Use of time series normalized difference vegetation index (NDVI) to monitor fall armyworm (Spodoptera frugiperda) damage on maize production systems in Africa

Abstract Fall armyworm (FAW) Spodoptera frugiperda (J.E. Smith), damage was monitored at a regional scale using time series data in Western and Southern African countries. The study employed the normalized difference vegetation index (NDVI) computed from Landsat 8 imagery using the Google Earth Engine (GEE) using image composites for the years 2013 to 2020 for the study areas. The index was then reclassified based on the NDVI threshold values into low, sparse, moderate, and dense classes. FAW prevalence data were then utilized to validate the correlation between the FAW infestation and NDVI values. FAW was associated with a decrease in vegetation productivity between the years 2016, 2017, and 2018 when the pest infestation was reported in the study areas. The validation results showed that there is a correlation between FAW infestation and NDVI (R20.83). Our study highlighted that NDVI can be used as a proxy to quantify pest damage to vegetation productivity.


Introduction
Fall armyworm (FAW), Spodoptera frugiperda (J.E.Smith) (Lepidoptera, Noctuidae), is endemic to the Americas, but following its first detection outside of its native range in 2016 (Goergen et al. 2016), it has rapidly spread across different continents (Sharanabasappa et al. 2018;Naeem-Ullah et al. 2019;Sun et al. 2021;IPPC 2022), thereby becoming a true global pest of many important crops, among them chiefly maize and sorghum.Studies have attributed this rapid spread to recent increases in global temperature which provide suitable environmental conditions to the pest (D ıaz-Alvarez et al. 2020; Nurzannah et al. 2020).Commonly, FAW thrives under tropical humid conditions to complete its development (Du Plessis et al. 2020).In Africa, the pest was first detected in Nigeria in 2016, followed by São Tom e and Pr ıncipe, Benin, Ghana, Togo and subsequently also in the Democratic Republic of Congo (DRC) (Goergen et al. 2016).To date FAW has been recorded in 47 out of the 54 African countries (Cabi 2022), driven by the strong dispersal ability of the adult moths which are capable of travelling up to 70 km per day (Ge et al. 2021).In 2018, FAW was first reported in Asia, i.e.India and Yemen, and by 2019 had widely spread with records from China, Malaysia, Indonesia, Pakistan, Japan, South Korea and Australia (Harrison et al. 2019).
Being an invasive species that mostly damages cereal crops, FAW has negatively impacted food and nutrition security, particularly among small scale farmers in Africa (Cock et al. 2017).For instance estimates in 12 African counties indicate that, maize yields have been reduced by 8.3 to 20.6 million tons annually following the FAW invasion (Day et al. 2017).Maize is one of the most important staple food crops in sub-Saharan Africa (SSA), covering approximately 36 million hectares and supporting the livelihoods for about 208 million people on the continent (Day et al. 2017;Assefa and Ayalew 2019).The economic losses by FAW in 12 major maize-growing countries of Africa are estimated at US$2.5 to 6.2 billion (Shylesha et al. 2018).
Earlier studies have demonstrated the use of various geo-spatial and remote sensing tools for effective monitoring of important crop pests like FAW and plant diseases (Acharya and Thapa 2015;Rani et al. 2018;Roberts et al. 2021;Rano et al. 2022).Some of these previous studies predicted habitat suitability of FAW using various modelling approaches at global scale (Tepa-Yotto et al. 2021;Guimapi et al. 2022;Paudel et al. 2022).However, the potential crop damage caused by FAW at finer scales remains to assessed and quantified by means of remote sensing technologies.Remotely sensed data, acquired either with handheld (Filho et al. 2020) or airborne (Bhattarai et al. 2019) sensors, have been used for assessing insect crop damage.Essentially, symptoms of crop damage by insect pests can be visual and/or invisible depending on the attack stage.To develop early warning tools using satellite systems for early insect pest damage detection, there are several sensors such as WorldView-3 that provide imagery of high (<3 m) spatial resolutions.However, to test such operational early warning systems, readily available satellite data should be utilized.Landsat 8, which operates with a multispectral sensor of 30 m spatial resolution, could provide unique data for FAW damage detection at local and regional scales.This could be achieved by integrating Landsat 8 based spectral features and geo-spatial analytics to characterize landscape structure in relation to FAW damage.This approach would enable the assessment of the spatio-temporal FAW damage status in crops like maize and sorghum, which are among the most preferred host plants of the pest (Day et al. 2017).Specifically, time series optical Sentinel-2 spectral bands and vegetation indices can provide a robust, efficient and cost-effective technique for monitoring FAW damage (Horelu et al. 2015;Prabhakar et al. 2020).Though a Sentinel-2 multispectral image has a high spatial resolution of 10 m it's, however, not suitable for long term time series analysis compared to Landsat images due to image availability since the satellite was launched in 2015, hence lacking historical images.
One of the most commonly used differential vegetation index for crop health monitoring is the normalized difference vegetation index (NDVI) (Tucker and Sellers 1986).Generally, NDVI is an indicator of crop vigor, stress and damage as a function of the reflectance at the near infrared (NIR) and red bands (Wall et al. 2008;Huang et al. 2014;Ji et al. 2021).For instance, Prabhakar et al. (2020) used NDVI calculated from Sentinel-2 data to show the viability of site-specific pest management and found that NDVI is quite effective for detecting FAW infestations with an R 2 of 0.80.In another study, Zhang et al. (2016) developed a mapping approach for tracking fall armyworm (FAW) infestation at a regional scale using a univariate model and two-date HJ-CCD (Huanjing satellite charge coupled device) NDVI.HJ-CCD is also an optical multispectral sensor of 30 m spatial resolution which generates images comparable to Landsat 8, however such images are not available in google earth engine software.This method was found to be effective and promising for use in a typical FAW outbreak event in Tangshan, Hebei Province, China.Furthermore, Zhang et al. (2016) expressed the need for robust and cost-effective remote sensing tools for monitoring FAW damage at a regional scale.Thus, this research aims to create a robust and cost-effective remote sensing instrument for FAW damage on a regional scale in Africa.The study fills the research gap as it utilizes Landsat 8 time series data for pre-and post-FAW invasion period at a regional scale, unlike previous studies that have used Sentinel-2 or other remote sensing technologies for crop damage detection at a site-specific/farm level (Horelu et al. 2015;Prabhakar et al. 2021).The objective of this study is to use multi-date Landsat 8 images and NDVI datasets to evaluate the spatial and temporal status of FAW damage in maize growing areas to cover the pre-and post-FAW infestation to evaluate the pest damage.Additionally, the study seeks to measure the crop damage caused by FAW in both West and South African study areas.The Landsat 8 image was considered for the study due to its narrower spectral bands, improved calibration and signal-to-noise characteristics, higher 12-bit radiometric resolution, and more precise geometry (Roy et al. 2016).

Study area
This study was focused on the following areas: Atlantique region of Benin, Centrale in Togo, Salima in Malawi, Chongwe in Zambia, and Limpopo in South Africa (Figure 1).These study areas were selected based on the presence of maize and sorghum cropping systems (Table 1) in the respective areas and their history of FAW attack.Atlantique is one of the 12 departments in Benin, located between latitude 6 20 0 00 00 N to 6 50 0 00 00 N and longitude 2 0 0 00 00 E to 2 20 0 00 00 E in the south of the country.It covers an area of 3,233 km 2 with an estimated population of 1.397 million, annual precipitation of 1,200mm and an annual temperature range between 28 and 32 C. Approximately 60% of its population depend on agriculture for food and income (Atindogbe et al. 2012).
The Centrale region of Togo is located between latitude 8 0 0 00 00 N to 9 0 0 00 00 N and longitude 1 0 0 00 00 E to 1 0 0 00 00 E with an area of 13,420 km 2 , annual rainfall between 1,200-1,500mm and an annual temperature range between 20 and 39 C. The main economic activity within the study site is agriculture with a population totaling 617,871 as of the 2010 census (Ataba et al. 2020).The Salima district of Malawi is located between latitude 13 30 0 00 00 S to 14 0 0 00 00 S and longitude 34 10 0 00 00 E to 34 40 0 00 00 E, covers an area of 2,196 km 2 , with a population of 478,368 (Government of Malawi, 2018), and a mean annual precipitation of 1,188mm as unimodal rainfall.With an average annual temperature of 22 C, the district has a tropical climate.The district is dominated by small holder agriculture, covering approximately 107,400 hectares (Musa et al. 2018).
The Chongwe district in the Lusaka province of Zambia is located between latitude 15 0 0 00 00 S to 16 0 0 00 00 S and longitude 28 0 0 00 00 E to 30 0 0 00 00 E, covers an area of 8,669 km 2 , with a population of 192,303, and precipitation ranges between 800-1,000mm with an annual temperature of 20.4 C (Jacqueline and Mubanga 2020) .Agriculture is the most important economic activity, primarily with crops like maize, cassava, and sweet potatoes (CSO, 2003).Finally, the Limpopo Province of South Africa lies between latitude 23 0 0 00 00 S to 25 0 0 00 00 S and longitude 27 0 0 00 00 E to 31 0 0 00 00 E, covering an area of 125,754 km 2 , an estimated population of 5,778,400, an annual temperature ranging between 24.6 À 28.2 C, 500 mm of total annual rainfall, with commercial farmers occupying >70% of the province's prime agricultural land, while smallholder farmers occupying the remaining arable surface (Cai et al. 2017).

Data description
Landsat 8 satellite data Landsat 8 multispectral top of the atmosphere images (LANDSAT/LC08/C01/T1_TOA) were utilized for the analysis.The Landsat 8 has a spatial resolution of 30 m with 11 spectral bands distributed between the visible, near-infrared (NIR), shortwave infrared (SWIR) and thermal infrared (TIRS) regions of the electromagnetic spectrum (EMS).The sensor has an altitude of 705 km and a temporal resolution (revisit time) of 16 days (USGS, 2021).

Data analysis
The methodology framework of this study consisted of four main steps (Figure 2).
Table 1.Cropping Surface of Maize and Sorghum for the study area in hectares (ha).

Region, Country
Maize cropping surface ha Sorghum cropping surface ha Step 1: Landsat 8 (Table 2) images were obtained and processed using Google Earth Engine (GEE) software (https://code.earthengine.google.com),which is a cloud computing software (Kumar and Mutanga 2018).The image processing involved cloud masking of which band quality band (BQA) was utilized, also atmospheric correction was utilized to remove the effects of atmospheric scattering and absorption on the images, as these factors can affect the NDVI values.A total of 415, 613, 909, 320, and 517 images were used for the Atlantique, Centrale, Chongwe, Salima, and Limpopo regions, respectively.
A composite image was then computed using the median values of the pixels for each year, covering the months January to December of 2013 to 2020 for the selected study  Table 2 lists the spectral bands of a satellite sensor, along with their band numbers, names, wavelength ranges in micrometers (lm), and spatial resolutions in meters.The sensor has eight multispectral bands (bands 1-7 and 9) and one panchromatic band (band 8).The multispectral bands cover a wide range of wavelengths from the visible to the near-infrared (NIR) and shortwave infrared (SWIR) regions of the electromagnetic spectrum.The spatial resolution of the multispectral bands is 30 meters, while the panchromatic band has a higher resolution of 15 meters.Additionally, the sensor has a Cirrus band (band 9) that is sensitive to high-altitude clouds.
regions to compute the NDVI using the Red (band 4) and NIR (band 5) bands (Table 1) of Landsat images (Equation 1).
where NIR and red are reflectance values at near-infrared and red bands 5 and 4 of Landsat 8, respectively.
Step 2: The NDVI values were reclassified into four classes, i.e.Low, sparse, moderate, and dense vegetation, using the reclassify tool in QGIS (QGIS.org,QGIS 2020) to quantify FAW crop damage in terms of vegetation change.The sparse and moderate vegetation productivity comprised of bare soil and grassland classes, while the dense vegetation entailed annual cropland, perennial cropland, forestland and woodlands (Akbar et al. 2019).The natural breaks threshold technique (Singh and Javeed 2021) was then used to reclassify the NDVI values into the four vegetation classes.The method was used to find class breaks that best group standard deviation data with comparable values while also maximizing the difference between classes (Table 3).
The NDVI threshold values listed for each vegetation class in the study regions are based on empirical observations and previous research.NDVI values provide an estimation of vegetation density and vigor, with higher values indicating denser and healthier vegetation cover.The threshold values are used to classify different vegetation classes based on their NDVI ranges.These values can vary depending on the specific region, climate, and vegetation types.Therefore, the values listed for each study region and vegetation class have been determined through previous research and field observations to best fit the local vegetation patterns and characteristics (Singh and Javeed 2021).
Step 3: The specific year of which the FAW was identified for each study site was obtained from literature.(Goergen et al. 2016;Koffi et al. 2020;Dassou et al. 2021).The vegetation density class for each study region for the period before and after the introduction of FAW was then computed using mulusce plugins in QGIS software, an algorithm subtracting annual NDVI values for pre and post FAW infestation expressed in hectares.The 2013 annual average NDVI values were used as a reference to quantify subsequent crop damage, since it was a year with high NDVI values in many of the study regions during pre-FAW introduction.
Step 4: The validation data were obtained from the Food and Agricultural Organization of the United Nations (FAO) collected in the selected study sites across Africa (FAO 2018).We utilized FAW data for the year 2018/2019 (January -December) for which geographic location together with field collected information about different crop types were available.Hence, the points were filtered to focus on maize farm field samples.The original excel data were then converted into a shapefile in QGIS software.The field data points were later overlaid on the NDVI raster data sets, of which the NDVI values were extracted for each point of FAW prevalence, using the extract values by points tool in QGIS.Linear regression was then applied to validate the correlation between NDVI values and the prevalence of FAW that was computed using the scouting data obtained from the field that signified the presence of FAW by rating the proportion of plants affected by the pest.

NDVI classification
Figure 3 highlights the variation in the classified NDVI images.In Atlantique, Benin (Figure 3a), there was a higher sparse and moderate vegetation density in 2013 than the densely vegetated class in 2016.In the other years, there was no consistent overall trend in vegetation density with the dense vegetation class more dominant in the region in 2014 and 2019.In the Centrale region of Togo (Figure 3b), the density of vegetation was decreasing over the years with density values in 2013 being considerably higher than in 2020.In Salima, Malawi (Figure 3c), the vegetation density was lower in 2017 than in 2013.A similar trend was observed in Limpopo (South Africa) (Figure 3d), and also in Chongwe (Zambia) (Figure 3e), the vegetation cover in 2018 was lower than in 2013.

Average annual NDVI time-series
In 2016, the variation in the yearly average NDVI for Atlantique (a) and Centrale (b) was minimal in both areas, as seen in Figure 4 The lowest average NDVI for Chongwe (c) was reported in 2014, then climbed in 2015 before continuing to decline until 2020.The lowest NDVI values were recorded in Salima (d) and Limpopo (e) in 2017 and 2018, respectively.

Area of change for the NDVI classes
We observed a decrease in moderate and dense vegetation classes by 4% and 19%, respectively, for the Atlantique (Benin) and Centrale (Togo) regions in 2016, and for the latter country also an increase in sparse (4%) and low vegetation (19%) (Tables 4 and 5).For Chongwe (Zambia) and Salima (Malawi) we noted a decrease in moderate and dense vegetation classes (Tables 6 and 7), and for Limpopo (South Africa) a decrease in dense, moderate and sparse vegetation between 2013 and 2018 (Table 8).

Validation
There was a significant (P, 0.05) positive correlation between the NDVI values and FAW prevalence with a strong R 2 of 0.83 (Figure 5), implying a higher FAW prevalence with lower NDVI values and vice versa.

Discussion
The study could conclusively show that NDVI can be used as a proxy for monitoring FAW induced crop damage.With multi-date NDVI data we could detect changes in vegetation density from before to after the FAW invasion in the study regions.For each of the latter, we used the year 2013 as the pre-FAW infestation reference year and hypothesized that after FAW was introduced, i.e. 2016 for Benin and Togo, 2017 for Malawi and Zambia, and 2018 for South Africa, drops in NDVI could be attributed to the impact of the pest (Goergen et al. 2016;Abrahams et al. 2017;Murray et al. 2019;Koffi et al. 2020;Dassou et al. 2021;Kasoma et al. 2021).This is consistent with other studies that utilized spectral vegetation indices such as NDVI to monitor crop health and vigor (Wall et al. 2008;Huang et al. 2014;Ji et al. 2021).For instance, Prabhakar et al. (2020) used NDVI to detect FAW damage in sorghum.The authors utilized Sentinel-2 derived NDVI to detect FAW infestation at farm level in southern India, and could correlate pest infestation with NDVI at an R 2 of 0.80 very similar to the R 2 0.83 in our study.Like other noctuid pests, FAW damages maize and sorghum crops and cause defoliation and stem damage, resulting in substantial yield losses (FAO 2018).These crop morphological and physiological symptoms can cause considerable changes in the crop spectral features, particularly in the red and NIR portions of the EMS.The red and NIR portion of the EMS, which is used to calculate the NDVI, are known chlorophyl and water sensitive features that can mimic the changes in the crop chlorophyl and water content due to pest damage (Yamanura and Patil 2021).
In Zambia, a 40% decline in NDVI magnitude was reported during the 2016/2017 cropping season and continued to decline until 2018 (Kansiime et al. 2019).This is corroborated with our findings, which show that the frequency of NDVI values decreased from 2017 to 2020.It is highly likely, that the massive FAW outbreaks in 2016/2017 in Zambia, with >124,000 ha of maize affected, could have been the reason for the observed NDVI reductions (Kansiime et al. 2019).Also, the level of NDVI reductions can vary depending on the FAW intervention measures applied.For instance, in Zambia, the NDVI values continued to decline following the 2016 invasion until 2020, possibly because of shortcomings in information dissemination to affected farmers, and inadequate control methods coupled with a lack of reliable access to insecticides (Kasoma et al. 2021).Yet, in countries like Benin, Togo, and Malawi, NDVI values increased after the FAW invasion, possibly because of more successful pest control intervention utilizing synthetic and biological insecticides which led to a considerable drop in FAW population levels (CABI 2018).Similar observations were made in Togo, with FAW levels decreasing from 68.5% to 55. 8% and 17.8% in 2016, 2017, and 2018, respectively, due to the increased use of synthetic insecticides (Koffi et al. 2020).
Nonetheless, it is worth noting that NDVI is not only affected by insect pest damages but also by other biotic and abiotic factors such as crop diseases, weeds, drought, flood, nutrient deficiencies etc. (Pei et al. 2019).For instance, the NDVI decrease in Chongwe (Zambia) and Limpopo (South Africa) in 2014 and 2015 prior to the FAW invasion were a result of El Niño oscillations (Archer et al. 2017).Moreover, the region experienced a severe drought from 2014 to 2015, which caused a considerable reduction in NDVI and crop yield (Liu and Zhou 2021).However, the effect of precipitation on NDVI cannot be confounded with other biotic and abiotic factors as the index steadily increases if rainfall is abundant and vice versa (Pei et al. 2019;Achille et al. 2021).Also, the impact of temperature changes on NDVI is minimal (Wang et al. 2003).

Conclusions
Our study demonstrates that NDVI can be utilized as a proxy for monitoring crop damage caused by FAW.Essentially, the study reclassified NDVI imagery to four greenness classes viz., low, sparse, moderate and dense vegetation, to quantify crop damage due to FAW infestation.The study hypothesized that the change in NDVI between 2013, 2014, 2015(pre FAW invasion), and 2016, 2017and 2018 (post FAW invasion) were caused by FAW damage.We found considerable reductions in NDVI in all study regions of Benin, Togo, Malawi, Zambia and South Africa post FAW invasion.Our results can be used as a guide for monitoring the effectiveness of pest control methods that are currently rolled out across Africa to mitigate FAW effects on important staple crops like maize and sorghum.They can also be used to estimate crop yield as a function of FAW damage in affected regions.Finally, we believe that future studies should look at testing our approach in other FAW affected regions of SSA and field validated with data on FAW damage and yield.
Zentrum f€ ur Entwicklungsforschung (ZEF), Bonn, Germany.It has been carried out in partial fulfillment of a Ph.D. thesis by the first author in close collaboration with the International Centre of Insect Physiology and Ecology (icipe) Nairobi, Kenya, and the International Institute of Tropical Agriculture (IITA) Cotonou, Benin.

Figure 2 .
Figure 2. Flowchart of the hierarchical methodology steps for assessing and quantifying the damage of fall armyworm on maize productivity using the normalized difference vegetation index (NDVI).

Table 2 .
Landsat 8 satellite images 11 spectral bands, which are used to capture different aspects of the Earth's surface.

Table 3 .
The threshold range of the normalized difference vegetation index (NDVI) values for the four vegetation density classes viz., water body, low, sparse, moderate, and dense vegetation for the five study regions.
The range of NDVI values for each vegetation class in five different study regions in table 3: Atlantique (Benin), Centrale (Togo), Salima (Malawi), Chongwe (Zambia), and Limpopo (South Africa).The Low vegetation class had the lowest NDVI values, ranging from -0.67 to -0.04, while the Dense vegetation class had the highest values, ranging from 0.31 to 0.84.The range of NDVI values for Sparse vegetation and Moderate vegetation classes varied between regions.

Table 4 .
Area and percentage of change in the normalized difference vegetation index (NDVI) classes between 2013 and 2016 for the Atlantique region of Benin (veg.¼ vegetation).

Table 5 .
Area and percentage of change for the normalized difference vegetation index (NDVI) classes between 2013 and 2016 for the Centrale region of Togo (veg.¼ vegetation).Difference (ha) and Change (%) columns in the table 5 shows the change in area and percentage change, respectively, for each vegetation class between 2013 and 2016.A positive or negative sign indicates whether the area increased or decreased, respectively.

Table 6 .
Area and percentage of change for the normalized difference vegetation index (NDVI) classes between 2013 and 2017 for the Chongwe region of Zambia (veg.¼vegetation).Difference (ha) and Change (%) columns in the table 6 shows the change in area and percentage change, respectively, for each vegetation class between 2013 and 2017.A positive or negative sign indicates whether the area increased or decreased, respectively.

Table 7 .
Area and percentage of change for the normalized difference vegetation index (NDVI) classes between 2013 and 2017 for the Salima region of Malawi (veg.¼ vegetation).

Table 8 .
Area and percentage of change for the normalized difference vegetation index (NDVI) classes between 2013 and 2018 for Limpopo, South Africa (veg ¼ vegetation).