Rice brown planthopper monitoring and detection by spectral reflectance: a review

Brown planthopper (BPH) has been one of the main pests of rice worldwide. Monitoring is important factor for determining attacks and estimating their effects. The traditional monitoring approach is usually conducted through visual observation and field scouting, with limitations such as subjectivity and time consumption. Remote sensing is an alternative pest monitoring method that covers a larger area in a shorter time. This paper discusses a remote-sensing method that uses a spectral approach to detect BPH attacks. Literature was filtered and processed using the PRISMA method. According to the spectral sensor, studies were classified into multispectral and hyperspectral sensors. Based on this scale, there are four studies on the panicle, leaf, canopy, and field levels. The model used single-wave reflectance and spectral indices as predictors. Various algorithms were used in the studies: linear regression, Principal Component Analysis, and Machine Learning to estimate the severity class, BPH Population density, and yield loss. A combination of spectral reflectance with other parameters, such as weather, fertilizer application, and infestation time, was conducted to improve the performance of the detection model. This review provides state-of-the-art spectral reflectance usage for detecting BPH attacks and opportunities for future development.


Introduction
Brown planthopper/Nilaparvata lugens (BPH) has been one of the most common pests of rice, especially in the Pacific and Asia Regions [1].The economic loss as the effect of its attack is huge.According to [2], for Indonesia itself, the economic loss in 2020 was more than IDR 39 Billion or USD 250 thousand.
Monitoring is important in determining attack and estimating the effect [2,3].The traditional monitoring approach is usually conducted by visual observation and field scouting with limitations, such as subjectivity and time consumption.Remote sensing is an alternative method to carry out pest monitoring to cover a larger area in a shorter time besides potential attack area analysis based on climate [4].Spectral analysis is one remote sensing method that can detect insect pest attacks, including BPH.BPH sucks rice sap by inserting its stylet into the vascular tissue of a rice stem [5].It induces changes in leaf chlorophyll and relative water content and can be observed by changes in spectral reflectance.
Some papers review the usage of remote sensing in detecting pests, as in [6,7], but the review is still general.According to a literature search, there has yet to be a single paper that reviews the use of remote 1230 (2023) 012088 IOP Publishing doi:10.1088/1755-1315/1230/1/012088 2 sensing to monitor BPH, especially by using spectral analysis.This paper reviews the using of spectral reflectance analysis in detecting and monitoring BPH attacks.

Material and Methods
In this research, the authors used the Preferred Reporting Items for Systematic Reviews and Meta-Analyses (PRISMA) [8] approach to search and select sample literature without meta-analysis.PRISMA has 3 stages: (1) identification; (2) screening; (3) includes.
Search conducted using defined keywords using PICO (Populations, Interventions, Comparators, Outcomes).The scope of the review discusses the usage of spectral information in detecting the attack of BPH in rice.A review has been done to answer primary and secondary review questions.The main review question is as follows: 1. How is spectral information used in detecting BPH attacks?Whereas secondary review questions are as follows: 1. How is the spectral information used in BPH attack detection extracted?2. What are spectral features used to detect BPH attacks? 3. What are spectral-based BPH attack detection models algorithm used existing and how the performance improved?4. What are the challenges and opportunities of future studies on this topic?Searching criteria were determined before searching, so the articles that would be reviewed were relevant to the aim of this research.Criteria articles included in this review were as follows: 1.
Study spectral-based BPH attack detection 2.
published in English on the Scopus database 3.
Published during the previous 20 years, i.e., 2001-2021, to ensure that the data is up-to-date and consistent with contemporary trends.The first stage PRISMA method was to do an identification library (identification) on the available database.We put together a list of articles from the Scopus database on September 23, 2022, using the string that was arranged based on the theme from the scope of the review formulated via PICO.The strings used in the search article are as follows: TITLE-ABS-KEY ((Nilaparvata OR planthopper) AND ( *spectral* OR reflectan* OR satellite* ) ) The initial screening stage was to filter the title, abstract and keywords reviewed based on criteria search that had been determined before.By keyword, at the beginning stage, the search obtained as many as 49 articles indexed by Scopus.The second screening stage was the feasibility evaluation, carried out by evaluating the remaining articles related to the topics studied by searching abstract and complete text articles (full text).Articles at the included stage differ among the number of published articles and the amount of passed article evaluation feasibility [9].Researchers consider this a suitable approach to produce a review of the strong, transparent, and extensive literature that can give interesting responses to destination research [10].A total of one article out of 20 were excluded during the full-text review because it did not focus on using spectral analysis to detect BPH attacks directly.Remaining articles were read one by one and done tabulation as well as discussion by objectives and review questions in the section scope.

Studies overview
Based on the article selection process, 20 articles were found with the distribution of the year of publication as shown in Figure 2.

Sensitive wavelength determination algorithm
In the studies that utilized a single or specific wavelength in their modeling, the spectral data measured using the spectrometer was first filtered to obtain the most sensitive wavelengths.There were various filtering algorithms, the earliest of which was to measure the correlation between the reflectance of each wavelength directly or the first and second derivatives of the reflectance curve with the level of attack as in [11][12][13][14][15]. Correlations were also measured against the cumulative reflectance of a given wavelength range.The use of the correlation method is considered to have a weakness because reflectance readings between waves in a measurement have Interco-linearity.Therefore, some subsequent studies also involved principal component analysis (PCA) in determining the sensitive wavelength for the onset of BPH [16,17].Greenness Index GI/GDR/RI(GREEN,RED) GREEN/RED [12,21,23,30] Enhanced Vegetation Index EVI 2.5((NIR-RED)/(NIR+6*RED-7.5*BLUE+1)) [17,22,25] Green NDVI GNDVI (NIR-GREEN)/(NIR+GREEN) [12,13,21] Zhou et al. [15] found that features such as Crest Amplitude, Trough Amplitude, Difference between the Crest and the Trough, Ratio of the Trough to the Trough, Edge Amplitude, Peak Area of the 1st Derivation, and Peak Abruptness of the 1st Derivation had significant (p-value<0.01)for distinguishing infested and non-infested canopies.However, processing these features is complex and difficult to transform in multispectral imagery, so further studies have yet to be conducted.
In addition to using raw reflectance as input for attack modeling, spectral indices were also used.At least 49 spectral indices were used by the 20 selected articles, a combination of several reflectances from hyperspectral or multispectral sensor readings.The combinations include difference, simple ratio, normalized difference, and a certain formulation involving reflectance from two or more wavelengths.The top 6 most used spectral indices can be seen in Table 2.
NDVI was the most used index in both hyperspectral and multispectral sensor types.However, there are variations in the spectral index calculation formula due to the availability of bands on the sensor.Despite having a common formula, the calculation and classification result potentially reduce the model's accuracy.The general formula of NDVI is the normalized reflectance of the RED and NIR wavebands.In hyperspectral sensors, the RED value is defined as the minimum reflectance value in the 640-740 nm range, while the NIR is obtained from the peak of the reflectance curve at wavelengths 740-1300 nm [12,17] while in multispectral imagery, RED and NIR are defined as bands that represent the red and nir wavelength ranges.In multispectral images, the representative bands have different ranges.For example, the RED value in the SPOT-5 satellite band ranges from 610-680 nm [26][27][28] while in PlanetScope in the range of 590-670 nm [23].The Terra MODIS satellite even has a dedicated NDVI channel in its imagery [24,25].Thus, despite the common denominator of NDVI, there can be differences in the values from each study modeling BPH infestation detection.In addition, this index is also often used for other purposes such as identification of land cover, plant species [32], plant phenology development, [33] other biotic and abiotic stresses [34][35][36][37], and crop yield estimation [38].
Several experiments had successfully developed specific indices to model BPH infestation levels from in situ reflectance measurements at the greenhouse [19] and field scales [13] named BPHI 1-3.Although both models still had predictive performance below NDVI, the indices were expected to detect BPH specifically.

Model algorithm
The models produced by the studies had been able to model the damage level [12,13,19,20], BPH population [11,24,25,27,29,31], distinguish between healthy and infested plants [17,22,23,28,39], and estimate crop losses [21,30].Generally, research using satellite imagery can only model the level of infestation in a little detail as the use of terrestrial sensors.Due to satellite imagery's limitations on spatial resolution, the images read were heavily contaminated by non-crop factors, especially in images taken in the early stages of cropping.These factors can be more easily separated in terrestrial spectral image observations, especially in laboratory-scale experiments.One of the efforts to bridge the model of in situ and satellite measurement results was carried out by Liu et al. [17], simulating in situ measurements to Landsat-7 imagery and getting good classification results of infested plants.
The damage level modeling is generally classified following the INGER guidelines [40].Spectral index values or other spectral features are linearly regressed with the attack class [12,13,19,20].R 2 and RMSE test the performance of these statistical regression-based models.The use of linear regression in classification has the risk of overfitting because the predicted factor is class / categorical.It is reflected in the R 2 values of some of these models, which are very high, above 95%.Therefore, the later studies involved machine learning algorithms such as Probabilistic Neural Network, Support Vector, and Random Forest.These were considered to reflect the category of plant condition better, although only able to model infested and non-infested plants at best [17,23,24,39].Meanwhile, studies that modeled populations also used logistic regression [24,25].

Improvement of model performance
Some studies use additional parameters to improve detection performance.Weather/climate parameters from terrestrial, satellite, or combined terrestrial stations are regression and machine-learning modeling variables.Crop treatments in the field, such as nitrogen fertilizer application and duration of pest exposure, are also known to influence the success of the modeling [29,31].
Ensuring that the rice field is under cultivation is also an effort to improve model performance and prevent detection errors.Ensuring that land is planted with rice is done by using inundation 1230 (2023) 012088 IOP Publishing doi:10.1088/1755-1315/1230/1/0120886 characteristics at the beginning of planting that can be observed through spectral indices that indicate water bodies such as Normalize Different Water Index (NDWI) [18].This method is constrained by the availability of data when cloud cover is high, so observations using radar satellite imagery (Sentinel-1) are used as an alternative [23].

Challenges and opportunities in future studies
One of the constraints of monitoring based on satellite optical imagery is the presence of cloud cover.Cloud cover at the observed location causes the observation coverage to be limited or empty for some time.It is not comforting in the preparation of time series observations.This obstacle greatly affects the sustainability of observations, especially in tropical areas where high levels of cloud cover coincide with the rainy season when farmers start planting rice.It is also the time of year when farmers are prone to attacks by various plant-disrupting organisms.This obstacle has been tried to overcome using statistical methods to cover dates that are empty of observations, such as the linear interpolation method [18,23].Linear interpolation has the disadvantage that patterns will be lost when the gap/window between two available images is shorter.Therefore, other methods, such as the ARIMA method, must be considered.
Distinguishing between BPH infestation and other stresses (biotic or abiotic) is also a challenge considering the pattern of crop damage detected by spectral imagery between stresses is similar.Detection errors can occur partly due to these similarities.In addition, in one location, there is rarely a homogeneous attack by one type of pest organism alone but a combination of several plant pest organisms.Additional information or modifications to the algorithm can be made.The two-step algorithm of confirming the presence of new crops and then analyzing the spectral images [23] can prevent misreading fallow or no-crop stadia as infestations.The additional step of filtering environmental compatibility, especially climate, with specific pest organisms or other plant pest attack characteristics is thought to improve detection accuracy more specifically, as done by Gosh et al. [41].The climatic data is expected to give more information through this method rather than simply using it as a direct predictor parameter in the model as in the previous studies.

Conclusion
This review shows the development and variation of spectral analysis to detect and monitor BPH attacks.The early studies focused on in situ observation and statistical methods, while the current studies also involved machine learning and satellite imagery.The hyperspectral sensor-based analysis were using direct specific wavelength reflectance and some spectral indices, while the multispectral one prefer usage of indices.The performance of the models were Some potential developments in spectral analysis to detect BPH attacks include overcoming missing data due to cloud cover in satellite data and distinguishing spectral readings of BPH attacks from abnormalities due to other stresses.One suggestion is to develop a multi-step algorithm involving characteristics of other stresses based on climate data.

Figure 1 .
Figure 1.Selection scheme articles that used for review

Figure 2 .
Figure 2. Number of articles by year of publication

Figure 3 .
Figure 3.The most common types of sensors

Table 1 .
PICO definitions and keywords used in search article

Table 2 .
Top spectral indices used in BPH attack modeling