3.1.

Based on Sensors

Remote sensing has the potential to provide significant support in oil palm monitoring and potential oil palm plantation predictions.⁷ Several studies in agriculture and plantations, particularly in oil palm plantations, have shown that remote sensing technology is capable of providing accurate performance estimates of the derivation of the relationship between spectral characteristics and vegetation indices in the form of canopy cover using the reflection of visual waves in red, green, blue, and infrared wavelengths.²⁴ The interaction between sensors and the Earth’s surface can be categorized into two modes: passive sensors and active sensors.²⁵

Passive sensors use solar radiation to reflect the Earth’s surface and detect surface reflections. They normally absorb electromagnetic waves in the visible (430 to 720 nm) and near-infrared (750 to 950 nm) ranges.²⁶ Passive satellite sensors include those onboard Landsat, SPOT, Pléiades, EROS, GeoEye, WorldView, and other satellites.¹¹ Active sensors depend on internal energy sources, whereas passive sensors rely on external energy source mainly sunlight.²⁷ These utilize electromagnetic waves in the visible and near-infrared ranges (a laser rangefinder or a laser altimeter), as well as radar waves [synthetic aperture radar (SAR)].²⁸ Because of their potential to penetrate clouds and collect data under all-weather day/night environments, active sensors have emerged as a promising technology in the field of remote sensing.²⁹ TerraSAR-X, Sentinel-1A, ALOS PALSAR, RADARSAT, LiDAR, and other active sensors are examples. Passive and active sensors may be used to map oil palms based on the phenology or image characteristics of oil palm plantations³⁰ by measuring and classifying reflected energy using spectral signatures.³¹

Several techniques have been established to map oil palm land cover throughout the tropics utilizing satellite remote sensing data.³² Oil palm mapping and change detection methods have been developed, with data sources derived from optical satellite Earth observations^33,34 to microwave datasets,^35,36 spatiotemporal resolutions from the regional to the national scale³⁷ and from single to multidecadal mapping,⁶ interpretation methods ranging from manual to semiautomatic and fully automatic identification,^6,38–40 and products ranging from oil palm land cover maps to more detailed datasets on plantation structure. These include methods based on phenology for the classification of oil palm plantations,^41–44 tree counting,^{34,36,38,40,45} age estimation,^10,21,46,47 yield estimation,^47–49 and a few studies have also focused on the detection of continuous oil palm changes,^50–52 allometric calculations of oil palm trees,^53–57 above-ground biomass estimation (AGB),^54,58–62 carbon estimation,^55,63–65 pest and disease detection.^66–68

Gutiérrez-Vélez and DeFries⁶⁹ used Landsat to investigate the usage of spatial changes in vegetation greenness to detect deforestation caused by large-scale oil palm expansion in the Peruvian Amazon. McMorrow¹⁰ estimated oil palm age using Landsat by measuring the reflectance value and showing the relationship between oil palm age. Sitoms et al. ⁷⁰ also used Landsat as well as Santos and Messina,⁷¹ Morel et al.,⁶² Vadivelu et al.,⁷² Lee et al.,³³ Fitrianto et al.,⁷³ Carolita et al.,²¹ Danylo et al.,⁷⁴ Shaharum et al.,⁷⁵ and Sarzynski et al.³⁰

High-resolution satellite datasets can also produce optimal performance in estimating the age of oil palm trees, as Chemura et al.¹ demonstrated by integrating multispectral data with WorldView-2 and performing regression analysis to reveal the oil palm age with biophysical parameters. IKONOS was used by Thenkabail et al.⁵⁵ and Agustin et al.⁴¹ Active sensors can also focus on the physics of scattering processes. This method has primarily been used for identification purposes and is capable of understanding backscatter value using SAR data. Nordin et al.⁷⁶ used AIRSAR data in the classification of oil palm age based on biophysical parameters. Other imagery that is commonly used is ALOS PALSAR. For example, Darmawan et al.⁴⁷ used ALOS PALSAR-2 to show a correlation between backscatter value and oil palm age. In another case study, Yayusman and Nagasawa⁷⁷ classified young oil palms and mature oil palms for a smallholder plantation. Morel et al.⁶¹ estimated biophysical parameters to identify oil palm age, and Tan et al.⁴⁹ estimated tree height and diameter at breast height. Furthermore, recent studies have investigated or made comparisons between sensors. For example, Avtar et al.⁷⁸ investigated a model of oil palm age using ALOS PALSAR, RADARSAT, and TerraSAR-X data. Carolita et al.²¹ made the comparison between active sensor (Sentinel-1A) and passive sensor (Landsat-8). Darmawan et al.⁷ investigated an oil palm phenology scattering model based on TerraSAR-X, Sentinel-1A, and ALOS PALSAR-2 data.

Recent studies have combined passive and active sensors with the aim of obtaining high accuracy; for example, Lee and Bretschneider⁷⁹ discovered that combining microwave and optical data could effectively map oil palm trees in a heterogeneous environment with an overall accuracy of 94%, relative to using Landsat (84%) and PALSAR (89%) data separately. Sarzynski et al.³⁰ used Landsat-8 and ALOS PALSAR data for oil palm age classification. Danylo et al.⁸⁰ mapped the extent and age of productive oil palm plantations using Landsat and Sentinel-1A data.

Aside from the radar–optical combination, the integration of light detection and ranging (LiDAR) and optical data has been used in oil palm mapping. LiDAR, which is recognized for its extremely high spatial resolution, can generate three-dimensional (3D) details (measuring canopy height and terrain elevation). It offers a 3D view of current oil palm plantation conditions when combined with an optical image, which is incredibly useful for plantation management and planning.⁴⁵ For example, Avtar et al.⁸¹ used multispectral unmanned aerial vehicle (UAV) LiDAR data for estimating biophysical parameters.

3.2.

Based on Biophysical Tree Parameters

In traditional management practices, oil palm phenology is documented when the oil palms are first planted in the field by labeling the field according to the year of planting.⁴⁹ As oil palms grow, they demonstrate an allometric formation, in which various tree parts grow at different speeds.⁵³ Individually, these biophysical structures of the oil palm may be calculated. They are classified as biophysical parameters, and some of them are known to be associated with the oil palms’ growth stages (young, mature, and old) and they may also be used to discriminate them into specific phenology categories. These biophysical parameters include the oil palms’ leaf area index (LAI), crown projection area (CPA), and height.⁵³ Shadow, roughness, and spectral response are examples of how these parameters are manifested.¹⁰

The concept of LAI is based on the area of one-sided leaf tissue per unit of ground surface.⁸² The basic concept of LAI is to identify the structure of trees by calculating the density of the surface of the leaves in a canopy, which could result in the effective interception of light, air, and water.^46,83 Phenology would increase the LAI, with more fronds and leaves around the crowns of the oil palms.⁸⁴ This relationship thus makes LAI a valuable feature in the estimation of oil palm phenology.⁸⁴ CPA is similar to the percentage of canopy cover but is expressed as the canopy area that is covered by an individual oil palm.¹⁰ To calculate the region of the oil palm’s crown, high-resolution satellite imagery is needed. Satellite data contain different responses in various spectral bands, which make it possible to distinguish, segment, and delineate an oil palm’s crown core, rachis edges, and background.¹ For example, the crown core appears brighter than the edges and background. Oil palm crowns can be delineated using object-based image processing, allowing CPA to be determined. In this area, a positive linear relationship with age is observed with the projected oil palm crown area.¹

The oil palm tree does not stop growing in height throughout its life. This feature of growth is very useful in estimating oil palm age.¹⁹ Throughout its life cycle, the height of the oil palm grows by 30 to 60 cm each year, depending on its physical condition and genetic factors.⁸⁵ Different methods, such as interferometry synthetic aperture radar (InSAR), can be used to obtain information about height.⁸⁶ Using only field-measured information, oil palm height showed a strong linear relationship with age.⁴⁹ Tree height has proven to be a potentially useful indicator for estimating oil palm phenology.⁵³ Another biophysical parameter, diameter at breast height, is insignificantly related to the phenology of oil palms since oil palm trunks do not grow after 2 years of maturity.⁸⁵

Several studies have used biomass to estimate the phenology of oil palms. Thenkabail et al.⁵⁵ used stem height to estimate the aboveground biomass of young oil palm trees (aged from 1 to 5 years). Khalid et al.⁸⁷ used total height as a descriptive variable to estimate biomass for mid-mature (23-year-old) Malaysian oil palm plantations. SAR for oil palm plantations is an important parameter, extracted from normalized backscatter, that is related to AGB. Nordin et al.⁷⁶ used a biomass index and created an empirical biomass model for oil palms. The results of SAR applications in tropical forests can be applied to oil palm plantations. Due to their standardized geometry, the use of the same plant species, and the identical ages of the trees planted in a single plantation, the resulting data are easy to interpret. Kee et al.⁴⁴ investigated oil palm tree discrimination using three parameters—LAI, biomass, and height—to assess the backscattering value of oil palms using data from the C-band and L-band. Lazecky et al.⁸⁸ also analyzed an oil palm plantation based on Sentinel 1 A (C-band) data.

3.3.

Based on Classification Techniques

A classification image algorithm can be used to sort an image into age categories to provide information regarding oil palm plantations.⁵⁵ Oil palm identification has been used successfully in a variety of image classification approaches. Image classification involves categorizing pixels based on different features, such as spectral signatures, indices, contextual details, and many more.⁸⁹ The techniques of image classification can be divided into supervised and unsupervised techniques.⁹⁰ Supervised classification techniques include the use of support vector machine (SVM), random forest (RF), spectral angle mapper (SAM), fuzzy adaptive resonance theory supervised predictive mapping, Mahalanobis distance, radial basis function, decision tree (DT), multilayer perception, naive Bayes, maximum likelihood classifier (MLC), and fuzzy logic analyses.⁵¹

There are numerous classification techniques available, with MLC being the most common parametric classifier due to its good classification performance. Santos and Messina,⁷¹ MLC for the identification of oil palm yields. Tan et al. ⁴⁹ separated the ages of oil palm trees, and Vadivelu et al.⁷² also performed oil palm age classifications. Yayusman and Nagasawa⁷⁷ also classified young and mature oil palms. On the other hand, unsupervised classification techniques include the affinity propagation (AP) cluster algorithm, fuzzy C-means algorithms, the $K$-means algorithm, iterative self-organizing data (ISODATA), etc.⁹¹ Thenkabail et al.⁵⁶ carried out unsupervised ISOCLASS classification. Tan et al.,⁵⁰ Razali et al.,⁹² and Danylo et al.⁷⁵ also used unsupervised classification (ISODATA) on a cloud computing Google Earth Engine (GEE) to reveal the age of oil palms in Southeast Asia. Other researchers have used a combination of classification methodologies to obtain the best level of accuracy. Morel et al.⁶³ effectively distinguished between forest and oil palm areas based on Landsat data using unsupervised $k$-means and maximum likelihood methods with a DT, as well as calculating aboveground biomass using ALOS PALSAR. A recent study performed by Carolita et al.²¹ showed a possible method of mapping oil palm age based on a SPOT-6 normalized vegetation differential index (NDVI) pixel-based classifier.

Object-based image analysis (OBIA) is another method of classifying oil palm phenology. Instead of analyzing individual pixels, OBIA of remote sensing data requires the grouping of pixels into homogeneous segments or objects.⁹³ These segments provide additional information about individual pixels, such as the mean, variance, and mean ratio values per band.⁹³ Accumulating pixels into segments further reduces the computing costs involved with working with finer-resolution satellite imagery, which is becoming more available as sensor quality improves. This type of image segmentation algorithm is classified into four categories: point-, edge-, and region-based, and combined algorithms.⁴⁵ Oil palm mapping can be carried out by combining pixels with identical values, clustering areas, and categorizing them based on texture, context, and geometry.⁴⁵ Rizeei et al.,⁴⁶ for example, were able to estimated oil palm tree counts and perform age predictions using WorldView-3 imagery and LiDAR data, utilizing an optimized OBIA method. Chemura et al.¹ used OBIA to estimate the age of oil palm plantations (years after planting) by integrating high-resolution multispectral remote sensing data and regression techniques.

In recent decades, machine learning classifiers have emerged as powerful classifiers and have been widely adopted for classification purposes due to their higher accuracy and performance compared to MLC.⁹⁴ Analyses using artificial intelligence to identify oil palm land cover, such as SVM, RF, classification and regression tree (CART), deep learning (DL), artificial neural network (ANN), and other machine learning algorithms, have been used in an attempt to obtain significant improvements.⁹⁵ Descals et al.⁹⁶ used cloud computing and the RF classification approach to discriminate between young and mature oil palms based on Sentinel 1 and Sentinel 2 data. Ordway et al.³⁹ also used RF models to classify imagery in 2000 and 2015 according to land cover types: forest, mature oil palms, immature monocultures, and other land cover, based on Landsat data. De Alban et al.⁹⁷ combined Landsat and L-band SAR data to improve land cover classifications using the RF classification method in Google Earth Engine. Shaharum et al.⁷⁵ also used GEE to characterize oil palm land cover for the first time across Peninsular Malaysia, using nonparametric machine learning algorithms such as SVM, CART, and RF, obtaining the best accuracy using SVM, at 93.16%.

To address automation issues, more recent works have used DL in computerized oil palm plantation analyses. For instance, the DL approach has been used in palm tree detection and counting.⁴⁰ The convolutional neural network (CNN), a widely used DL model, has achieved great performance in many studies in the computer vision field, such as image classification. Li et al.³⁸ proposed using DL to detect plants instead of manual detection methods. They used data from a manual count to train and improve the performance of a CNN system. Mubin et al.⁴¹ also used a CNN to distinguish between young and mature oil palm trees.

4.1.

Based on Wavelength

Depending on the wavelength, electromagnetic radiation may be available from a natural source, or it must be transmitted by an instrument. Typical passive sensor systems on space platforms include panchromatic systems, multispectral systems, and hyperspectral systems. A multispectral sensor is a multichannel detector with a few spectral bands.²⁵ Each channel is sensitive to radiation within a narrow wavelength band. Visible light refers to energy in the blue (0.4 to $0.5\mu \mathrm{m}$), green (0.5 to $0.6\mu \mathrm{m}$), and red (0.6 to $0.7\mu \mathrm{m}$) bands of the electromagnetic spectrum.⁹⁸ The middle-infrared region (often referred to as the short-wavelength infrared) includes energy with a wavelength of 1.3 to $3\mu \mathrm{m}$. The thermal infrared region contains light at 3 to $5\mu \mathrm{m}$ and 8 to $14\mu \mathrm{m}$. McMorrow¹⁰ used Landsat TM data to estimated age parameters by analyzing the spectral band and infrared index. Sitoms et al.⁷¹ analyzed a phenology model using the spectral index. Tan et al.⁵⁰ performed comparisons using bands 1, 2, 3; NDVI; and surface reflectance data from Landsat and UK-DMC. For high spatial resolutions, Carolita et al.⁹⁹ succeeded in estimating age using NDVI based on SPOT 6 data.

Active sensors use frequencies from 1 to 90 GHz. The wavelengths of active sensors are intrinsically linked to the penetration capabilities of the transmitted microwave signal, such that longer-wavelength signals can penetrate deeper into vegetation canopies and soils.¹⁰⁰ An active sensor can obtain information in the electromagnetic spectrum bands K (1.1 to 1.7 cm), X (2.4 to 3.8 cm), C (3.8 to 7.5 cm), L (15 to 30 cm), and P (30 to 100 cm),¹⁰¹ with the polarization of horizontal-to-horizontal (HH), vertical-to-vertical (VV), horizontal-to-vertical (HV), or vertical-to-horizontal (VH), which have varying ranges and azimuth resolutions.¹⁰¹ SAR releases electromagnetic radiation in the microwave range of wavelengths and receives backscatter radiation when interacting with objects on the Earth’s surface.²⁹ Backscattering has been used in many studies to describe vegetation dynamics, based on current microwave wavelengths in the X-, C-, and L- bands.²⁷ For example, Tan et al.⁵⁰ applied the gray-level co-occurrence matrix (GLCM) technique to ALOS PALSAR-2 data to determine the ages of oil palm trees. Carolita et al.²¹ developed a regression model based on the correlation backscattering value from SAR Sentinel-1A data (band C) and age information collected from oil palm age-block data. Okarda et al.²¹ discovered that the relationship between age and backscattering has a strong correlation for HH and HV polarized signals, and so the value of backscattering can be useful for analyzing the height and age of oil palm trees.

4.2.

Based on Biophysical Tree Parameters

Plant biophysical parameters can be used to determine the phenological characteristics of oil palms and are also recognized to be related to the growth stages of oil palms (young, mature, and old) and can be categorized into various age classes.^10,11 Assessing and quantifying biophysical parameters of vegetation cover are extremely important problems when monitoring land cover and associated changes, identifying vegetation stress, and assessing yield. These parameters can be used to quantify the state of crops within agriculture monitoring tasks under the global agriculture monitoring initiative, and they have already proved to be efficient for crop yield and production predictions and estimates.¹⁰² Remote sensing data from space are the only source of information that enable the regular and consistent estimation of biophysical parameters at the regional, national, and global scales.¹⁰³ Both optical and SAR imagery can be used to extract biophysical values.

Phenology information is a good indicator for yield prediction as it influences the quality and quantity of the fresh fruit bunches.¹ To determine the phenology, many researchers have investigated biophysical parameters. These parameters are used to identify the biophysical characteristics of oil palms. Oil palm phenology begins at the age of zero; in this stage young oil palms are not much affected by competition before the leaves start to overlap and yields per palm in the first few years may be more or less constant up to quite high densities.¹⁶ Thus, the optimum is very high when production starts, but this falls to below 150 palms/ha within 10 years.⁸² At any one age, quite a wide range of optimal densities has been recorded.¹⁶ Oil palm fruits can be harvested throughout the year. In oil palms older than about 10 years, leaf area does not change much, so the optimum might be expected to remain fairly constant after reaching a minimum $\sim 12$ to 13 years after planting, and the optimum density increases slightly in later years.¹⁰⁴ This has been attributed to better light distribution through the canopy as differences in palm height become more marked, and light penetration through the canopy was found to increase in year 13 from its minimum in years 9 to 11, with the variation in palm height increasing with age.¹⁰⁵ Another factor may be that the LAI is maximal at about 10 years and then falls slightly.

An allometric equation is often used in forestry to relate the measurable parameters of plants (such as age, diameter at breast height, and tree height) to carbon stocks.⁵⁴ An allometric equation can be developed based on a strong statistical relationship between the carbon stocks obtained from the plot-based measurement of certain parameters. Moreover, these parameters provide an important piece of information to complete the allometric equation for the estimation of biomass.⁴⁹ This further indicates the carbon stock of oil palms and its environmental effects.¹⁹ In oil palms, parameters such as age and trunk height are applied in an allometric equation for determining carbon stocks.⁶⁰ For example, Shashikant et al.⁶⁰ used different allometric equations to investigate the relationship between radar backscattering and AGB data on age differences. Kanniah et al.¹⁰⁶ estimated the LAI and fPAR values of an oil palm ecosystem in Malaysia using remote sensing. The data derived from near-infrared signals from UK-DMC 2 and LAI were computed on the basis of hemispherical photos. The model was used to increase the LAI from the plot to the region level. The results revealed a strong relationship between the LAI data and the validation data. The authors discovered that the LAI and fPAR values from UK-DMC 2 and MODIS matched well after comparing these results to MOD15A data.

4.3.

Based on the Type of Regression Model

A wide range of studies have been conducted using different remote sensing approaches to estimate the age of oil palms, which can be grouped into two main categories: radiative transfer models and statistical regression models. The most common method for estimating phenology is regression analysis.¹⁰⁷ A regression analysis or other statistical technique is used to determine a correlation or relationship between two sets of information.¹⁰ Regression models are used to understand how changes in the predictor values are associated with changes in the response mean. There are a variety of regression methodologies based on the type of response variable, the type of model that is required to provide an adequate fit to the data and the estimation method, such as whether it is linear, multiple linear, or nonlinear.¹⁰⁸ The most common models are simple linear and multiple linear models. Nonlinear regression analysis is commonly used for more complicated data sets in which the dependent and independent variables show a nonlinear relationship. There are three purposes for regression analysis: (1) prediction, (2) design of the model, and (3) assessment of parameters.¹⁰⁹ This procedure has been specifically used to analyze the relationship between extracted features and ground-measured biophysical parameters.¹⁰

In the case of oil palm phenology, many studies have used linear and nonlinear regression approaches, such as Nordin et al.⁷⁶ using AIRSAR; Tan et al.,⁵⁰ Avtar et al.,⁷⁹ Darmawan et al.,⁴⁸ and Darmawan et al.⁷ using ALOS PALSAR, who found that the relationship between age and biophisycal parameters presented a nonlinear logarithmic equation, with $R^2$ ranging from 0.4 to 0.8. McMorrow¹⁰ used Landsat data and Okarda et al.²¹ used ALOS PALSAR data, showing a linear regression relationship between age and spectral index equation, with $R^2=0.7$. On the other hand, a nonlinear quadratic equation, with an average $R^2$ value of 0.8 using optical and SAR data, has been demonstrated by Tan et al.,⁵⁰ Carolita et al.,²¹ and Rizeei et al.,⁴⁶ and Darmawan et al.,⁷ using Sentinel-1A data.

McMorrow¹⁹ using Landsat TM data showed a negative nonlinear correlation between the brightness of Landsat TM imagery and the age of the oil palms, meaning that a smaller reflectance value indicated a younger age oil palm. Carolita et al.²¹using SPOT showed that variations in NDVI can describe oil palm growth using the formula $y=-0.0004x^2+0.0107x+0.3912$, where $x$ is the age of the oil palm and $Y$ is the NDVI, with $R^2=0.657$. Comparison of the utilization of Landsat 8 imagery with regression analysis techniques indicates that the age of oil palm plants can be analyzed using spectral index NDVI data with $R^2=0.85$. Chemura et al.¹ also estimated the age of oil palms in Ghana by integrating multispectral data generated from WorldView-2 high-resolution satellite imagery and regression analysis techniques. Correlations between age and CPA can be used to evaluate oil palm age models. According to the results of this study, there is a strong relationship between age and CPA in oil palms over the age of 13 years. Correlations with biomass parameters were observed by Nordin et al.⁷⁷ using AIRSAR data. Different polarizations [horizontal ($H$) and vertical ($V$)] of radar remote sensing data were utilized to investigate the ages of oil palms.⁴⁷ Both studies using the ALOS PALSAR series agreed that (1) HH was statistically related to the age of oil palms better than HV and (2) the backscattering value increased with age due to the increase in biomass.

Shashikant et al.⁵⁹ tested the relationship between radar backscattering and AGB data with certain allometric equations in relation to age variants. Tan et al.⁵⁰ used ALOS PALSAR-2 data to detect the age of oil palm trees using the GLCM technique. Pohl and Loong¹¹⁰ researched the application of InSAR topography and developed high-profile oil palms; this information was useful for mapping the age profile of oil palms. They analyzed the relationship between tree height and the ages of oil palms, applying a linear regression model. This relationship resulted in a height-to-age empirical model to form an equation that can derive age values from the heights of oil palm trees in Malaysia.

The relationship between age and backscattering on mineral soils and peatlands has a strong correlation for HH and HV, as found by Okarda et al.²⁰ According to Okarda and colleagues, the importance of backscattering may be useful in analyzing the height and age of oil palm trees.²⁰ Carolita et al.²¹ produced a regression model with a correlation between backscattering value from SAR Sentinel-1A data (band C) and age obtained from oil palm age-block data. In the C-band, the researchers found that tree cover pixels were considered oil palms if the VV/VH difference was greater than 7.4 dB and VH backscattering was less than $-13\mathrm{dB}$.¹¹¹ However, smaller wavelengths such as the C-band have been used successfully to predict phenology.¹¹¹ Darmawan et al.⁷ aimed to explore the potential analysis of phenology characteristics using SAR data with X-, C-, and L-band measurements. The results showed that the most promising region was the C band, with VV polarization, with $R^2=0.85$.

5.1.

Identification and Mapping of Oil Palm Phenology Based on Remote Sensing

Previous researchers have investigated classification methods to distinguish between oil palm plantations and other land cover elements and have then used these classification methods to distinguish between age groups (Table 1).

Table 1

Oil palm phenology using classification methods.

Many classification techniques in machine learning, such as RF, SVM, and ANN, have improved in terms of accuracy, with error levels around 80% to 90%. Hyperspectral imagery can provide improved levels of detail on oil palms for phenology analyses using biophysical parameters. However, all techniques should be applied when adopting methods for one location and applying them to other locations as there are various factors that affect the accuracy of the classification outcomes, such as scale, phenology, and variations in topography and background signals. There is no single type of data that is suitable for all oil palm areas.¹¹⁵

Machine learning with RF, SVM, and CNN techniques showed higher accuracy than other techniques. Gutiérrez-Vélez and DeFries⁶⁹ used RF classification based on vegetation index with an accuracy of 96.3% using Landsat imagery data. Using the same technique, Lee et al.³⁴ identified immature oil palms based on Landsat data with a cloud computing GEE, showing an accuracy of 91.2%. Amirruddin et al.¹¹⁴ made a comparison of RF and DT based on hyperspectral imagery with accuracies of 98.77% for RF and 90.48% for DT. Sarzynski et al.³¹ used RF based on a fusion of Landsat 8 and ALOS PALSAR-2 techniques, with an accuracy of 84%. A CNN method was developed by Mubin et al.⁴¹ based on WorldView-3 for the identification of young and mature oil palms, with an overall accuracy of 95.11% for young oil palms and 92.96% for mature oil palms. Furthermore, an SVM technique showed an average accuracy of 84.91% for Rizeei et al.⁴⁵ and Shaharum et al.⁷⁶ achieved 93.16% accuracy.

5.2.

Developing Regression Models of Oil Palm Phenology Based on Remote Sensing

Aside from classification methods, one approach to determining the ages of oil palms is through modeling algorithms that use regression to show the correlation between oil palm age parameters based on wavelength and biophysical parameters (Table 2).

Table 2

Oil palm phenology models.

All remote sensing data can be applied in developing oil palm phenology models. The choice depends on the biophysical parameter used for estimating oil palm age. For passive sensors, Landsat imagery is commonly used to extract oil palm parameters using spectral indexes, such as NDVI and a quadratic nonlinear regression model. Furthermore, the use of WorldView imagery was applied by Chemura et al.,¹ using the biophysical parameter CPA, with $R^2=0.88$. Many studies have chosen L-band penetration in active sensors because it provides the ability to detect and measure biophysical parameters with moderate coefficient determination, using aboveground biomass, for example, and the model involves logarithmic nonlinear regression. Nordin et al.,⁷⁷ for example, demonstrated a coefficient determination of 0.8 using AIRSAR data with biophysical parameters relating to biomass. This means that biomass has a strong correlation with standing age. Tan et al.⁵⁰ analyzed the biophysical parameters of height tree and diameter at breast height of oil palms based on ALOS PALSAR data, with $R^2=0.9$. On the other hand, comparisons of Sentinel-1A data with VV polarization conducted by Darmawan et al.⁷ showed quadratic linear regression with a higher $R^2$ value of 0.85. Nonlinear regression was applied for the purpose of better coefficient determination ($R^2$), with quadratic and logarithmic types giving $R^2$ values ranging from 0.5 to 0.8, from moderate to strong correlations.

6.1.

Combining Multisensor Remote Sensing Data

The use of active and passive sensors has proven to be successful in classification applications and algorithm models with high accuracy have been developed. Passive sensors are generally used to analyze oil palm phenology simultaneously at the regional scale and can also be used to identify vegetation indices and biophysical parameters such as LAI and forest canopy density (FCD). Active sensors can also be used for identifying oil palm age using biophysical parameters, such as aboveground biomass, tree height, and canopy height.

Active sensors can only obtain complex data under various transmitting and receiving polarizations and biophysical parameters can only be extracted through regression models or machine learning.¹¹⁷ With the advancement of active sensors (SAR), multipolarization SAR methods are evolving, which involve adaptions to the shape and orientation of vegetation components. According to Izzawati et al.,¹¹⁸ the key variables influencing the accuracy of tree height estimations are canopy shape, tree density, and slope. According to Pohl et al.¹¹⁹ and Tan et al.,⁵⁰ active sensors can provide important information on tree height to complete the list of biophysical parameters used to estimate the productivity of a plantations using polarimetry and interferometric techniques (PolInSAR). These techniques can be used to estimate tree height by subtracting a digital terrain model from an InSAR height estimate or using dual PolInSAR data generated from X-band, C-band, L-band, and P-band data.¹²⁰ PolInSAR seems to be a very effective instrument for ecology and forestry because it can provide highly accurate estimations of canopy height and forest biomass,¹¹⁹ but it has not yet been implemented in oil palm phenology studies.

Furthermore, the use of passive and active data to identify oil palm plantations emphasizes the usefulness of multisensory and data fusion techniques, which combine various data features to improve accuracy.³⁰ According to Rashid et al.,⁹⁴ for oil palm classification, the fusion of data from multisensors is highly recommended. Mohd Najib et al.¹²¹ have also combined multisensor data using ALOS PALSAR-2 and Landsat TM to detect mature and young oil palms and their results showed an accuracy of 98.36% in Peninsular Malaysia. Sarzynski et al.³¹ evaluated the use of a semiautomated approach with an RF method as a classifier and combined optical and radar datasets to classify oil palm land-cover in 2015 in Sumatra, Indonesia, using Google Earth Engine with an overall accuracy of 84%. Danylo et al.¹⁸ used Sentinel 1 to generate an oil palm plantation map, and a Landsat time series was used to analyze an oil palm plantation, first detected using NDVI and BSI. In other plantations, the fusion of data from various remote sensing sources has already been undertaken; for example, optical and SAR analyses of paddy fields were effectively combined for rice identification and mapping,¹¹⁹ with high accuracy.

6.2.

Most Suitable Classification Method and Regression Model

Machine learning approaches have been shown to be able to utilize different kinds of remote sensing data. The recent evolution in artificial intelligence, such as development of DL, promises to bring remote sensing technology to new standards of performance, with improvements in detection and classification.¹²² This information is critical for determining crop health and productivity and serves as an essential measure for agriculture planning and management.¹²³ In relation to oil palm management practices, recent studies have proposed the use of remote sensing technology using various data sources and advanced image processing methods based on machine learning and DL. According to Chong et al.,¹¹ in a general sense, for the discrimination of oil palms and other plantations, classification is carried out by classifying pixels of similar attributes/values depending on the used classifier. Machine learning methods, based on RF classifiers, are very effective in the identification and mapping of oil palms. According to Sarzynski et al.,³¹ RF is a class of DTs that can increase the precision of classifying and detecting plantations, for example, in oil palms, depending on the data they collect. Furthermore, according to Shaharum et al.,⁷⁶ the RF algorithm derived oil palm information better than the other classifier used, and the RF algorithm enhanced classification when the RF trees were ensembled into a forest.⁶⁹ Moreover, according to Attarchi and Gloaguen,¹²⁴ nonparametric RF models can accommodate a large number of correlated input variables and biased data, while avoiding overfitting. As a result, RF models are highly useful for mapping heterogeneous landscapes. In addition, incipient developments in deep machine learning approaches, such as CNNs, are gaining attention due to their promising capability for automatically extracting valuable contextual information from remote sensing data, with potential usefulness in large-scale classifications.⁹⁷ Li et al.³⁸ proposed using DL to detect plants instead of manual detection methods. According to Toh et al.,¹²⁵ DL classification requires a large dataset for optimal classification training. In addition to being very accurate, according to Bonet et al.,¹²⁶ the DL approach, such as the use of CNNs, is very flexible as it can be used straightforwardly to identify similar plants, or those of different varieties, decreasing the number of misclassified oil palms. All these classification techniques show improvements in accuracy up to 90%.

In one study, OBIA was found to be more accurate than pixel-based analysis. Chemura et al.¹ adopted WorldView-2 data to investigate oil palm age using an OBIA classification method to identify the crown area, combined with a regression technique. Their results showed an overall accuracy of 72.8%, a segmentation goodness of fit of 0.69, and a correlation with field-measured crown area of 0.9, which indicated the successful delineation of individual tree crowns using OBIA.

The modeling of oil palm phenology based on regression models does not always display linear behavior. Oil palm phenology normally exhibits a nonlinear type of relationship.¹²⁷ Nonlinear machine learning techniques are a type of highly efficient regression approaches that have been used successfully in the analysis of biophysical parameters. According to Frost,¹²⁸ nonlinear regression also requires a continuous dependent variable, but it provides a greater flexibility in the fitting of curves than linear regression. According to Chong et al.,¹¹ the biophysical characteristics of the tree have also proved to be potentially reliable indicators for the estimation of oil palm age. Most nonlinear models have one continuous independent variable, but it is possible to have more than one.¹²⁹ According to Ali et al.,¹³⁰ such techniques can learn and estimate even complicated nonlinear mappings, using advanced machine learning methods and the information contained in a series of comparison samples. ANNs are among the most frequently used techniques in the field of biophysical parameter retrieval, and these have been widely investigated in many application domains.

Recent studies have involved the use of ANNs in oil palm studies. Hilal et al.¹²⁷ showed that ANN was a significant modeling instrument, showed terrific potential for the optimization of prediction models, with high precision in many related areas of study. All regression models can show improved coefficient determination ($R^2$) values, ranging from moderate (0.4) to strong (0.9) relationships. Furthermore, machine learning techniques do not always produce high-precision results because good results depend on the machine-learning model set-up, the training samples, and the input parameters.¹²⁷ In addition, in the case of paddy fields, the best classification methods are SVM and RF, and CNN and recurrent neural network techniques will require phenology and textural feature information to improve the accuracy of their data.¹²⁶

6.3.

Future Direction in Oil Palm Phenology Based on Remote Sensing

According to Refs. 106 and 131, LAI can identify phenology based on optical sensor spectral values derived from NDVI. According to Refs. 45, 95, and 125, machine learning classifiers (RF, SVM, DT) have emerged as effective classifications and have been extensively utilized for classification applications due to their better accuracy and performance. According to Refs. 95, 127, and 132, machine learning-based regression algorithms such as RF, CNN, and SVR could be very useful for predicting oil palm production including phenology. Furthermore, rather than a single algorithm, an ensemble of numerous algorithms should be considered to improve the robustness of the prediction model. Sarzynski et al.³⁰ demonstrated improved accuracy based on machine learning RF classification in the analysis of oil palm plantations using multisensor and data fusion techniques. Descals et al.¹³³ used a DL CNN for semantic segmentation with 98% overall accuracy. According to Ref. 132, the detection and classification of oil palm trees using DL need to be performed using high-resolution images and requires field survey data to study the important biophysical characteristics of oil palm trees and to evaluate the reliability of the results.

According to Refs. 81, 103, 119, 125, 131, and 134, regression models of oil palm phenology must be created to be developed based on biophysical tree parameters, researchers can make estimates about oil palm phenology by analyzing the canopy layer, fruit bunch, and tree height using allometric techniques. Regression model is used to analyze the relationship between biophysical characteristics extracted from an image with data from measurements in the field. Regression model-based algorithms using machine learning could be very effective in making predictions about oil palm phenology.

In the oil palm industry, UAV-based imaging provides low cost flexible data acquisition with less weather constraints and higher spatial/temporal resolution, as compared with high-resolution satellite data.⁸¹ There are various applications of UAV, such as monitoring canopy structure and condition, mapping biomass, and precision agriculture.¹³⁵ Yarak et al.¹³² applied high-resolution imagery from a UAV, combined with the DL approach, for the automatic detection and health classification of oil palm trees using three important biophysical characteristics: crown size, crown color, and crown density. According to Pohl,⁶⁹ classification of polarimetric and interferometric (PolInSAR) data to estimate heights (biomass) information and to provide information on oil palm frond shapes and changes as well as moisture parameters of plants and soil.

In the future, remote sensing technology is expected to evolve into high-tech industries, in which systems are coupled with artificial intelligence and big data¹³⁶ to examine the relationship between environmental and socioeconomic indicators.¹³⁷ For future direction, we recommend some parameters, remote sensing data sources, and methods in Table 3. According to Table 3, we have recommendation for further studies on oil palm phenology based on remote sensing to identify, classify, and mapping can be use parameter such as LAI, canopy density, and NDVI derived from Landsat, high resolution images, and UAV LiDAR based on machine learning such as RF, SVM, DT, and OBIA method. For developing regression model can be use LAI, canopy height, biomass derived from SAR, PolInSAR, and UAV LiDAR based on nonlinear regression using machine learning method. This recommendation, we believe will be more accurate for identification, classification, and mapping, also developing regression model. However, the accuracy is influenced by testing and validation data, climate factors, precision errors, soil characteristics, spatial mixtures, and other external factors.

Table 3

Recommendation for further studies on oil palm phenology based on remote sensing.