates the value pc for each class in order to categorize the measurement vector X of an unknown pixel into a class. The pixel is then allocated to the class with the highest value.
Let’s now examine the necessary calculations. The covariance matrix and mean vectors for class 1 are V1 and M1, respectively, while p1 is calculated in the first run. Next, V2 and M2 are used to calculate p2. This holds true for every m class. The class that generates the largest or maximum PC is given the pixel or measurement vector X. Each calculation step uses a measurement vector X with n elements (the number of bands under analysis). For instance, each unknown pixel would have a measurement vector X of if all six bands were being examined.
BVi,j,l
BV ij,2
BVij, 3
X = BVij,4
BVij, 5
BVij,6
Because it does not presume that each class has equal probabilities (equal probability contours are depicted in Figure 6), the Bayes decision rule is the same as the maximum likelihood choice rule. In order to improve classification accuracy, a priori probabilities have been effectively employed to incorporate the impacts of relief and other terrain characteristics. Compared to the shortest distance or parallelepiped classification methods, the greatest likelihood and Bayes classifications need a lot more calculations per pixel. They don’t always yield better outcomes.
Evaluation of Classification Accuracy:
It is necessary to gather some in situ data or a priori knowledge about specific terrain features in order to quantitatively evaluate classification accuracy. This information may then be compared with the classification map that was produced using remote sensing. Therefore, two classification maps must be compared in order to evaluate classification accuracy. 2) the presumed genuine map, and 1) the map obtained from remote sensing. In situ research or, more frequently, the interpretation of remotely sensed data acquired at a better resolution or on a wider scale can be used to create the presumed genuine map.
Overall Accuracy Evaluation of Classification Maps
Finding out if a remotely sensed classified map satisfies or surpasses specific predefined classification accuracy standards is essential to assessing the map’s overall correctness. The agreement between the two maps in terms of total area or each category is assessed by the overall accuracy assessment. Typically, they don’t assess construction faults that arise in the different categories.
Accuracy Evaluation of Site-Specific Classification Maps
This kind of error study contrasts the pixel-by-pixel accuracy of the categorization map obtained from remote sensing with the presumed actual land use map. First, in a supervised classification, a site-specific error evaluation can be performed using just the training pixels that were used to train the classifier. This only indicates that the classifier is trained using pixel locations i, j that are carefully assessed on both the presumed actual map and the categorized map from remote sensing data products. This assessment might be regarded as representative of the study region if training samples are dispersed at random throughout the study area. if they behave biasedly because the analyst knows in advance where specific forms of land cover are present in the scene. Due to this bias, pixels located within the training sites—the data locations used to train the classifier—generally have higher classification accuracy than pixels located elsewhere on the map.
The accuracy of the classified map, on the other hand, can be assessed if additional test locations inside the study region are found and appropriately labeled before classification, even if they are not utilized in the classification algorithm’s training. In general, this process produces a more reliable categorization accuracy evaluation. However, in addition to the challenge of figuring out how many pixels are needed in each test site class, these test sites also require extra ground truth. On the other hand, even if they are not included in the training of the classification algorithm, the accuracy of the classified map can be evaluated if more test locations inside the study region are located and suitably labeled prior to classification. Generally speaking, this procedure yields a more trustworthy assessment of categorization accuracy. These test sites, however, require additional ground truth in addition to the difficulty of determining the number of pixels required in each test site class.
The class assigned to each pixel in the remote sensing derived map and the assumed true map must be identified after the criterion for objectively identifying the location of specific pixels to be compared is established. The results are tabulated and reported in a contingency table (error matrix), where overall classification accuracy and misclassification between categories are identified.One of the most significant features of such matrices is their ability to summarize errors of omission and commission. These procedures allow quantitative evaluation of the classification accuracy, and their proper use enhances the credibility of using remote sensing derived land use information. t takes the form of a m x m matrix, where m is the number of classes under investigation. The rows in the matrix represent the assumed true classes, while the columns are associated with the land use.
Kappa Coefficient:
Since 1983, discrete multivariate approaches have been frequently utilized to statistically assess the accuracy of maps and error matrices obtained from remote sensing. Since the remotely sensed data is discrete rather than continuous and is binomially or multinomially distributed rather than regularly distributed, these methods are suitable. A Khat statistic, which is a measure of accuracy agreement, is produced via kappa analysis. Calculating the Khat statistic is done as
Khat = N”Ʃr xii – (Ʃr xi+ * x+i) / N2 – Ʃr (xi+ * x+i) r
where N is the total number of observations, r is the number of rows in the matrix, xii is the number of observations in row I and column I, and xi+ and x+i are the marginal totals for row I and column I, respectively.
