A Batch Pixel-Based Algorithm to Composite Landsat Time Series Images

Abstract

Compositing is a fundamental pre-processing for remote sensing images. Landsat series optical satellite images are influenced by cloud coverage, acquisition time, sensor types, and seasons, which make it difficult to obtain continuous cloud-free observations. It limits the potential use and analysis of time series images. Therefore, global change researchers urgently need to ‘composite’ multi-sensor and multi-temporal images. Many previous studies have used isolated pixel-based algorithms to composite Landsat images; however, this study is different and develops a batch pixel-based algorithm for composing continuous cloud-free Landsat images. The algorithm chooses the best scene as the reference image using the user-specified image ID or related parameters. Further, it accepts all valid pixels in the reference image as the main part of the result and develops a priority coefficient model. Development of this model is based on the criteria of five factors including cloud coverage, acquisition time, acquisition year, observation seasons, and sensor types to select substitutions for the missing pixels in batches and to merge them into the final composition. This proposed batch pixel-based algorithm may provide reasonable compositing results on the basis of the experimental test results of all Landsat 8 images in 2019 and the visualization results of 12 locations in 2020. In comparison with the isolated pixel-based algorithms, our algorithm eliminates band dispersion, requires fewer images, and enhances the composition’s pixel concentration considerably. The algorithm provides a complete and practical framework for time series image processing for Landsat series satellites, and has the potential to be applied to other optical satellite images as well.

1. Introduction

More than 8 million scenes have been gathered in Landsat series satellite data; this number is increasing at a rate of more than 400,000 scenes each year [1]. These bring considerable differences in many scientific communities [2], such as climate change, fire outbreaks, and landscape monitoring [3,4,5]. However, because of the massive amount of data, it is difficult to easily gather, handle, and evaluate. Therefore, automatic processing techniques are necessary [6].

Landsat offers almost 50 years of continuous observations with a free data access policy [7] for long-term remote sensing applications and global change studies. Therefore, multi-sensor and multi-temporal images should be composited to provide continuous sky-clear time series images [8,9]. Pixel-based image compositing is replacing traditional scene-based compositing [10] because of the creation of cloud-free, radiometric, and phenology-consistent image composites that are continuous over large areas depending on the set of user-defined criteria [11]. The average number of Landsat images included in the analysis increased from 10 in 2008 to 100,000 in 2020 [9], and the successful launch and application of Landsat 9 doubled the temporal resolution of the Landsat series satellites [1,12]. Based on such developments, dozens of new studies are conducted, which are providing humankind with a better understanding of the planet and human activities. Forests [13], agricultural mapping [14], water management [15], land cover/land change [16], disasters [17], climate monitoring [4], and environmental conservation [18] are among the research fields in which Landsat data and data processing algorithms play important roles. Furthermore, frequently, there are specific restrictions on the lengths of the observation times in practical analyses. For example, the observation period for the study of the reaction of ecological vegetation to natural floods is limited to 1–3 months [19]. In the study of a forest fire assessment, there is a need to determine the narrow period before and after the fire event [20]. However, due to Landsat’s poor temporal resolution and uncertain high cloud coverage, it is challenging to obtain the desired images within the required time frame.

Not only do enormous amounts of data and multiple sensors need image compositions, a wide range of requirements also need better compositing algorithms.

For the study of image compositing algorithms, earlier compositing methods were only developed for coarse resolution sensors, such as AVHRR [21] and MODIS [22]. However, these algorithms are not suitable for Landsat data since the sensor has a small field of view of 15${}^{\circ }$ and the repeat cycle is 16 days. Therefore, Landsat data do not provide sufficient numbers or angles of surface samplings to invert the bidirectional reflection models [23]. Fortunately, few studies were conducted about Landsat in recent years. Roy et al. [24] were the first to propose an image composite approach for Landsat ETM+ data, primarily based on a combination of maximum NDVI and the greatest brightness temperature criteria. This benchmark work has been applied in the United States and the final composites are available for free download. Potapov et al. [25] used median values of the near-infrared (NIR) band as the criteria to choose the best data; this method outperformed the standard maximum NDVI compositing method. Griffiths et al. [26] created a mechanism for calculating scores for each Landsat observation, with image compositing criteria set by weighted scores that were generated based on acquisition year, acquisition day of the year, and distance of a specific pixel to the cloud. Furthermore, White et al. [11] developed a pixel-based image compositing approach based on the sensor type, day of year, distance to the cloud, or cloud shadow and opacity. Zhu et al. [27] proposed estimating time series models for each pixel and spectral band using all available clear Landsat observations to forecast daily compositing Landsat images. Google Earth Engine (GEE) is a cloud computing platform that provides massive satellite data processing interfaces and algorithms, including Landsat. It implements a general algorithm known as median [28], based on Potapov’s [25] ideas, as well as a composition algorithm for Landsat called ‘SimpleComposite’ [28] (referred to as simple), which integrates the ideas by White [11] and Roy [24].

However, there are two huge difficulties with the algorithms mentioned above. First, it is uncertain whether to evaluate the different compositing results quantitatively. Most evaluations rely on visual inspections of natural color composites due to the lack of a standard set of reference images for comparison [8]. Second, isolated pixel-based algorithms are unconcerned about the association of pixels on the same image, which may lead to a temporal or band dispersion of pixels in the compositing results.

In this study, a batch pixel-based composition algorithm is proposed for a temporal or band dispersion of compositing results by taking five factors into account, including cloud coverage, observation time, acquisition year, observation season, and sensor types. Furthermore, rather than the traditional visual inspection, temporal dispersion is provided, which is used as an indicator to quantitatively evaluate the compositing results. This research may help to understand the cloud removal processing in a better way and create a more reasonable composition algorithm for Landsat time series images.

The structure of this paper is organized as follows. Following the introduction, Section 2 describes research materials and main details of the new batch pixel-based composite algorithm (referred to as batch). Section 3 describes the research results and compares the differences between the algorithm in this study and the median and simple algorithms provided by GEE using all Landsat 8 satellite observations in 2019 as sample data. Section 4 describes the advantages and shortcomings of this algorithm and Section 5 provides a summary of the study.

2. Materials and Methods

2.1. Data

The remote sensing data used in this study were all collected from GEE. They include Landsat 4, 5, 7, 8 in the Landsat series. Observations of Landsat 1, 2, 3 are also available but these were not considered due to the low quality and small amounts of observations, which could limit the feasibility of the application [29].

Three online sensors, ETM+ of Landsat 7, OLI of Landsat 8, and OLI2 of Landsat 9 continually contribute fresh observations, as shown in Figure 1. These summarize and illustrate the sources and timeliness of the data, which are now accessible on the GEE platform. For Landsat 7 and Landsat 8, there are around 1200 new images added per day [2], and Landsat 9 provides 740 new scenes per day [12]. It is expected that the GEE platform will soon provide these images as well as the accessibility to the general public.

In this study, all high-quality observations (the image attribute $IMAGE\_QUALITY$ equals 9) of Landsat 8 worldwide between 1 January 2019 and 31 December 2019 were collected as test sample data. These are fairly distributed over global land areas with different latitudes and longitudes, seasons, and ecological zones for which batch, median, and simple compositing algorithms are performed at each location to obtain more reliable statistical data for comparisons.

At the same time, twelve visualization locations of interest were selected at various latitudes and cloud coverage zones (

Table A1. List of sampling locations for visualization.

Location	Lon.	Lat.	Path/Row	Mean Cloud Cover (%)	Available Image Count
a	118.3367136	31.2981391	120/38	28.35	5
b	121.6889709	31.3093027	118/38	58.53	6
c	104.0712565	30.6049022	129/39	74.93	8
d	24.4716735	0.0630941	176/60	69.55	7
e	133.4560485	−15.9006582	104/71	13.01	9
f	107.4404235	65.3931218	138/14	37.65	9
g	−152.7158265	62.2972684	72/16	68.59	10
h	−97.8720765	19.3706768	25/47	29.22	9
i	−59.2002015	0.7661963	231/59	65.27	5
j	−60.6064515	−34.8341590	226/84	20.20	8
k	21.6591735	−28.2488155	174/80	0.30	10
l	−114.6070330	32.5517492	38/37	0.80	9

and Figure 2). The data acquisition time range of the visualization points was restricted between 1 May and 30 September 2020. Visual data have different date ranges than test data. The first involves limiting the scenes of available images to overcome the cloud coverage disturbance and the second to check the performances in various years. A square buffer as a mask within 5 km was used, centered on the corresponding location for image cropping. For better image details, a small buffer distance is used to reduce the field of vision (such as moving ships).

2.2. Method

2.2.1. Automated Algorithm for the Batch Composition

The batch pixel-based composition algorithm proposed in this study consists of three steps (Figure 3): (1) Calculation of the reference image to obtain the main part of the compositing result; (2) Calculation of the priority index based on the combined five factors to obtain the missing part in order of priority; (3) Composition of the main and missing parts into one image.

2.2.2. Reference Image

The reference image is the image that the user believes to be the best, which constitutes the majority of the final compositing result. It can be selected automatically from a set of images that fit the user input parameters or it can be uniquely defined by specifying an image ID. The following represents its computational logic: the user-specified image ID is preferred as the reference image; however, if it is not specified, then images are first matched according to the user-specified year, season and sensor types. Likewise, the set of matched images is sorted according to the amount of cloud coverage and finally the one with the least amount of clouds is specified as the reference image. After removing the cloud, cloud shadow, and invalid pixels [30] from the reference image, the part that remains is the main part of the compositing result. The rest of the part, other than the main part, is considered the missing part. All available observations except the reference image, including the available sensor types, years, and seasons, might be used to fill the missing part.

2.2.3. Priority Index

The purpose of the priority index is to find the optimal substitute for the missing pixels. To calculate the priority of the substitute, a criterion based on five priorities is defined.

(1)

The smaller the observation date difference between the substitute and the reference image, the higher the priority;

(2)

The smaller the acquisition year difference between the substitute and the reference image, the higher the priority;

(3)

Less cloud coverage;

(4)

Images from the same season;

(5)

Images from the same sensor type.

Based on the aforementioned criteria, the priority index is defined as follows.

$$v=\beta \times \sqrt{\sum _{i=1}^5{\alpha }_i{v}_i^2}/\sqrt{\sum _{i=1}^5{\alpha }_i}(v,v_i\in [0,1];{\alpha }_i,\beta \subseteq \{0,1\})$$

(1)

where v is the priority index with the value range of [0,1] and $\beta$ is the overall quality control coefficient of the remote sensing image, which is used to filter the observations with the poor overall image quality. When the observation parameter $IMAGE\_QUALITY$ is 9, $\beta =1$, otherwise it is considered as $\beta =0$. $v_i$ is the ith priority factor of the above-mentioned five priority factors, and ${\alpha }_i$ is the weight of the corresponding ith priority factor, which is used to control whether the factor participates in the calculation of the priority index or not. ${\alpha }_i$ takes the value of 0 or 1. The value of 1 means that the corresponding factor participates in the priority index calculation; otherwise, it does not.

The detailed calculation method of each priority factor is as follows.

$$v_1=1-\frac{\left|d_x-d_0\right|}{365}(v_1\in [0,1];d_0,d_x\in [1,365])$$

(2)

where, $d_x$ and $d_0$ are the DOY attribute of the computed image and the reference image, respectively.

$$v_2=1-\frac{\left|c_x-c_0\right|}{100}(v_2\in [0,1];c_0,c_x\in [0,100])$$

(3)

where $c_x$ and $c_0$ are the $CLOUD\_COVER$ attribute of the computed image and the reference image, respectively.

$$v_3={\left(\frac{1}{5}\right)}^{\left|y_x-y_0\right|}(v_3\in [0,1])$$

(4)

where $y_x$ and $y_0$ are the acquisition years of the computed image and the reference image, respectively. The larger the distance between the two, the lower the priority level. Since the Food and Agriculture Organization of the United Nations uses a five-year criteria for land use management [31], the power function with a base of 0.2 can better describe the priority of various years.In the function, the current year’s maximum priority value is 1. The priority value declines dramatically with the distance deviation, tending to 0 after more than 4 years.

$$v_4=\left\{\begin{array}{c}1,seaso{n}_x=seaso{n}_0\\ 0,seaso{n}_x\ne seaso{n}_0\end{array}\right.(v_4\subseteq \{0,1\})$$

(5)

where $seaso{n}_x$ and $seaso{n}_0$ are the acquisition seasons of the computed images and the reference images, respectively.

$$v_5=\left\{\begin{array}{c}1,senso{r}_x=senso{r}_0\\ 0,senso{r}_x\ne senso{r}_0\end{array}\right.(v_5\subseteq \{0,1\})$$

(6)

where $senso{r}_x$ and $senso{r}_0$ are the satellite sensor types of the computed images and reference images, respectively.

2.2.4. Composition Properties

To describe the status of the original image data, some attributes are attached to the compositing results, which include the following:

(a)

The compositing integrity;

(b)

The number of scenes used in the composition, which corresponds to the temporal dispersion;

(c)

The time and contribution ratio of each scene used in the composition;

(d)

Whether the compositing integrity is achieved or not.

The definition of integrity is:

$$Integrity=1-n_{loss}/n_{all}$$

(7)

where $Integrity$ is the compositing integrity. $n_{loss}$ and $n_{all}$ are the number of missing pixels and total pixels of the image respectively, whereas the former is divided by the latter to obtain the missing percentage. The compositing success rate can be described as:

$$Success=coun{t}_{success}/Coun{t}_{all}$$

(8)

where $Success$ is the compositing success rate. Successful composition depends on the integrity objective (The integrity objective is a user-defined threshold. A total of six thresholds are used in this study, which are $90\%$, $95\%$, $99\%$, $99.9\%$, $99.99\%$, and $99.999\%$, respectively.). If the compositing result meets the integrity objective it is considered a successful composition; otherwise, it is considered a failed composition. $coun{t}_{success}$ and $coun{t}_{all}$ represent the number of successful and total observations for a given PATH/ROW. For all of the positions with successful compositions, the required scene count is aggregated. According to the data characteristics, the required scenes count is divided into six groups: 1–1, 2–3, 4–6, 7–11, and >=12. The average cloud coverage of each group is calculated.

2.2.5. Methods of Sampling and Comparison

To visually quantify the differences between different methods, several commonly used indices, such as NDVI [32], NDBI [33], and NDWI [34] are generated by band computations, which are widely used to express vegetation, built-up areas, and water separately. The following are the definitions for the indices mentioned above:

$$NDVI=\frac{(NIR-Red)}{(NIR+Red)}$$

(9)

$$NDBI=\frac{(SWIR-NIR)}{(SWIR+NIR)}$$

(10)

$$NDWI=\frac{(Green-NIR)}{(Green+NIR)}$$

(11)

where $NIR$, $Red$, $SWIR$, and $Green$ are the reflectance of digital values of the Landsat images’ near-infrared band, red band, mid-infrared band, and green band, respectively.

After generating the three remote sensing indices as mentioned earlier, 5000 pixels are chosen randomly for comparison from the compositing results of the 12 visualization locations listed in Table A1 and Figure 2.

2.2.6. Implementation and Source Code of the Algorithm

The automated batch pixel-based Landsat image compositing algorithm has been implemented and released as a GEE application. It provides compositions for Landsat images from the previous 40 years throughout the world and nine kinds of vegetation index (VI) products such as NDVI, NDWI, NDBI, and so on. Furthermore, all outputs can be easily compared to isolated pixel-based algorithms, such as the median and simple, and differences can be seen in algorithm outputs. The source code and documentation for the GEE implementation are available at https://leejianzhou2080.users.earthengine.app/view/referencecomposite (accessed on 20 December 2021). In addition, we made all intermediate data used in the application available to the public. Users are encouraged to adopt our techniques to produce compositing products using remote sensing data.

3. Results

3.1. Spatial Distribution of Observation Frequencies and Cloud Coverage

Landsat 8 had a total of 137,594 high-quality scenes in 2019, distributed over 9331 sites (Figure 4a–c), with an average of 14.7 high-quality images per site. However, the spatial distribution varied widely, with high latitudes and polar regions being significantly lower than low and middle latitudes. The average number of observations at different latitudes was 11.5; in the Arctic, this number dropped below 5. In Antarctica, there are no available observations. This implies that different strategies are needed in different regions: polar regions need to lower both compositing integrity and success rate expectations and high latitudes need to lower integrity appropriately.

The management of cloud coverage is pivotal in remote sensing image composition. It is one of the biggest obstacles in image processing [35]. With an average cloud coverage of 31.5% ± 30.9% during 2019, Figure 4d–f show generally large cloud coverage and a wide range of variability. The frequency of observations with less than 10% cloud coverage is much lower than the total frequency of observations (the magenta lines in Figure 4e,f), and the average values in longitude and latitude directions are 4.2 and 4.7, respectively, accounting for less than 30% of the total average observations.

3.2. Global Comparison of the Three Composition Algorithms

To investigate the differences in integrity and success rates of the three compositing algorithms, six different compositing integrity goals are set ranging from 0.9 to 0.99999. To achieve the targeted integrity, compositing calculations are performed for 9331 observed locations with a limit of a maximum of 30 images of each location through the batch, median, and simple algorithms. Figure 5 and

Table A2. Compositing results of six integrity goals using three different algorithms.

Goals (%)	Methods	Observation Number to Achieve the Integrity Goal															Total
		1	2	3	4	5	6	7	8	9	10	11	12	13	14	≥15
90.000	Batch	8311	803	177	31	9	0	0	0	0	0	0	0	0	0	0	9331
	Median	205	375	1004	1740	1914	1475	996	543	376	226	183	103	90	44	57	9331
	Simple	4603	1065	630	751	781	491	370	221	133	80	66	37	42	23	38	9331
95.000	Batch	7852	983	373	80	33	8	2	0	0	0	0	0	0	0	0	9331
	Median	174	221	631	981	1589	1470	1401	906	598	388	272	189	171	103	237	9331
	Simple	3943	1263	719	525	720	619	476	304	240	140	102	60	63	37	120	9331
99.000	Batch	4397	3148	1098	371	177	71	26	21	11	4	4	1	1	0	1	9331
	Median	158	25	184	206	592	849	1306	1228	1148	878	664	470	359	261	1003	9331
	Simple	539	2642	1542	759	546	521	562	522	410	321	230	166	120	85	366	9331
99.900	Batch	681	2518	2082	1370	976	668	394	225	146	82	50	40	26	17	56	9331
	Median	158	3	136	43	211	163	477	506	873	900	1031	868	784	610	2568	9331
	Simple	159	285	746	1308	1408	1072	748	520	531	462	430	314	301	220	827	9331
99.990	Batch	342	1520	1616	1241	996	802	683	554	364	293	193	163	123	85	356	9331
	Median	158	0	129	14	191	60	325	204	537	524	805	744	871	750	4019	9331
	Simple	155	118	337	569	1015	1157	1097	778	632	483	493	403	410	300	1384	9331
99.999	Batch	169	1009	1284	1092	907	813	660	589	519	423	322	229	202	184	929	9331
	Median	158	0	127	7	182	39	273	118	438	329	672	538	803	675	4972	9331
	Simple	155	94	190	332	663	915	1043	946	880	554	515	394	452	343	1855	9331

show the relationship between compositing integrity and success rate.

From the results, the performance of the batch algorithm is far better than the median and simple algorithms. Taking the first integrity goal as an example (90%, Figure 5a), the compositing results of the top three images are significantly different: (1) If the three methods only use one image, the success rates of the batch, median, and simple algorithms are 89.1%, 49.3%, and $2.2\%$, respectively; (2) With the use of two images, the success rate of the batch algorithm increases to 97.7% while the median and simple increase to 60.7% and 6.2%, respectively; (3) With the use of three images, the success rate of the batch algorithm reaches 99.997% while the median and simple are far lower than this value. The results are similar for other integrity goals as well (Figure 5b–f).

The use of fewer observations gives larger concentrations of data as a result. The use of one scene with 100% integrity can give the ideal result to the user. However, in practice, it is frequently not possible, and the greater the integrity requirement, the more observations are used, and vice versa. Our experimental results show less than 2% possibility to achieve 99.999% integrity with one scene (Table A2). After looking at the difficulties in obtaining 100% integrity in realistic data processing, 99% integrity seems to be a better compromising option. As a result, Figure 6 shows the geographical distribution of the number of observations needed by the various algorithms to achieve 99% integrity. The batch algorithm requires just three observations to achieve the 99% integrity goal for the majority of sites (blue region in Figure 6a); however, the median and simple algorithms require much greater frequency. On the other hand, larger frequencies are primarily found in coastal and inland basin areas (the red areas in Figure 6a), indicating that more observations are needed in places where cloud cover is more. According to the relationship between the required count and cloud cover in Figure 7, the batch method’s average cloud coverage grows as the required image count rises. It demonstrates that it can discriminate between various observations and the processing of cloud coverage is more logical.

We may have more strict observation period requirements in practical analyses, such as in studies of forest fires [36,37] and floods [3], where the research duration is generally limited to 1–2 months.

Table 1. Comparison of the compositing success rate with an observation time limit of 2 months.

Methods	Integrity (%)
	90.000	95.000	99.000	99.900	99.990	99.999
Batch	99.9%	99.5%	96.6%	71.3%	50.6%	38.1%
Median	35.6%	21.5%	6.1%	3.6%	3.2%	3.1%
Simple	75.5%	69.1%	58.8%	26.8%	12.6%	8.3%

shows the compositing success rate using the three algorithms when the observation duration is restricted to about 2 months. If the user requires 90% integrity, the batch method has a success rate of 99.9%, whereas if the user requires 99% integrity, the batch has a success rate of 96.6%.

The batch pixel-based algorithm in this study has a substantially better compositing success rate than the isolated pixel-based algorithm, and significantly reduces the number of necessary satellite observations.

3.3. Visualization Results of Different Algorithms

Figure 8, Figure 9 and Figure 10 show the band 4/3/2 compositing results of twelve locations obtained from the three algorithms, as well as the available image count and average cloud coverage at that location. The following comparison focuses on three different aspects.

First of all, in terms of cloud coverage, the difference between the three algorithms is relatively small for locations with extremely low cloud coverage, such as locations e, j, k, and l. The difference between algorithms is also small for locations with exceptionally heavy cloud coverage, which may be due to a lack of data sources, such as locations d and g. The batch algorithm performs noticeably better than other algorithms with a middling cloud cover, such as locations a, b, c, f, h, and i.

Secondly, from the perspective of geographical location, locations f and g are near the Arctic area, whereas locations d and i are close to the equatorial zone. It can be seen that the batch sometimes performs well (location f and i) and sometimes as bad as other algorithms (location d and g). It indicates that the influence of latitude on the compositing results is far less as compared to the influence of cloud coverage.

Thirdly, in terms of seasonal factors, j and k are in the southern hemisphere’s winter, whereas a, b, c, and l are in the northern hemisphere’s summer. It can be seen that the compositions in winter are clearer than in the summer. This requirement is encountered by all three methods. The most probable explanation for it might be that in the winter cloud cover is much higher than in the summer. Despite considerable cloud coverage in the summer, the batch method obtains better results than other algorithms for locations a, b, and c.

From the context of visual comparison, the compositing results of the batch method are the clearest and unclouded. The batch algorithm minimizes the impact of cloud coverage and cloud shadow. In contrast, the results of the median method are randomly contaminated by a small amount of cloud cover. Similarly, the results of the simple method are affected by a large amount of cloud cover.

Band computations, such as NDVI, NDBI, and NDWI, were calculated from 5000 pixels, which were randomly selected from the visualization images as shown in Figure 11. The majority of the post-banding data of the batch algorithm were derived from the one-view image and a small amount of missing data were derived from the observations where the substitutes were located (Figure 11a,d,g). In contrast, the median (Figure 11b,e,h) and simple (Figure 11c,f,i) algorithms have scattered data sources and different bands from different observations, which brings uncertainty to the results of the band computation.

In comparison to the isolated pixel-based algorithm, two advances were obtained in this work based on the novel batch pixel-based algorithm. (1) The composite result shows a considerable reduction in pixel dispersion. (2) Band dispersion is no longer an issue.

4. Discussion

4.1. Principles of the Compositing Algorithms

In principle, isolated pixel-based algorithms are a group of algorithms that use a single pixel as the unit of computation to determine the best value for each pixel in each band using a statistical aggregation approach. For example, when the median algorithm calculates the B1 band, it first generates a set of B1 bands from all alternative images, and then generates a subset of all pixels at each position. Onward, the median or mean value of each subset is picked as the best value for that position. Furthermore, the simple algorithm appears similar to the median. The only difference between the two is that the median is replaced with the minimal value of the cloud possibility fraction. As a result, the isolated pixel-based algorithm’s primary criterion for pixel selection is numerical statistics. It does not distinguish the overall observation quality of the image and does not examine the relationship between pixels on the same image. This might lead to issues such as the unpredictability of pixel dispersion and band dispersion. The unpredictability of pixel dispersion in this context shows that distinct pixels are derived from separate observations in which dispersion seems unpredictable and increases quickly as the alternative set increases. The unpredictability of band dispersion exhibits that pixels in different bands are derived from different observations. For example, the red band may be derived from a bright April observation but the NIR band may be produced from a foggy June observation and the changing time and weather conditions may have unpredictable effects on the band computations.

The algorithm used in this study is different from the pixel-by-pixel selection described earlier, in which the best options are picked in batches rather than individually. In detail, it chooses the reference image that best meets the aim of the user’s research as the main portion of the result. Further, it selects the best substitute from the list of alternatives if the main portion is insufficient. Furthermore, a well-designed model (see Section 2.2.1) determines the priority of the alternatives automatically, taking into account of the observation quality, cloud cover, observation time, and sensor types of alternative images. As a result, the pixel dispersion is minimized because the reference image accounts for the majority of the final output, and the algorithm does not suffer from band dispersion because all bands of the alternative are processed as a whole.

4.2. Effects of Cloud Coverage, Latitude, and Season

The two primary factors affecting image compositing are “available observations count” and “cloud coverage”. The available count is determined by satellite parameters and observation conditions, which are directly affected by cloud coverage.

Latitude and season are two sub-factors that also make a difference. Seasonal variations frequently affect the available clouds and cloud coverage. For instance, in the northern hemisphere, the available count is substantially lower in the summer than in the winter due to the frequent surface atmospheric activity. The cloud coverage is significantly higher than it is in the summer. Furthermore, many remote sensing applications have specific limits for the season or observation period, which greatly affect the availability of remote sensing data.

From our statistics of all the scenes of Landsat 8 in 2019, the available counts in polar regions are significantly less than the lower altitude regions. The cloud coverage in equatorial regions and coastal regions is higher while in desert regions and inland regions it is relatively low.

On the other hand, in areas with extremely insufficient or excessive cloud coverage, there is basically little difference in the composite results; in other areas, the difference between the three algorithms is noticeably prominent. It should be emphasized that cloud cover continues to affect the compositing algorithms even after the cloud removal. It is possible that this is due to the cloud detection algorithm’s limitations for thin clouds, cloud shadows, etc. [38,39].

4.3. Significance of This Study

The algorithm in this study gives more reasonable results for the composite processing of multispectral optical satellite images and is especially valuable for the composite processing of long time-series images, which applies to Landsat series satellites, but can also be used with Sentinel-2, MODIS, and other sensors. It is especially suitable for remote sensing applications that require high compositing integrity, such as phenology [5], land use classification [13], and climate change [40] studies.

Moreover, the present algorithm is customizable and the parameter settings make it possible to meet the needs of various remote sensing applications. If the user is sensitive to the acquisition time, a relatively small number of observations can be used. If the user is sensitive to the integrity of the composition, a large enough observation frequency can be used to obtain higher compositing integrity.

The major contributions of this study are:

(1)

The paper proposes a new automated compositing algorithm for Landsat multi-temporal series images, which is implemented on the GEE cloud platform and is characterized by wide adaptability and high data concentration.

(2)

It suggests a method for quantifying composition assessment using temporal dispersion instead of human visual inspection and it solves several shortcomings and limitations of isolated pixel-based compositing algorithms.

4.4. Limitations and Potential Improvements

The compositing algorithm proposed in this study works well in most cases, but if the probability of local cloud coverage is too high in some extreme cases, it will lead to a small percentage of valid pixels in the reference image, resulting in a large percentage of missing pixels and a high degree of pixel dispersion in the compositing results. As a result, such compositing results may cast doubt on some specific analyses. The fundamental cause of this problem is the inadequate quality of the observations. In order to overcome it, the user may need to reduce the integrity of the composition to improve data concentration. In addition, to address this problem, extra observations might be combined from other similar satellites, such as Sentinel-2 and MODIS [41,42,43].

On the other hand, the algorithm in this study majorly depends on cloud and cloud shadow detections. It uses the Fmask [27,30] for cloud masking in the demonstration application mentioned in Section 2.2.6. In other words, it is the post-processing of the cloud-removal process. It follows that the compositing results of this algorithm, similar to other compositing algorithms, are limited by the influence of cloud and cloud shadow identification [35].

5. Conclusions

This study focuses on the compositing algorithm for optical satellite time-series images and proposes a batch pixel-based compositing algorithm for Landsat with the implementation and source code on the GEE platform. The algorithm uses all valid pixels from the reference image as the main part of the results, then selects substitutes in batch for the missing portions using a priority coefficient model. In the end, the main part and the missing parts are merged into the compositing result. Using all Landsat 8 observations from 2019, this study compares our algorithm and the isolated pixel-based algorithm provided by the GEE platform, highlighting solutions to the following problems: (1) Unpredictable pixel dispersion; (2) Unpredictable band dispersion; and (3) Reducing cloud and cloud shadow interference. Therefore, this algorithm has less bias, requires fewer observations, lowers pixel dispersion, and improves the future analysis of Landsat images. The approach provides a full and useful framework for the composition of long time-series images from Landsat series satellites with the potential to composite other optical satellite images.