Land cover change detection (LCCD) plays an increasingly important role in various practical applications, such as urban land resource management, urban expansion, and damage assessment [1
]. With the development of remote sensing techniques, obtaining remote sensing images over a certain area at different times is convenient. Therefore, the land cover change on earth’s surface can be detected through bi-temporal remote sensing images [1
In recent years, various methods based on remote sensing images, such as image difference [14
], image ratio [11
], change vector analysis (CVA) [16
], and principal component analysis-based change detection methods (CD_PCA_Kmeans) [1
], have been developed. In addition, pixel-based post-classification change detection can provide the “from to” change information compared with the binary change detection technique [18
]. Most of these methods only focus on the remote sensing images of median-low spatial resolution, although these methods have been applied in practice. Furthermore, several parameters should be determined for generating a binary change detection map (BCDM), and parameter determination is experience-dependent and time-consuming [23
In addition to the abovementioned methods, remote sensing image with high spatial resolution can be obtained conveniently through satellite or aerial platforms and can capture more ground details than remote sensing images with median-low spatial resolution. However, the high spatial resolution does not mean higher classification or change detection accuracy than median-low spatial resolution images due to the wide intra-class and minimal inter-class variances in the remote sensing images of high spatial resolution [24
]. Therefore, the spatial features of high spatial resolution image, such as the features proposed in Semi_FCM [3
], Markov random field [26
], morphological profiles [29
], and multi-resolution level set (MLS) [2
], are considered for change detection to solve the abovementioned problem. Contextual feature extraction typically depends on various parameters, such as size, shape, direction [30
], and training samples, despite improving the detection performance and accuracy. Furthermore, the contextual feature extraction algorithm is typically complex, and designing these algorithms depends considerably on experience [31
In recent years, numerous remote sensing images have been obtained conveniently, and LCCD based on remote sensing images are increasingly required [33
]. Numerous methods, such as hierarchical unsupervised change detection method [35
], urban area change detection [36
], and unsupervised change detection based on morphological profiles [13
], have been developed with the increase in data availability. However, some methods have been insufficient in terms of practical application, especially for inexperienced practitioners. Therefore, automatic or semi-automatic methods should be developed given the increasing requirements of change detection based on various images.
As mentioned previously, opportunities for improvements in terms of accuracy and usability of the LCCD system still exist despite exerting considerable efforts to develop the change detection techniques for remote sensing images. For example, few methods are generally used for change detection of median-low and high spatial resolution images. However, various detection targets in remote sensing images have different appropriate observational spatial resolutions [38
]. Remote sensing images with various spatial resolutions are required for LCCD. Therefore, a change detection system with high generality should be developed, indicating suitability for remote sensing images with high and low-resolution remote sensing images. In addition, the degree of automation for practical application should be improved.
In this work, we proposed a change detection system with a nearly automatic degree and high generality. The proposed system consists of three proposed techniques: (1) automatic noise reduction based on the Gaussian filter (GF); (2) parameter determination assistance with threshold range prediction (TRP); and (3) a modified change detection region growing algorithm (RGA) for optimizing the initial change detection result. The proposed system is data-independent and nearly automatic. Therefore, the proposed system is capable of detecting land cover change semi-automatically and has a high generality.
We conducted the experiments on four image scenes to evaluate the effectiveness of the proposed system. Two images are real multitemporal aerial datasets with a very high resolution (0.5 m per pixel) for landslide detection. The other two images, captured by Landsat-7 and Landsat-5, are open testing datasets with a resolution of 30 m per pixel. Further details are presented in the Experimental Section.
This rest of this paper is organized as follows. Section 2
introduces the proposed system. Section 3
describes the datasets and experimental comparisons. Section 4
discusses the experiments. Section 5
presents the conclusion.
The two co-registered remote sensing images are
, where H and W are the height and width of the image, respectively, that are acquired over the same area at different times
. In Figure 1
, the proposed system is composed of three main blocks intended for the following: (1) noise reduction; (2) parameter determination; and (3) improvement of change detection result. To achieve these goals, the corresponding approaches are labeled ①, ②, and ③ in Figure 1
. These approaches are described in details in subsequent sections.
The preprocessing of bi-temporal images plays an important role in change detection [40
]. In the proposed system, geometry co-registration and radiometric correction are conducted using Erdas 8.7 software. In addition, image difference, which is one of the most simple and popular methods, is adopted to generate the change magnitude image (CMI) [3
]. Moreover, the change magnitude of image difference is defined as
are the gray pixel values obtained at times
, correspondingly, and i and j are the row and column positions, respectively. A large value of
typically indicates that the probability of change at position (i,j) is high. Thus, a bright pixel obtained in image difference normally corresponds to a high probability of change. However, multitemporal images for change detection are typically different in radiation, and the difference is caused by different atmospheric conditions, solar angles, and soil moisture. Therefore, obtaining a reliable change magnitude image is necessary.
2.1. Automatic Noise Reduction Based on GF
As mentioned previously, CMI is usually corrupted by noise and resulted in an unreliable CMI. In this section, it considers that GF is frequently used to reduce the noise in the original image or intermediate results, as discussed in previous studies [44
]. However, the common drawback of using GF is the requirement for “prior” knowledge about the amount of noise corruption, although Gaussian-like noise distribution is frequently encountered in acquired data [47
]. This knowledge is valuable for the optimal choice of parameter. However, this information is frequently unavailable in practical applications. Therefore, an automatic noise reduction approach should be developed. In this section, an automatic noise reduction approach based on GF and OTSU [48
] is proposed to further refine the noise in CMI. The flowchart of this proposed approach is illustrated in Figure 2
, and its principle is explained as follows.
In the proposed automatic noise reduction method, the parameter, which is the filter radius (
) of GF, is tuned by a gradual procedure that considers the relationship between the filtered image and its corresponding OTSU-binary result. In the proposed approach,
is increased from small to large gradually, and each corresponding filtered image is automatically divided into a binary image through OTSU. The corresponding
can be considered the optimized filtered radius when the previous predicted binary threshold (
) is equal to the current threshold (
) because OTSU predicts an optimal threshold by maximizing the two-class variance [49
]. The two classes are changed (foreground) and unchanged (background) pixels, especially for change detection. If the filtered result with an increasing
will not vary with the binary threshold obtained by OTSU when combining GF and OTSU comprehensively, then the current (
) can smoothen the noise optimally and provide a stable performance of changed and unchanged pixels for an input CMI.
Based on the above principles, the details of the proposed technique are presented in Figure 2
, as follows: (1) the GF is adopted to smoothen the CMI with an initial filtering radius
; the filtered image is
; (2) OTSU is adopted to predict a binary threshold (
) based on the
; (3) an iteration is presented,
is updated by the value of
, and the updated filtering radius is used to filter the image difference (
); (4) in the next iteration, new threshold
) is obtained by OTSU based on GF with updated filtered radius; and (5) this iteration is terminated when the two predicted thresholds are equal, the optimal filtered radius of GF is determined, and the optimal filtered image (
) is generated.
2.2. Threshold Range Prediction (TRP) for Generating the Initial Change Detection Region
Obtaining the change region through binary threshold based on filtered change magnitude image (FCMI) remains difficult, although the first proposed technique can reduce much noise in CMI. Furthermore, several previous studies demonstrate that determining a threshold for dividing the CMI into a BCDM is time-consuming and experience-dependent [23
]. Therefore, the threshold range should be predicted automatically, especially for inexperienced practitioners. The flowchart of the second proposed technique is illustrated in Figure 3
In Figure 3
, several key points are worth noting: First, curve fitting based on the histogram is conducted on the filtered image; in this study, moving least squares (MLS) is adopted as the curve fitting method [54
]. MLS is used as a ten-order polynomial function to fit points and describe the trends of pixel distribution in the proposed TRP. Second, if f(x) is assumed as the curve fitting function, then S(x) is defined as the slope value function; the value of S(x) can be calculated by the first derivative of f(x). Finally, a window sliding technique is used to detect the top and bottom points at the curve of S(x); further details on this technique can be tracked in [55
Specifically, the schematic of the proposed TRP presented in Figure 4
indicates that the pixels in FCMI are observed as “changed pixels–uncertain pixels–unchanged pixels”. The noise error induced by misregistration distributes along the edge between the changed and unchanged areas [56
]; the “uncertain pixels” distribute between the changed and unchanged areas. However, the difference between the three kinds of pixels is fuzzy and uncertain in a practical application. The distribution trend of pixels should be depicted to obtain the threshold range. Therefore, firstly, the distribution of pixels in FCMI is depicted using a histogram technique, such as in [14
]. Then, a simple and classical curve fitting technique, namely, MLS method, is adopted to smoothen the noise and depict the distributed trends accurately [54
]. Finally, the slope value of the curve fitting line is calculated by using a derivative, and the bottom and top values related to the threshold range can be captured.
Two examples using the FCMI of Mexico and Sardinia Island (Italy) datasets to illustrate the performance of the proposed TRP are depicted in Figure 5
. The yellow pixels in Figure 5
are within the predicted range obtained by the proposed technique. The yellow pixels are distributed around the bright pixels, which indicate a high possibility of change.
As discussed in previous studies, threshold determination is performed by using empirical strategies or manual trial-and-error procedures, which are time-consuming and affect the accuracy and reliability of change detection [60
]. Therefore, the threshold range should be predicted for determining a rational threshold. Compared with several existing threshold decision processes for change detection [14
], the novelty of the proposed TRP is that it can automatically determine the threshold range for selecting a binary threshold. Moreover, the process of the proposed TRP is driven by the data itself and is independent of any assumption or prior knowledge.
2.3. Modified RGA for Land Cover Change Detection (LCCD)
Opportunities for improving the detection accuracy and performance still exist, although threshold range assistance is valuable for selecting a rational threshold for dividing the FCMI into a BCDM in terms of convenience. Therefore, a modified RGA is proposed in this section as the third technique in the proposed framework to improve the performance of change detection. The flowchart of the RGA is presented in Figure 6
In BCDM, the spatial change pixels continuously construct a changing area, namely, “initial growing region (IGR)”, in the proposed system. First, a single changed or unchanged pixel in the BCDM is observed as a noise pixel that can be removed automatically. To improve the performance of the BCDM further, the FCMI and BCDM are overlaid together, and the pixels for the FCMI within an IGR are used to build a corresponding credible interval (CI):
are the mean value and standard deviation of region r, respectively. The pixels on the boundary of r are seed pixels, as the green pixels displayed in Figure 7
b. Similar gray presents the adjacent change probability or ground material in Figure 7
. If the value of a seed pixel is within the CI, then the pixel will be marked as “changed”, and this iteration will be terminated until no pixel can be found and can meet this condition. An example of the proposed RGA is demonstrated in Figure 7
to clarify this growing procedure.
The proposed framework notably improved the detection accuracy in the performance and quantitative comparisons of the experiments. Furthermore, the experimental results demonstrated that: (1) the proposed framework can achieve an enhanced change detection performance with high accuracies; and (2) the proposed framework is generally suitable for aerial images with high spatial resolution and satellite images with median-low spatial resolution. The advantages of the proposed framework indicated considerable potential applications in LCCD using remote sensing images compared with the widely used methods, such as the related methods in the experiments.
In addition, the parameter setting is more convenient in the proposed framework than in CD_PCA_PCA [1
], CD_MLS [2
], and Semi_FCM [3
]. One free-parameter is required in the proposed framework. Furthermore, the second technique of the proposed framework automatically provided a range for its parameter determination, which is valuable for parameter selection, especially for an inexperienced practitioner. In Figure 16
, the pairwise green line marked the threshold range predicted by the proposed framework for each experiment. The first minimum and maximum were captured automatically and referenced to the prediction threshold range. The binary threshold could be determined easily with the suggested auxiliary range, although the predicted range cannot directly determine a rational binary threshold for dividing a CMI into a BCDM. Furthermore, the relationship between detection accuracy and parameters varied within the suggested range, as illustrated in Figure 17
. MA is increased with the increment of parameter (T) in each experiment. By contrast, FA and OE are decreased with the increment in parameter (T) in each experiment. Furthermore, FA and OE inclined gradually to the horizontal line with the increment in parameter (T) and the constraints of the predicted range. This finding is essential for the parameter determination of the proposed framework.
The growth performance of the third proposed technique, namely, RGA, is illustrated in Figure 18
to demonstrate the effectiveness of this method. In Figure 18
, the pixels of the growing area were along the change area and were continuously spatial despite the small size of the growing area. The quantitative comparison between the results of the second and third techniques (RGA) is presented in Table 5
; this comparison was performed to investigate the advantages of the proposed method. In Table 5
, the detection accuracy was relatively improved in terms of the four experimental datasets when the proposed RGA is applied to optimize the initial binary change detection result, which was obtained using the second technique.
In this work, a semi-automatic change detection framework is proposed for LCCD. The proposed framework consists of three techniques for improving the performance of change detection and promoting the automation degree for practical applications. The first technique is the automatic noise reduction based on GF, which aims to reduce the noise in a prepared CMI. Then, the TRP strategy is proposed as the second technique to improve the automation degree and assist the parameter determination of the proposed framework. Finally, the modified RGA is suggested as the third technique for improving the detection accuracy because missing alarm pixels typically distribute around a change region when a binary threshold is divided from the FCMI into the BCDM.
Four experiments were conducted on four remote sensing image scenes that are related to different actual land cover change events, including landslides and fire disasters. The experimental results demonstrate the effectiveness of the proposed framework and obtained superior detection results compared with widely used methods, such as CD_PCA_Kmeans [1
], CD_MLS [2
], and Semi_FCM [3
]. The advantages of the proposed framework can be summarized as follows: (1) the proposed framework is favorable in terms of change detection accuracy and performance; (2) the proposed framework is suitable for detecting land cover change using the remote sensing images with high and median-low spatial resolutions based on the experiment results because the detection procedure of the proposed framework is data-independent; and (3) parameter setting requires minimal tuning because it only maintains one free parameter. Furthermore, the free parameter can be determined through a procedure assisted by a predicted range. Overall, the superiority of the proposed framework indicates considerable potential applications for LCCD using remote sensing images.
In the future study, an extensive investigation of the proposed system will be conducted on additional types of remote sensing images, such as unmanned aerial vehicle image with very high spatial resolution. Theoretically, further investigations conducted through a method with different sourcing images will enhance the robustness of the method. Furthermore, additional investigation will broaden the use and applicability of the proposed method.