Bare Land Referenced Algorithm from Hyper-Temporal Data (BRAH) for Land Use and Land Cover Age Estimation

Abstract

Determining the age of land use and land cover (LULC) using satellite imagery has long been one of the challenging tasks in remote sensing research. Accurately determining age, especially crop age, is essential for plot management, biomass calculations, and carbon sequestration. This research proposes a method for determining the age of LULC using hyper-temporal satellite data. The method is based on the assumption that “the starting point for the age count is when the latest bare land status disappears at any location”. To create a geospatial layer (referred to as the BR layer) that can be used to determine the age of any land cover at a specific location, we conditionally stacked such statuses obtained from the analysis of numerous satellite imagery data. The algorithm was tested at two study sites in Thailand, where rubber plantations dominated land use. The study revealed that all the rubber ages determined using BRAH fell accurately within the range of the local government survey data. The manuscript provides a straightforward explanation of the algorithm, including the pseudocode, accuracy assessment, implementations, robustness, and limitations.

1. Introduction

Imagine yourself observing a time-lapse video showcasing the complete life cycle of a bean sprout, from seed to maturity. In this scenario, determining the age of the bean sprout is a straightforward task by simply referring to the beginning of the video to identify the date it started. In the current era of big data, we are fortunate to have access to a wealth of information that enables us to simulate similar events and accurately estimate the age of land cover. Land cover age refers to the temporal information associated with different land cover types in an area. It helps study landscape dynamics, monitor land-use changes, and evaluate environmental impacts. Understanding land cover age provides insights into land management and conservation. Satellite data plays a crucial role in this process, presenting unique challenges when it comes to estimating land cover age. This challenge is particularly pronounced in the field of agricultural land management, where crop age serves as a key factor in predicting yields, establishing product pricing, and effectively managing warehouses. Moreover, plant age holds significant importance in understanding biomass dynamics and carbon sequestration, making it a valuable parameter for natural resources and environmental management [1].

Over the course of time, numerous methods have been developed and widely used for determining or calculating the age of land cover. In urban areas, the nature of land cover tends to remain relatively stable or undergo spatial changes at a lower frequency compared to other types of land cover. As a result, determining the age of built-up structures such as roads and houses is often a straightforward and accurate process. This is typically achieved by accessing relevant information from the appropriate authorities. However, in the case of water areas, forested regions, agricultural land, and open spaces, the task of determining the age of land cover becomes more challenging. These types of land cover exhibit high-frequency spatial changes that are not consistently documented, which poses significant complexities in accurately assessing their age.

This phenomenon is particularly evident in the context of agricultural land, where the age of crops undergoes dynamic changes over time. The duration of crop cycles can vary significantly, with some crops completing their life cycles within a few months while others require several decades from planting to harvest. Furthermore, in third-world countries, agricultural practices often revolve around small-scale farming conducted in private gardens or plantations. Consequently, acquiring accurate data on the age of plants necessitates conducting surveys of individual plots, which can be time-consuming and costly. Compounding the challenge is the fact that harvesting or cultivation activities for different areas may occur independently, making it challenging to maintain a comprehensive record of the current age of crops.

Throughout the years, the most widely adopted and reliable approach for estimating land cover age, particularly in the case of rubber plantations, has been the utilization of high-resolution aerial data. This method relies on visual interpretation of age ranges, considering disparities in spectral reflectance, shape, and dispersion patterns [2]. Researchers have also explored the correlation between spectral changes in various wavelengths and land cover age, employing data and techniques to enhance prediction accuracy. For instance, Chen et al. (2018) developed a tree growth model using annual time series from Landsat to estimate the age of rubber (Hevea brasiliensis) plantations. Their model achieved an average age estimation error of 1.53 years for the entire study area [3]. The authors recommended incorporating fine-scale phenological variations from multi-source time series to address the challenges posed by mixed plant species. Furthermore, Janatul et al. (2018) demonstrated that the distinctive phenological characteristics of rubber plantations can be leveraged to differentiate between age classes and other species [4]. Another study by [5] successfully established a regression model between vegetation indices derived from hyperspectral data and the surveyed age of rubber plantations, showcasing a strong statistical correlation with low root mean square error (RMSE) value.

In the case of oil palm plantations, the age of the stands can be visually discerned from the palm trees. However, spectral-based age classification encounters limitations when the palm canopy reaches full maturity, causing saturation in spectral reflectance [6]. Hamsa et al. (2018) investigated the relationship between texture measures derived from SPOT-5 images and oil palm age, hypothesizing that texture profiles could discriminate between 12 age classes of oil palm [7]. Similarly, another study [8], utilizing radar images in a forested area, demonstrated the relationship between texture profiles and stand age. The study supported the hypothesis that texture could serve as an alternative to the conventional intensity-age relationships that tend to saturate mature stands. These aforementioned age estimation methods have been applied to various types of crop land [9,10]. It is important to note that each method carries distinct advantages and disadvantages depending on the specific location and context. The accuracy of age estimation is influenced by various factors such as the ground truth method (including training and testing data), the resolution of satellite data (spectral, spatial, radiometric, and temporal), and the optimization of internal model parameters [11].

Land use change refers to the process of transitioning from one land state to another, such as the transformation from a forested area to agricultural land, from agricultural land to urban areas, or the conversion of land into a water body. In the realm of agricultural land use planning, changes resulting from human activities typically involve a preparatory phase known as the bare land process [12,13,14]. Based on this understanding, we propose a novel algorithm called the Bare Land Referenced Algorithm from Hyper-temporal Data (BRAH), which relies on this assumption. To estimate the age of a crop field using this algorithm, it is necessary to identify the crop type and gather enough satellite imagery to observe the inception of crop growth. This straightforward approach is highly robust and practical, particularly in the current era of abundant data resources. Upon obtaining a new satellite image, it can be incorporated into the existing bare land referenced layer, thereby ensuring the algorithm remains up-to-date. This article provides a comprehensive discussion of the pseudocode, implementation strategies, robustness analysis, limitations, and recommendations pertaining to the utilization of the algorithm.

2. Materials and Methods

2.1. Study Sites and Satellite Data

2.1.1. Study Sites

The BRAH algorithm was implemented in two specific sites for evaluation purposes. Site 1 encompasses Ban Na, Cham Kho, and Krasae Bon Districts in Rayong Province, located in the eastern part of Thailand. Site 2 is situated in Khiri Rat Nikhom and Phun Phin Districts, within Surat Thani Province in the southern part of Thailand, as illustrated in Figure 1. These sites predominantly consist of agricultural areas, with a particular emphasis on rubber plantations owned by local villagers.

The secondary data used in this study consisted of two main components. Firstly, the precise locations of the rubber plantations were obtained. Secondly, the corresponding ages of these plantations were categorized into three age classes: less than 1–6 years, 7–15 years, and 16–25 years, corresponding to the stages of planting, harvesting, and logging, respectively. Both sets of data were sourced from the local government authorities responsible for the respective sites. Please note that the accuracy and reliability of the secondary data provided by the local governments were assumed to be of high quality for the purposes of this research.

2.1.2. Satellite Data

For this research, satellite data was acquired from the USGS EarthExplorer website. The data encompasses a time span from 1994 to 2019, comprising a total of 74 acquisition dates (

Table 1. Satellite image data in this research.

Path/Row	Sensor	Acquisition Date
129/54	LANDSAT 5 TM	1994-01-10	1996-04-07	1999-02-26	2004-12-23	2008-03-21
Site 1		1994-03-31	1997-03-07	2001-12-12	2005-03-29	2009-08-15
		1994-09-23	1997-12-20	2001-03-02	2005-07-19	2010-12-24
		1995-05-05	1998-01-21	2001-04-19	2006-04-01	2011-02-15
		1995-05-21	1998-04-27	2003-11-19	2006-08-23	2011-09-04
		1995-06-22	1999-01-08	2003-12-05	2007-03-03	2011-11-09
		1996-01-16	1999-02-09	2004-03-26	2008-03-05
	LANDSAT 8 OLI	2013-04-20	2014-04-07	2017-03-14	2018-05-04	2019-02-16
		2013-07-25	2015-05-28	2017-04-15	2018-09-09	2019-03-04
		2014-03-22	2016-03-11
128/51	LANDSAT 5 TM	1996-02-10	1997-01-13	2000-01-04	2004-12-16	2010-01-15
Site 2		1996-12-10	1997-12-29	1999-02-02	2006-12-22	2011-01-02
		1996-12-26	1998-01-30	2000-01-20	2008-12-11	2011-01-18
		1997-01-11	1999-01-17	2003-12-30
	LANDSAT 8 OLI	2014-01-10	2014-12-12	2015-12-31	2017-02-19	2018-01-05
		2014-01-26	2015-12-15	2016-02-17	2017-12-20	2018-02-06

). It is important to note that only satellite images with minimal or no cloud cover within the study areas were selected for analysis. To ensure data accuracy and consistency, all acquired satellite images underwent essential preprocessing steps, including geo-metric correction, radiometric calibration, and atmospheric correction. These corrections were performed to account for any distortions, variations in sensor response, and atmospheric interference, respectively. By applying these preprocessing techniques, the satellite imagery was prepared for subsequent analysis and interpretation.

2.2. Bare Land Referenced Algorithm from Hyper-Temporal Data (BRAH)

The implementation of the Bare Land Referenced Algorithm from Hyper-temporal Data (BRAH) can be achieved through the following pseudocode, as depicted in Figure 2. Each step of the algorithm is explained as follows:

2.2.1. Bare Land Extraction from Each Satellite Image

• Setting up criteria for classifying bare land state

To begin implementing the algorithm, a set of criteria is established to classify the bare land state. In this particular implementation, the Normalized Difference Vegetation Index (NDVI) is utilized for extracting bare land. The NDVI is calculated using Equation (1):

NDVI = (NIR − R)/(NIR + R),

(1)

where NIR and R represent the near-infrared and red reflectance values, respectively, obtained from the atmospheric-corrected satellite image. It is important to note that the classification of land cover as bare land is not limited to NDVI thresholding alone. Other indices or classifiers can also be applied based on the specific requirements of the study. Additionally, in the case of study site 1, the classification of bare land is further refined by incorporating the Normalized Difference Bareness Index (NDBaI) thresholding in conjunction with NDVI. The NDBaI is calculated using Equation (2):

NDBaI = (SWIR − TIRS)/(SWIR + TIRS),

(2)

where SWIR and TIRS represent the shortwave infrared and thermal infrared reflectance values, respectively. The criteria for extracting the bare land index are determined once from the satellite image along with the corresponding ground truth data. It is important to ensure that the criteria are consistent across different acquisition dates or sensors by applying atmospheric normalization to the images before further processing.

2.

Classifying bare land state

To simplify the classification of the bare land state, a straightforward decision tree approach using thresholding was employed. This approach involves creating a separate layer for each single-date image, where pixels are classified as either bare land or non-bare land based on predefined threshold values.

2.2.2. Labelling

Once the bare land state has been classified for each layer, the data points identified as bare land are labeled with the date of the original image data. This labeling process results in a layer where each bare land pixel is replaced with the corresponding image acquisition date, while non-bare land pixels are assigned empty data or a value of 0. This layer is referred to as the Bare land layer.

2.2.3. Accumulating the Bare Land Layers to Create a Bare Land Referenced Layer

To generate a bare land referenced layer, a virtual layer is created with the same spatial extent as the Bare land layer. The individual bare land data layers are then intersected together. In cases where multiple data points exist at the same pixel location, the algorithm selects the data point with the most recent date and assigns it to the corresponding location in the virtual layer. Through this process, a Bare land referenced layer, known as the BRAH layer, is obtained by combining multiple bare land layers.

2.2.4. Using the BRAH Layer

The BRAH layer serves as a valuable tool for estimating the age of land cover. This estimation is accomplished by subtracting the date of the land cover map of interest from the date of the corresponding location in the BRAH layer. If the date of the land cover map is older than the BRAH layer’s date, the calculation results in a negative value, indicating that there is no available image in the database showing a bare land state for that location. Negative values are discarded. Positive values, on the other hand, represent the age of the land cover at the specific location relative to the BRAH layer. These positive values can be stored and utilized as the estimated age of the land cover.

3. Results

3.1. Study Site 1

3.1.1. Bare Land Referenced Layer Creation

For the implementation of the BRAH algorithm in study site 1 (Khiri Rat Nikhom and Phun Phin District, Surat Thani Province, Thailand), the Normalized Difference Bareness Index (NDBaI) and the Normalized Difference Vegetation Index (NDVI) were utilized as criteria for classifying bare land. Through a comparison between the field data and image reflectance, it was determined that the optimal threshold values for NDBaI ranged from −0.9987 to −0.9920, while for NDVI, the range was 0.10 to 0.40. Figure 3 depicts the resulting bare land referenced layer obtained for study site 1. This layer provides a spatial representation of the bare land distribution within the designated area.

3.1.2. Rubber Stand Age Estimation Based on the Bare Land Referenced Layers

The rubber stand age estimation was performed by subtracting the date of the rubber plantation map with the date of the BRAH layer at the corresponding location (referred to the explanation in Section 2.2.4).

For study site 1, a comprehensive quantitative analysis was conducted, revealing significant insights into the land cover dynamics. The analysis unveiled that the total extent of rubber plantations within this site spanned a vast area of 693.02 square kilometers. Among the rubber stands in study site 1, the top three most common ages were 11 years, 8 years, and 7 years. These age classes covered areas of 137.76 square kilometers, 85.17 square kilometers, and 79.65 square kilometers, respectively. These findings shed light on the spatial distribution patterns of rubber stand ages within study site 1. A visual representation of this distribution can be observed in Figure 4.

3.2. Study Site 2

3.2.1. Bare Land Referenced Layer Creation

Study site 2 located in Ban Na, Cham Kho, and Krasae Bon District, Rayong Province, Thailand, the classification of bare land in the BRAH algorithm solely relied on the Normalized Difference Vegetation Index (NDVI) as the criterion. Through a comparison between the field data and image reflectance, it was determined that the optimal threshold value for NDVI in study site 2 ranged from 0.10 to 0.25. Figure 5 displays the resulting bare land referenced layer obtained for study site 2. This layer provides a visual representation of the bare land distribution based on the NDVI thresholding approach.

3.2.2. Rubber Stand Age Estimation Based on the Bare Land Referenced Layers

The rubber stand age estimation of study site 2 was performed using the same method as study site 1. In the case of study site 2, due to the satellite imagery used, the age range in this site was narrower compared to study site 1. It should be noted that the initial age values obtained from the algorithm had a higher level of precision, such as day or month resolution. However, for the purpose of comparison with the surveyed rubber age data, the age units were rounded to the nearest year. The age map for study site 2, illustrating the estimated ages of the rubber plantations in the area, is presented in Figure 6.

4. Discussion

4.1. The Accuracy Assessment of the BRAH

The accuracy assessment of the Bare Land Referenced Algorithm from Hyper-temporal Data (BRAH) relies on two main variables: the temporal resolution of the data and the method used to determine the bare land stage.

The temporal resolution refers to the frequency at which the satellite images are captured and used in the process of creating the bare land referenced layer. Higher temporal resolution, such as having daily or even hourly images, would provide a more accurate age estimation. However, the choice of temporal resolution depends on the specific land cover type and its lifecycle. In the case of rubber plantations, which have a long lifecycle of more than 25 years, a temporal resolution of one or two images per year was considered sufficient in this study for age estimation. It is important to note that the temporal resolution requirements may vary for different applications, such as public green space monitoring, which might require higher temporal resolutions like weekly or monthly.

The process of determining the bare land stage is another critical factor affecting the accuracy of BRAH. In this study, a simple reflectance thresholding method was applied to the atmospheric normalized corrected data. While this approach is time-consuming, there are more sophisticated land cover classification methods available that have been proven to be highly accurate [15,16,17]. Some methods can automatically detect land cover types without relying on predefined spectral thresholds [18,19]. The choice of the method for determining the bare land stage depends on the specific application and the practicality of the reflectance threshold for the land cover type.

In conclusion, the accuracy of BRAH is influenced by the temporal resolution of the image data and the method used to determine the bare land stage. Both variables are subjective and depend on the specific application. However, it is important to emphasize that the temporal resolution of the data is the most critical factor. Higher temporal resolution data, such as hyper-temporal images captured hourly or daily, would be ideal for accurate age estimation of different land cover types. It is recommended to clearly specify the temporal resolution in each application that utilizes the BRAH algorithm.

4.2. The Practical Accuracy Revealed by the Classical Accuracy Assessment

In the experiment conducted on study site 1, a traditional accuracy assessment was performed using the following steps:

• The age estimation layer, as shown in Figure 5, was reclassified into three age classes based on the secondary data: less than 1–6 years, 7–15 years, and 16–25 years. These age classes correspond to the planting, harvesting, and logging stages of the rubber plantation, respectively, as shown in Figure 7.

• The sample size for the accuracy assessment was determined using Cochran’s formula [20], resulting in a random sample size of 385 locations. This sample size was selected with a 95% confidence level. The samples were drawn randomly without replacement from the age estimation layer with 125, 135, and 125 samples for less than 1–6 years, 7–15 years, and 16–25 years, respectively.
• The age class obtained from the local government survey data for each sample was compared to the age class obtained from the age estimation layer at the corresponding location. In this assessment, all 385 samples were correctly classified, resulting in an accuracy of 100%.

It is important to note that the high accuracy observed in this assessment is not surprising, given the characteristics of the BRAH algorithm and the age estimation process. The age resolution obtained from the BRAH algorithm is far greater than the resolution of the recent survey data conducted by the local government. Additionally, when downscaling the age estimation into age classes, there is an increased possibility of true positive occurrences in the accuracy assessment due to the benefit of changing resolution.

However, the accuracy and reliability of the local government data can vary depending on several factors, such as data collection methods, data management practices, and potential limitations in data quality control. The assumption about the accuracy and reliability of the data was not intended to imply an inherent guarantee of its quality. Instead, it was based on the premise that local government data is often considered a primary and trusted source for calculating important parameters in various studies within our country.

The accuracy of the age calculation with the BRAH algorithm can be improved with a higher frequency of image acquisitions. When more frequent images are available, it allows for better tracking of land cover changes over time, including the detection of bare land stages and the estimation of land cover age. With higher temporal resolution, the algorithm can capture more precise and accurate information about the timing and duration of bare land stages, resulting in more accurate age calculations for land cover types such as rubber plantations. By having access to a greater number of images taken at regular intervals, the BRAH algorithm can capture more subtle changes in land cover and provide a more detailed understanding of land dynamics. This can lead to improved accuracy in estimating the age of land cover, as the algorithm has a richer temporal dataset to work with. Therefore, acquiring images at a higher frequency can enhance the performance and accuracy of the BRAH algorithm in estimating land cover age and monitoring land dynamics over time.

4.3. Robustness, Limitations, and Suggestions

• Robustness

The BRAH algorithm exhibits robustness and versatility in its applicability to various land cover types, accommodating both high and low-speed changes. When provided with sufficient data, the algorithm produces highly accurate and reliable age classification by effectively monitoring changes from the initial stage to the desired time. While misclassification of bare land remains a potential error, it is worth noting that bare land classification is considered one of the simpler tasks in land cover classification. Alternatively, to mitigate classification errors, the trend of change can be observed by analyzing the time-series reflectivity curve of the land.

Furthermore, the bare land referenced layer can be continuously updated with new images, eliminating the need for complete regeneration with each update. The update mechanism is outlined in the final section of the pseudocode (Algorithm 1). Consequently, the BRAH algorithm is well-suited for deployment as an algorithmic solution within large databases or cloud systems where satellite imagery is consistently incorporated. Additionally, it can seamlessly run in parallel with other time-series-based tasks, particularly if the determination of land cover age is required.

The algorithm exhibits flexibility in accurately determining the age of different land cover types. As previously mentioned, hyper-temporal data enables the determination of age for all land cover types. However, lower-frequency image acquisition can also be utilized, albeit with decreased precision in age calculation.

Algorithm 1. Procedure BRAH

Procedure NDVI  # as an example # Load the multispectral image and access the metadata    Image_object = load_image(“image.tif”)    Image_date = load_metadata(“xxxx-xx-xx”)

# Define a matrix to store the image date (a) corresponding to the bare land status and thresholds    Date_matrix[a](i,j)    input criteria_max,criteria_min

# Calculate index    Red_band = Image_object.get_band(band = “Red”)    NIR_band = Image_object.get_band(band = “NIR”)    NDVI_layer = (NIR_band − Red_band)/(NIR_band + Red_band)

# Thresholding through all pixels in NDVI_layer    FOR (i,j) in range(NDVI_layer):    ## Append the Image_date to the Date_matrix if the value fall within the thresholds      IF   NDVI_layer(i,j) > criteria_min & NDVI_layer(i,j) < criteria_max          Date_matrix[a](i,j) = Image_date      ELSE          Date_matrix[a](i,j) = []      ENDIF    ENDFOR END Procedure NDVI

# Batch process    User_input = input(“Do you have more scenes to create Date_matrix? (y/n)”)    IF   User_input == “n”:        break    ELSE        a = a+1      #assumed the maximum count is k    ENDIF END WHILE

# Create an empty date array matrix for bare land reference layer    BRRef_layer = []    k = count_matrix(Date_matrix)

# Accumulating the Date_image layers to create the bare land reference layer    FOR  l = 1 to i        FOR m = 1 to j           BRRef_layer(l,m) = max (get Image_date from Date_matrix from a to k)        ENDFOR    ENDFOR

# (Optional) update the BRRef_layer    FOR  p = 1 to l        FOR q = 1 to m           IF   BRRef_layer(p,q) is [] or < Image_date of newDate_matrix              BRRef_layer(p,q) = newDate_matrix(Image_date)           ENDIF        ENDFOR    ENDFOR END BRAH

The BRAH algorithm utilizes a data-driven approach, incorporating historical data and time-series analysis to estimate the age of land cover. It explicitly considers the temporal aspect by analyzing changes over time, making it suitable for applications that require an accurate estimation of land cover age. However, the effectiveness of the BRAH algorithm is contingent upon the availability of sufficient historical data. On the other hand, existing spectral classification methods rely on statistical algorithms or clustering techniques to assign class labels based on spectral characteristics. These methods primarily focus on analyzing spectral characteristics at a specific point in time without explicitly considering temporal changes. Regarding flexibility, the BRAH algorithm demonstrates the ability to accurately determine the age of different land cover types by emphasizing the acquisition of relevant temporal information, enabling its applicability to various environments. In contrast, spectral classification methods offer flexibility but are more reliant on the spectral signatures associated with different land cover classes.

• Limitations and suggestions

The presence of floating matter that obscures land cover directly impacts the classification of bare land states, as it alters the reflectivity characteristics. To enhance classification efficiency, the utilization of microwave-based imaging techniques could be explored. Please note that, the inclusion of Synthetic Aperture Radar (SAR) imagery has the potential to mitigate errors in land cover classification in hazy conditions.

Moreover, setting an appropriate threshold for NDVI indices to classify the bare land stage across all images proves challenging due to variations in climate, lighting conditions, and other environmental factors between satellite images captured at different time periods. Expertise and knowledge in preprocessing satellite image data are crucial for adjusting these values. It is important to note that land cover characteristics directly influence the determination of the NDVI threshold. In this case, the NDVI value was set between 0 and 0.2, signifying the characteristics associated with the bare land stage. However, this range should not be interpreted to mean that NDVI values outside this range cannot represent bare land. The range was calculated based on a limited sample of data, thus lacking representativeness for the entire population.

Lastly, inadequate temporal resolution can introduce inaccuracies in the construction of the bare land referenced layer, particularly in cases where spatial phenology undergoes rapid evolution (e.g., areas with crops featuring short harvest cycles). The limitations of satellites with low temporal resolution hinder data capture during the bare land stage, leading to unidentified areas within the bare land layer. Furthermore, areas older than the date range covered by the image data cannot be calculated using the algorithm.

5. Conclusions

The Bare Land Referenced Algorithm from Hyper-temporal Data (BRAH) utilizes satellite imagery and relies on the assumption that the bare land stage can be observed and used as a reference point for estimating the age of a land cover. One of the strengths of the BRAH algorithm is its versatility. It can be applied to different types of land cover with varying rates of change. The algorithm is adaptable to different temporal resolutions, allowing for flexibility in determining land cover age. Moreover, the BRAH algorithm can be integrated into large databases or cloud systems, enabling real-time monitoring and continuous updates to the bare land referenced layer. However, it is important to consider the limitations of the BRAH algorithm. Misclassification errors, especially in the classification of bare land, can occur due to environmental variations. However, these errors can be mitigated by increasing the frequency of data acquisition. Obtaining more frequent satellite imagery allows for a better understanding of temporal patterns and reduces the risk of misclassification. By addressing its limitations and refining its performance, the BRAH algorithm may have the potential to contribute to land cover analysis and monitoring in the big data era.