Ecological Indicators 117 (2020) 106655

Contents lists available at ScienceDirect

Ecological Indicators
journal homepage: www.elsevier.com/locate/ecolind

Modelling desert locust presences using 32-year soil moisture data on a T

large-scale
⁎
Diego Gómez , Pablo Salvador, Julia Sanz, José Luis Casanova
Remote Sensing Laboratory (LATUV), University of Valladolid. Paseo de Belen 11, 47011 Valladolid, Spain

A R T I C LE I N FO A B S T R A C T

Keywords: The desert locust is the world's most dangerous migratory pest according to the Food and Agriculture
ESA CCI SM Organization of the United Nations (FAO), and its recession area extends over more than 30 countries. While
Desert locust promising assertions have been made to relate desert locust presences with remotely sensed soil moisture (SM),
Machine learning they have not yet been tested robustly in large-scale studies. The aim of this work was to evaluate the potential of
Monitoring and preventive
soil moisture data to detect desert locust presences (solitarious nymphs) using the European Space Agency
Soil moisture
Climate Change Initiative Soil Moisture (ESA CCI SM) product from 1985 to 2017. Firstly, 5 machine learning
algorithms were fitted using various pre-processing options and variable creation scenarios. Subsequently, the
best performing models were fine-tuned using the k-fold cross validation technique and validate the results with
an independent dataset taking random dates and locations. The best results were obtained by the weighted k-
nearest neighbours (kknn) and random forest (rf) models. The kknn performance was ROC-AUC = 0.79,
KAPPA = 0.61 and accuracy = 0.83; and the rf obtained a ROC-AUC = 0.91, KAPPA = 0.56 and accu-
racy = 0.81. In general, both models agreed that SM values above 0.11 m3/m3 led to increase the possibility to
observe nymph of desert locust with a time delay between 35 and 79 days depending on the model. Furthermore,
it was found that model performances increased when the time interval of the variables was smaller, so that we
suggest the use of mean SM values over 4 days period to link presences and SM values. These results prove the
validity of our methodology to identify favourable breeding areas by means of ESA CCI SM dataset using ma-
chine learning approaches over the entire recession area of desert locust, and it could be used in any of the
affected countries for this pest. Future improvements in ESA CCI SM product (e.g. higher spatial resolution) may
lead to improve model accuracies with monitoring and preventive purposes.

1. Introduction Gregarious individuals are more active, present bright colours, high
fecundity and attract other locusts (Maeno and Tanaka, 2009; Feliciano,
Desert locust (Forskål, 1775) is one of the most dangerous migratory 2012). In contrast, solitarious individuals can be found at low densities
pests that threaten local economies and social stability in North Africa and avoid other conspecifics. Ott et al. (2010) discovered that solitar-
and Middle East countries (FAO, 2019). Despite the efforts to control ious locusts present smaller brains, having disproportionally larger
locust effects, these insects are still a major threat for food security primary visual and olfactory neuropiles, possibly to gain sensitivity. In
(Lecoq, 2003). For instance, the locust plague between 2003 and 2005 terms of behaviour, the phase transition may occur within hours (Ellis,
infested 20 countries and required 13 million of litres of pesticides. The 1962), what emphasize the necessity to better understand the spatial
control cost alone was around US$0.5 billion while the crop losses distributions of solitarious adults to monitor and control possible out-
reached US$2.5 billion (FAO, 2018). breaks (Maeno et al., 2016). Many authors agree on the necessity to
Desert locusts present high adaptability to the extreme ecological locate and treat locust populations at low densities using insecticides,
conditions of their habitat, having phase polyphenism which implies before or during gregarization (Cressman, 1996; Lecoq, 2001; Brader
drastic changes when population density increases, either in adult or et al., 2006; Magor et al., 2008). Furthermore, these tasks are re-
nymph stage (Simpson et al., 2011). Cisse et al. (2013) proved that commended to be conducted during hopper or nymphal stage, so that
phase transition (solitarious to gregarious) would occur under sparse their mobility is more limited and they cannot threaten agricultural
vegetation densities rather than green and denser conditions. lands (Maeno et al., 2016). In order to prevent locust outbreaks, Food

⁎
Corresponding author.
E-mail address: diego@latuv.uva.es (D. Gómez).

https://doi.org/10.1016/j.ecolind.2020.106655
Received 27 May 2019; Received in revised form 4 June 2020; Accepted 22 June 2020
Available online 29 June 2020
1470-160X/ © 2020 Elsevier Ltd. All rights reserved.
D. Gómez, et al. Ecological Indicators 117 (2020) 106655

and Agriculture Organization of the United Nations (FAO) developed non-restricting habitat conditions (Simpson et al., 1999) and nymph
the Desert Locust Information Service (DLIS). Their aim was to early stage to present lower mobility (wingless) (Showler et al., 2008). We
detect outbreaks by means of monitoring techniques based on Earth filtered the data to reduce the uncertainties noted by Renier et al.
observation and field work, so then preventive controls over recession (2015). The total number of presence samples was 34356.
areas (low density breeding areas) could be applied with a limited We did not use the absence records included in the SWARMS da-
amount of pesticides (leqoq, 2003; Piou et al., 2018). tabase for the following reasons: firstly, the number of presence/ab-
The Schistocerca WARning and Management System (SWARMS) is sence samples was very unequal and secondly, the solitarious in-
the database used by FAO-DLIS for desert locust global monitoring and dividuals are usually at very low densities and may go unnoticed for
early warning. It contains geo-located desert locust data collected by survey teams during recession periods (Meynard et al., 2017). In con-
national survey and control teams of affected countries since 1985. trast, we randomly generate a grid of “pseudo-absences” as reported in
Some authors such as Renier et al. (2015) noted some uncertainties in other studies using species distribution models (Engler et al., 2004;
the data records such as missing geo-location or coincident sample lo- Zaniewski et al., 2002). We used the same methodology applied on
cations and dates, so the information should be analysed and filtered Gómez et al. (2018, 2019) to create pseudo-absences. Based on soli-
before is used. In addition, solitarious individuals are usually at very tarious nymph geo-locations between 1985 and 2017, we selected the
low densities and may go unnoticed for survey teams during recession most suitable areas for nymphs to avoid geographical bias, which
periods (Meynard et al., 2017). otherwise could misguide model predictions (Barnes et al., 2014). We
Satellite remote sensing is a cost effective way of monitoring in- found that 50 km radius buffering mask from each recorded sighting
accessible or complicated regions with a good spatial and temporal enabled us to remove those zones with no reported occurrences. We
resolution (Melesse et al., 2007), and it has been widely used in pre- randomly generated 34,356 data points within the buffer zone with
vious work to identify suitable environmental conditions for desert lo- ArcMap 10.4 software (ESRI, 2016) and assigned them random dates
cust (Tucker et al., 1985; Ceccato et al., 2007; Waldner et al., 2015; from 1985 to 2017 using R software (R Core Team, 2017).
Renier et al., 2015; Piou et al., 2017; Gómez et al., 2019). Traditionally, SM data was obtained from the ESA CCI SM v4.4 dataset. It is a
precipitation has been one of the most used variables to forecast multi-decadal and global satellite-observed SM dataset generated via
breeding sites (Hielkema & Snijders, 1994); nevertheless, its probability the climate change initiative (CCI) of the European Space Agency
of detection ranges from 20 to 70% in arid and semiarid areas, tending (ESA). This product combines single active and passive sensors into
to overestimate rainfall occurrences (Dinku et al., 2010). Given the link three harmonized products: a merged active, a merged passive, and a
between rainfall and soil moisture (SM), several studies described the merged from active and passive sensors. We used the latter product
benefits of studying soil moisture so as to identify favourable desert which combines active and passive data. It is more complete since uses
locust breeding conditions (Tucker et al., 1985; Liu et al., 2008; Gómez the pixel from either the active or passive source, or the average value
et al., 2018; Piou et al., 2019). Egg-laying location, egg-survival, and of both depending on the performance of the vegetation optical depth
egg-hatching rate are dependent on the SM conditions between 5 and from the Advanced Microwave Scanning Radiometer for EOS (AMSR-E)
10 cm below the surface (Symmons, 2001; Liu et al., 2008). Other C-band observations (Liu et al., 2012). ESA CCI SM imagery provides
environmental factors such as temperature or wind also condition the the volumetric SM content of the first 5 cm of the top soil in m3∕m3
egg development phase (Bennett, 1976) which can last up to 95 days units, with a spatial resolution of 0.25 deg and a daily coverage
(Symmons and Cressman, 2001). Despite satellite remote sensing only worldwide since 1978 (Liu et al., 2012; Dorigo et al., 2017; Gruber
retrieves SM content from 1 to 5 first cm of the soil, and desert locusts et al., 2017). Generally, it provides good estimations of SM in com-
lay eggs usually at a depth down to 10 cm; Albergel et al. (2008) de- parison to land surface models or in-situ observations, and was vali-
monstrated the strong relationship of the top SM with deeper soil dated against ground measurements or other sensors (Gruber et al.,
layers. Furthermore, sufficient SM content favours vegetation growth 2017). However, it presents some uncertainties with certain surface
what implies food supply and shelter for desert locusts at nymph and conditions such as dense vegetation or organic soils (Dorigo et al.,
adult stages (Bennet, 1976; Nevo, 1996). 2017). The list of satellites and sensors that form this product is the
The use of machine learning methods can improve the traditional following: (1) Active data from C-band scatterometers on board of ERS-
predictive performance of species distributions models and their capa- 1, ERS-2, MetOp-A, and MetOp-B satellites (generated by the “TU
city to incorporate complex interactions among variables (Elith et al., Wien”); (2) Passive data from Nimbus 7 SMMR, DMSP SSM/I, TRMM
2006; Booth et al., 2014), making them able to work with large eco- TMI, Aqua AMSR-E, Coriolis WindSat, GCOM-W1 AMSR2, and SMOS
logical datasets (Robinson et al., 2014). (generated by VU University Amsterdam in collaboration with NASA).
The aim of this study is to evaluate the potential of SM data to detect
desert locust presences on a large scale (31 countries), in a solitarious 2.2. Methods
phase at nymph stage. Our methodology is based on machine learning
approaches that relate locust ground-based observations and 32-year The present study aims to develop a methodology to predict desert
satellite imagery. This study covers the period between 1985 and 2017. locust presences at nymph stage and solitarious phase identifying the
most suitable SM conditions across 32 years and 31 countries. We used
2. Materials and methods a derived approach used in Gómez et al. (2018). In the latter study, the
proposed model was able to identify breeding locations based on SM
2.1. Study area and materials patterns in Mauritania. The current study is more ambitious in terms of
study-area extension (Fig. 1), and we used a more up-to-date SM dataset
The study site covers the desert locust recession area (Fig. 1) which (ESA CCI SM version 4.4) and two-year longer study-time. Thus, there
comprises > 30 countries in the northern Africa and south-east Asia were more data to build and evaluate the predictive models and their
(Anyamba et al., 2005). According to the Koppen classification, there results can be applied to any other recession areas beyond Mauritanian
are principally 3 types of climate in order of importance: “BWh” hot borders.
arid dessert, “BSh” hot arid steppe and “Csa” hot summer dry (Kottek Fig. 2 shows the sequence of the proposed method as a flow chart.
et al., 2006). Mean annual precipitation rates vary across countries and Given the importance of SM to favour a successful egg lying, hatching
latitudes, but generally are below 300 mm/year (Fick and Hijmans, and development, our predictor variables were formed based on the
2017). antecedent SM conditions of each pixel in which presence/pseudo-ab-
We used the SWARMS dataset to locate locusts in the study area. sence sites were located. For these pixels, the SM value was extracted
The target population were locusts at solitarious phase to be related to since 95 days prior the survey time and aggregated in temporal

2
D. Gómez, et al. Ecological Indicators 117 (2020) 106655

Fig. 1. A) Nymph presences in solitarious phase (green dots) and B) pseudo-absence sites (black dots) across the study area from 1985 to 2017. C) Average of the
annual precipitation (mm) in the region of study (Fick and Hijmans, 2017).

intervals of 4, 6 and 8 days (3 temporal scenarios) using the minimum training dataset was partitioned in 5 subsets so that the models were
or mean data value across each time interval (2 descriptive statistics). trained on 4 subsets, and their accuracy evaluated on the subset not
By this method, we may differentiate critical events in the locust life- used for training. Thus, different hyper-parameter combinations were
cycle such as egg-laying, egg-hatching, and early stages of the nymph tested to select the set of parameters that reported the lowest error.
phase individuals (Gómez et al., 2018). In addition, it permits to deal Finally, we evaluated the generalization capacity of the models using
with punctual missing data that some ESA CCI SM v4.4 images have due the independent test set (30% of the original data) and assess the re-
to the satellite revisit times. lative importance of each variable in the final results.
From now onwards, we refer to these variables with the acronym
“SM” and a sub-index number. For instance, if we aggregate SM values
in 4 days, the SM1 would correspond to the SM value obtained between 3. Results
95 and 92 days before the record date; SM2 between 91 and 88 days,
and so on until the closest day to the survey time which has the highest The best model performance in terms of ROC-AUC was obtained by
sub-index. In addition, we fitted 5 machine learning algorithms using the rf algorithm, when the variables were created using 4-day time
the Caret package (Kuhn, 2008): logistic regression model “glm”, eX- intervals either using the minimum or mean SM values of the given
treme Gradient Boosting “xgbTree” (Chen et al., 2015), Weighted k- periods (Table 1). The algorithms kknn and rf achieved from fair to
Nearest Neighbors “kknn” (Schliep et al., 2016), Feed-Forward Neural good KAPPA scores with values close to 0.60 (Table 2). In general, the
Networks and Multinomial Log-Linear Models “nnet” (Ripley et al., pre-processing options across algorithms and time intervals offered si-
2016), and random forest “rf” (Breiman, 2001). During the model milar results in terms of ROC-AUC and KAPPA. Nevertheless, the option
building phase for each algorithm, 4 different pre-processing options “scale, center and pca” reported lower scores across algorithms. It is
were tested to ensure the best performance of each model and scenario: worth stressing that models were in most cases better using 4-day
(1) center and scale, (2) scale, (3) range and (4) scale, center and variables with mean SM values.
principal component analysis “pca”. We randomly split the dataset into We optimized rf and kknn algorithms by means of k-fold cross va-
70% to build the models and the remaining 30% of the data to evaluate lidation. Aiming to improve model results, we used the “center and
the predictions. This process was repeated 10 times with predefined scale” option to pre-process the data and mean SM values every 4-day
seeds and the results were averaged. In order to measure model per- time interval (Table 1-2). The best hyper-parameters for the kknn ap-
formances, we used the Receiver Operating Curve Area - Under the proach were the following: kmax = 9, distance = 2 and kernel = op-
Curve “ROC-AUC” (Hanley & McNeil, 1982) and Kappa (Monserud & timal. This kknn model achieved good predictive performance with
Leemans, 1992) statistics. high ROC-AUC (0.79) and KAPPA (0.61) values, accuracy = 0.83 (95%
We selected the model with the highest performance in terms of CI = 0.82–0.83), sensitivity = 0.67 and specificity = 0.79 (Table 3). In
ROC-AUC and the model that obtained the highest KAPPA metric. the rf approach, we optimised the number of variables randomly sam-
These 2 models were subsequently fine-tuned through k-fold cross va- pled as candidates at each split “mtry”, being mtry = 13 the most
lidation (k = 5). Firstly, the data was split into 2 datasets: 70% for optimal choice. This rf model gave a higher ROC-AUC (0.91) and
training and 30% for testing with independent data. Subsequently, the slightly lower KAPPA (0.56) in comparison to kknn model. Further-
more, rf results showed an accuracy = 0.81 (95% CI = 0.80–0.82),

3
D. Gómez, et al. Ecological Indicators 117 (2020) 106655

of both methods to detect solitarious nymphs of desert locust with high

accuracies. It is worth noting that models were better identifying actual
pseudo-absences (specificity) than actual positives that are correctly
identified as such (sensitivity).
Fig. 3 shows the relative importance of each variable (%) in the final
outcome of the models. Variables without graphical bar representation
mean irrelevant contribution to the model. While the most important
variables (> 50% of influence) in the kknn approach were clearly from
SM1 to SM15 (between 95 and 35 days prior the survey time), the
variable importance in the rf model showed a more irregular pattern
with less variables > 50% of importance (SM1, SM2, SM3, SM4,
SM13, SM24). The distribution of SM values for presence records and
pseudo-absence records across the different time-based variables is
shown in Fig. 4. The variable importance pattern observed for kknn
coincides with the trend of SM median values in presence records
(Fig. 3, Fig. 4B), so that the aforementioned variables (SM1-SM15)
would coincide with median SM values > 0.11 m3/m3. Fig. 4A in-
dicates the little variability of SM across time in the study area for the
sites assigned as pseudo-absences. Those locations maintain similar SM
values because they characterize typical values of the study region. On
the contrary, the SM values recorded before hopper sightings (Fig. 4B)
present a different pattern. The latter show under which conditions
hoppers are more likely to be present than on average. Therefore, the
different distribution values of SM between presences (Fig. 4B) and
pseudo-absences (Fig. 4A) would validate the applied methodological
approach to generate absences of desert locust.

4. Discussion

In this study, we fitted 5 machine learning algorithms for classifi-

cation purposes to derive favourable SM conditions to desert locust in
solitarious phase and nymph stage.
Fig. 2. Workflow of the proposed methodology to study the link of ESA CCI SM Our results are in line with previous studies that highlighted the
with desert locusts using a machine learning approach. importance of the antecedent soil moisture conditions to increase the
probability of nymph́ s occurrence (Popov, 1958; Tucker et al., 1985;
Pener & Simpson, 2009). Rainfall has been traditionally regarded as a
sensitivity = 0.59 and specificity = 0.93. Both approaches got high
strong predictor to locate favourable locust environments (FAO, 2016),
accuracy (> 0.81) and good ROC-AUC values (> 0.79). In the kknn
but remotely sensed precipitation is usually overestimated due to sub-
method, KAPPA and sensitivity scores were better, while rf out-
cloud evaporation (Dinku et al., 2011). While promising assertions have
performed kknn on specificity. These results demonstrate the suitability
been made to relate desert locust presences with remotely sensed soil

Table 1
Model results per algorithm, pre-processing option (a: center and scale; b: scale; c: range; d: scale, center and range), time interval and descriptive statistic (minimum
or mean SM) in terms of ROC-AUC metric.
Statistic Mean Min

Algorithm Pre-processing 4-day 6-day 8-day 4-day 6-day 8-day Average

glm a 0.67 0.62 0.61 0.65 0.56 0.55 0.61

b 0.67 0.62 0.61 0.65 0.56 0.55 0.61
c 0.67 0.62 0.61 0.65 0.56 0.55 0.61
d 0.67 0.62 0.61 0.65 0.56 0.55 0.61
kknn a 0.79 0.77 0.75 0.75 0.73 0.72 0.75
b 0.79 0.77 0.75 0.75 0.73 0.72 0.75
c 0.79 0.77 0.75 0.75 0.73 0.72 0.75
d 0.77 0.74 0.74 0.78 0.71 0.70 0.74
nnet a 0.75 0.72 0.70 0.73 0.68 0.66 0.71
b 0.73 0.71 0.69 0.72 0.68 0.65 0.70
c 0.75 0.72 0.70 0.73 0.68 0.65 0.71
d 0.75 0.71 0.70 0.73 0.69 0.65 0.71
rf a 0.91 0.89 0.88 0.91 0.90 0.89 0.90
b 0.91 0.89 0.88 0.91 0.90 0.89 0.90
c 0.91 0.89 0.88 0.91 0.90 0.89 0.90
d 0.91 0.88 0.85 0.88 0.86 0.84 0.87
xgbTree a 0.81 0.77 0.75 0.80 0.77 0.75 0.78
b 0.81 0.78 0.76 0.80 0.77 0.75 0.78
c 0.81 0.77 0.75 0.80 0.77 0.75 0.78
d 0.81 0.76 0.74 0.78 0.74 0.71 0.76
Average 0.79 0.75 0.74 0.77 0.72 0.71 0.74

4
D. Gómez, et al. Ecological Indicators 117 (2020) 106655

Table 2
Model results per algorithm, pre-processing option (a: center and scale; b: scale; c: range; d: scale, center and range), time interval and descriptive statistic (minimum
or mean SM) in terms of KAPPA metric.
Statistic Mean Min

Algorithm Pre-processing 4-day 6-day 8-day 4-day 6-day 8-day Average

glm a 0.06 0.09 0.11 0.05 −0.01 −0.01 0.05

b 0.06 0.09 0.11 0.05 −0.01 −0.01 0.05
c 0.06 0.09 0.11 0.05 −0.01 −0.01 0.05
d 0.06 0.08 0.11 0.05 −0.01 0.00 0.05
kknn a 0.60 0.55 0.51 0.53 0.48 0.44 0.52
b 0.60 0.55 0.51 0.53 0.48 0.44 0.52
c 0.60 0.55 0.51 0.53 0.48 0.44 0.52
d 0.57 0.49 0.44 0.48 0.43 0.40 0.52
nnet a 0.36 0.32 0.30 0.33 0.27 0.23 0.30
b 0.33 0.31 0.28 0.29 0.26 0.24 0.29
c 0.36 0.32 0.30 0.34 0.27 0.23 0.30
d 0.35 0.31 0.30 0.33 0.27 0.22 0.30
rf a 0.57 0.56 0.54 0.54 0.58 0.57 0.56
b 0.56 0.56 0.54 0.54 0.58 0.57 0.56
c 0.57 0.56 0.54 0.54 0.58 0.57 0.56
d 0.57 0.54 0.50 0.51 0.49 0.47 0.51
xgbTree a 0.45 0.40 0.37 0.43 0.39 0.36 0.40
b 0.45 0.40 0.37 0.43 0.39 0.36 0.40
c 0.46 0.40 0.37 0.43 0.39 0.36 0.40
d 0.45 0.38 0.35 0.40 0.34 0.29 0.37
Average 0.40 0.38 0.36 0.37 0.33 0.31 0.36

moisture (Liu et al., 2008; Gómez et al., 2018; Piou et al., 2019), they These results concur with previous rf approaches such as Gómez et al.
have not yet been tested robustly in large-scale studies (not more than 4 (2018) and Piou et al. (2019) whose ROC-AUC values were 0.94 and
countries). 0.76 respectively, but it is worth highlighting that our study covered a
Our machine learning models were built using the whole desert much wider area.
locust recession area (31 countries), with 34,365 solitarious nymph The most relevant variables in our kknn model were found from
sightings and the longest SM dataset available to date. The ESA CCI SM SM1 to SM15 (about 60 days). This fact implies that median values
dataset was found to be appropriate because it is the most complete and above 0.11 m3/m3 increases the possibilities to present nymphs after
consistent global SM dataset available (Wagner et al., 2012). Further- 35 days (Figs. 3-4). The interpretation of the variable importance de-
more, the large dimensions of the study area permitted the use of rived from the rf model is not as straightforward because it presents
presence records in great numbers for solitarious nymphs (1985–2017). some time discontinuities in comparison with the kknn approach,
These types of individuals provide critical signs of breeding-area loca- however it agrees to highlight the importance of the variables from SM1
tions because they move very little at solitarious phase (Hemming et al., to SM4 (about 16 days). Hence, median SM values above 0.11 m3/m3
1979) what is of great importance to detect locust outbreaks before they for at least 16 days would lead to increase the possibilities to detect
became gregarious (Lecoq, 2003). Altogether, it ensures robustness to desert locust after 79 days.
our models, which were built and independently evaluated over the Similar results concerning time response can be seen in the litera-
whole recession area (31 countries) so that they may be used in any of ture. For instance, it was found that an area becomes favourable for
the aforementioned countries (Fig. 1). The use of pseudo-absences may breeding purposes when minimum SM values > 0.07 m3/m3 during at
be controversial in certain fields because they bring some uncertainty least 6 days, with a time response of 72 days (Gómez et al., 2018).
into the model results (Hastie & Fithian, 2013). Nevertheless, their use Moreover, Piou et al. (2019) observed that mean SM values > 0.09 m3/
is generally justified for providing a set of conditions available in the m3 for at least 20 days would increase the chance to observe locusts
study region that need to be included in this type of models (Phillips 70 days later. Some uncertainties regarding time response and optimal
et al., 2009). Furthermore, our results demonstrate the validity of the values are expected due to different soil features in the study area
used methodology to locate pseudo-absences given the low SM varia- (texture, bulk density, structure, porosity or permeability), vegetation
bility shown across time in comparison with the higher variability on growth and locust dynamics (migrate to suitable areas, egg laying and
presence sites (Fig. 4). incubation or growth) (Wang et al., 2007; Piou et al., 2019). The dif-
The best performances were acquired by kknn and rf algorithms, ferent time response between favourable SM conditions for breeding
dividing the survey time into ranges of 4 days, and selecting the mean and the presence of nymphs may be related to the number of days used
SM as the variable value. Based on Thuiller et al. (2009), the model in the variable creation step, which in our case was 4 days. And in this
evaluation for kknn indicated a good performance in terms of KAPPA sense, model performance was increased when the time interval of the
(0.61), and ROC-AUC (0.79), with an overall accuracy of 0.83. The rf variables was smaller; we therefore suggest regarding mean SM values
approach obtained a very high performance in terms of ROC-AUC over a 4-day period to link presences and SM values. The use of these
(0.91), overall accuracy (0.81), and fair outcome of KAPPA (0.56). locust individuals (solitarious nymphs) provide additional information

Table 3
Model results after hyper-parameter optimization in terms of accuracy, 95% CI (Confidence Interval), sensitivity, specificity, ROC-AUC and KAPPA.
Algorithm Accuracy 95% CI Sensitivity Specificity ROC-AUC KAPPA

kknn 0.83 (0.82,0.83) 0.67 0.91 0.79 0.61

rf 0.81 (0.80,0.82) 0.59 0.93 0.91 0.56

5
D. Gómez, et al. Ecological Indicators 117 (2020) 106655

Fig. 3. Variable importance of the 20 most important variables of the kknn (blue) and rf (red) approaches. The median SM value of each variable, only for presences,
is represented in green (m3/m3).

Fig. 4. Distribution of SM values for pseudo-absence (A) and presence (B) records across the different time-based variables.

about environmental limitations and location of breeding areas such as approach may be used in any of the countries making up the desert
described in Tucker et al. (1985) and Hemming et al. (1979). The locust recession area as our models were built and validated within its
models were built and validated using the most complete dataset of 31 countries (Fig. 1). In addition, it can also improve tasks such as lo-
locust records (SWARS - FAO) using the longest SM dataset available cust monitoring, management and preventive control in a cost-effective
over the entire desert locust recession area. To date, none has yet in- manner.
tended to relate SM with locust presences in large-scale studies. The use
of representative SM values (minimum or mean) by intervals of time
5. Conclusions
has been verified as a valid approach, finding that the most optimal
time-interval was 4 days.
The aim of this study was to assess the capability of ESA CCI SM
These results prove the validity of our methodology to identify fa-
data to identify favourable desert locust breeding sites on a large-scale
vourable breeding areas by means of ESA CCI SM dataset. This
basis across 32 years. While promising assertions have been made to

6
D. Gómez, et al. Ecological Indicators 117 (2020) 106655

relate desert locust presences with remotely sensed soil moisture, they References
had not yet been tested robustly in large-scale studies. Given the wide
variety of ecological conditions in the study area, our machine learning Albergel, C., Rüdiger, C., Pellarin, T., Calvet, J.C., Fritz, N., Froissard, F., Martin, E., 2008.
models were built using the whole desert locust recession area (31 From near-surface to root-zone soil moisture using an exponential filter: an assess-
ment of the method based on in-situ observations and model simulations. Hydrol.
countries), with 34,365 solitarious nymph records and the longest SM Earth Syst. Sci. 12 (6), 1323–1337.
dataset available to date. After a model selection process, we chose two Anyamba, A., Small, J., Tucker, C. J., Cressman, K., Love, T. B., & Linthicum, K. J. (2005).
robust machine learning models (kknn and rf) capable to detect SM Remote sensing of eco-climatic conditions associated with the 2004 Desert Locust
outbreak in Northwest and Sahelian Africa. In Pecora 16 Symp. Global Priorities in
patterns and to identify solitarious nymph of desert locust. The kknn Land Remote Sensing. http://www.asprs.org/a/publications/proceedings/pecora16/
approach obtained a good performance in terms of KAPPA (0.61), ROC- Anyamba_A.pdf (Accessed at 17/04/2019).
AUC (0.79) and accuracy (0.83). The predictive capacity of the rf model Barnes, M.A., Jerde, C.L., Wittmann, M.E., Chadderton, W.L., Ding, J., Zhang, J., Lodge,
D.M., 2014. Geographic selection bias of occurrence data influences transferability of
was better off in terms of ROC-AUC (0.91), although slightly worse in invasive Hydrilla verticillata distribution models. Ecol. Evol. 4 (12), 2584–2593.
accuracy (0.81) and KAPPA (0.56). In general, both models agreed that Bennett, L.V., 1976. The development and termination of the 1968 plague of the Desert
SM values above 0.11 m3/m3 led to increase the possibility to observe locust, Schistocerca gregaria (Forskål)(Orthoptera, Acrididae). Bull. Entomol. Res. 66
(3), 511–552.
nymph of desert locust with a time delay between 35 and 79 days de-
Booth, T.H., Nix, H.A., Busby, J.R., Hutchinson, M.F., 2014. BIOCLIM: the first species
pending on the model we used. It was worth noting that model per- distribution modelling package, its early applications and relevance to most current
formances increased when the time interval of the variables was MAXENT studies. Divers. Distrib. 20 (1), 1–9.
smaller, we therefore suggest regarding mean SM values over 4-days Brader, L., 2006. Towards a More Effective Response to Desert Locusts and Their Impacts
on Food Security, Livelihoods and Poverty. In: Multilateral Evaluation of the 2003–05
period to link presences and SM values on the ground. Desert Locust Campaign. Food and Agriculture Organisation, Rome.
These results prove the validity of our methodology to identify fa- Breiman, L., 2001. Random forests. Machine learning 45 (1), 5–32.
vourable breeding areas by means of ESA CCI SM dataset using machine Ceccato, P., Cressman, K., Giannini, A., Trzaska, S., 2007. The desert locust upsurge in
West Africa (2003–2005): Information on the desert locust early warning system and
learning approaches. We believe that the use of solitarious nymphs as the prospects for seasonal climate forecasting. Int. J. Pest Manage. 53 (1), 7–13.
target population increases the success of this type of studies because Chen, T., He, T., Benesty, M., Khotilovich, V., Tang, Y., 2015. Xgboost: extreme gradient
they are easier to track (more limited mobility) and indicate breeding boosting. R Package Version (4-2), 1–4.
Dinku, T., Ceccato, P., Cressman, K., Connor, S.J., 2010. Evaluating detection skills of
sites, what would improve monitoring and preventive management to satellite rainfall estimates over desert locust recession regions. J. Appl. Meteorol.
avoid locust outbreaks. The proposed models were validated with in- Climatol. 49 (6), 1322–1332.
dependent data over the entire recession area (31 countries), what Cisse, S., et al., 2013. Effect of vegetation on density thresholds of adult desert locust
gregarization from survey data in Mauritania. Entomol. Exp. Appl. https://doi.org/
ensures their capacity to be used in any of the affected countries unlike 10.1111/eea.12121.
other proposed models which are based on more local environmental Cressman, K., 1996. Current methods of desert locust forecasting at FAO 1. EPPO Bulletin.
conditions. Future improvements in ESA CCI SM product (e.g. higher https://doi.org/10.1111/j.1365-2338.1996.tb01500.x.
Dinku, T., Ceccato, P., Connor, S.J., 2011. Challenges of satellite rainfall estimation over
spatial resolution) may lead to improve model accuracies with mon-
mountainous and arid parts of east Africa. Int. J. Remote Sens. 32 (21), 5965–5979.
itoring and preventive purposes. Dorigo, W., Wagner, W., Albergel, C., Albrecht, F., Balsamo, G., Brocca, L., Haas, E., 2017.
ESA CCI Soil Moisture for improved Earth system understanding: State-of-the art and
future directions. Remote Sens. Environ. 203, 185–215.
Funding Elith, J., et al., 2006. Novel methods improve prediction of species’ distributions from
occurrence data. Ecography. https://doi.org/10.1111/j.2006.0906-7590.04596.x.
Ellis, P.E., 1962. “The behaviour of locusts in relation to phases and species”, Colloques
This research did not receive any specific grant from funding Internationaux du Centre National de la Recherche Scientifique, No. 114.
agencies in the public, commercial, or not-for-profit sectors. Physiologie, Comportement et Écologie des Acridiens en Rapport avec la Phase.
CNRS, Paris.
Engler, R., Guisan, A., Rechsteiner, L., 2004. An improved approach for predicting the
distribution of rare and endangered species from occurrence and pseudo-absence
CRediT authorship contribution statement data. J. Appl. Ecol. 41 (2), 263–274.
ESRI 2016. ArcMap 10.4. Redlands, CA, USA: Environmental Systems Research Institute
Diego Gómez: Conceptualization, Data curation, Formal analysis, https://doi.org/10.1016/j.patrec.2005.10.010.
FAO and WMO, “Weather and desert locusts” (2016), http://www.fao.org/ag/locusts/
Investigation, Methodology, Resources, Software, Validation, common/ecg/2350/en/2016_WMOFAO_WeatherDLe.pdf (19 April 2018).
Visualization, Writing - original draft, Writing - review & editing. Pablo FAO, 2018. http://www.fao.org/in-action/countries-take-responsibility-for-regional-de-
Salvador: Data curation, Methodology, Software, Resources, sert-locust-control/en/ (Accessed at 16/12/2018).
FAO, 2019. http://www.fao.org/news/story/en/item/1180586/icode/ (Accessed at 16/
Validation, Writing - review & editing. Julia Sanz: Project adminis- 04/2019).
tration, Resources, Supervision. José Luis Casanova: Project admin- Feliciano, P., 2012. Gregarious locusts. Nat. Genet. 44 (2), 120.
istration, Resources, Supervision, Writing - review & editing. Fick, S.E. and Hijmans, R.J. (2017). Worldclim 2: New 1-km spatial resolution climate.
http://worldclim.org/version2 (Accessed at 17/04/2019).
Gómez, D., Salvador, P., Sanz, J., Casanova, C., Taratiel, D., Casanova, J.L., 2018.
Machine learning approach to locate desert locust breeding areas based on ESA CCI
Declaration of Competing Interest soil moisture. J. Appl. Remote Sens. 12 (3), 036011.
Gómez, D., Salvador, P., Sanz, J., Casanova, C., Taratiel, D., Casanova, J.L., 2019. Desert
locust detection using Earth observation satellite data in Mauritania. J. Arid Environ.
The authors declare that they have no known competing financial 164, 29–37.
interests or personal relationships that could have appeared to influ- Gruber, A., Dorigo, W.A., Crow, W., Wagner, W., 2017. Triple collocation-based merging
ence the work reported in this paper. of satellite soil moisture retrievals. IEEE Trans. Geosci. Remote Sens. 55 (12),
6780–6792.
Hanley, J.A., McNeil, B.J., 1982. The meaning and use of the area under a receiver op-
erating characteristic (ROC) curve. Radiology 143 (1), 29–36.
Acknowledgements Hastie, T., Fithian, W., 2013. Inference from presence-only data; the ongoing controversy.
Ecography 36 (8), 864–867.
Hemming, C.F., Popov, G.B., Roffey, J., Waloff, Z., Rainey, R.C., 1979. Characteristics of
Authors would like to show our gratitude to the ESA Climate Change
desert locust plague upsurges. Philos. Trans. Royal Soc. London. B, Biol. Sci. 287
Initiative and the Soil Moisture CCI project-team for providing free (1022), 375–386.
access to the Combined Soil Moisture dataset. We would also like to Hielkema, J.U., Snijders, F.L., 1994. Operational use of environmental satellite remote
sensing and satellite communications technology for global food security and locust
acknowledge Keith Cressman and the FAO-DLIS team from FAO to fa-
control by FAO: The ARTEMIS and DIANA systems. Acta Astronaut. 32 (9), 603–616.
cilitate us SWARMS database and make possible this research, as well as Kottek, M., Grieser, J., Beck, C., Rudolf, B., Rubel, F., 2006. World map of the Köppen-
all the current and past locust field workers and National Centres for Geiger climate classification updated. Meteorol. Z. 15 (3), 259–263.
Locust Control of the affected countries, to collect information about Kuhn, M., 2008. Building predictive models in R using the caret package. J. Stat. Softw.
28 (5), 1–26.
the desert locust and its environment.

7
D. Gómez, et al. Ecological Indicators 117 (2020) 106655

Lecoq, M., 2001. Recent progress in Desert and Migratory Locust management in Africa. Popov, G. B. (1958). Ecological studies on oviposition by swarms of the Desert Locust
Are preventative actions possible? J. Orthoptera Res. https://doi.org/10.1665/1082- (Schistocerca gregaria Forskal) in eastern Africa. Ecological Studies on Oviposition by
6467(2001)010[0277:RPIDAM]2.0.CO;2. Swarms of the Desert Locust (Schistocerca gregaria Forskal) in eastern Africa, (31).
Lecoq, M., 2003. Desert locust threat to agricultural development and food security and R Core Team, 2017. R: A language and environment for statistical computing. R
FAO/International role in its control. Arab Soc. Plant Protect. 21, 188–193. Foundation for Statistical Computing, Vienna, Austria https://www.R-project.org/.
Liu, Z., Shi, X., Warner, E., Ge, Y., Yu, D., Ni, S., Wang, H., 2008. Relationship between Renier, C., Waldner, F., Jacques, D., Babah Ebbe, M., Cressman, K., Defourny, P., 2015. A
oriental migratory locust plague and soil moisture extracted from MODIS data. Int. J. dynamic vegetation senescence indicator for near-real-time desert locust habitat
Appl. Earth Obs. Geoinf. 10 (1), 84–91. monitoring with MODIS. Remote Sensing 7 (6), 7545–7570.
Liu, Y.Y., Dorigo, W.A., Parinussa, R.M., de Jeu, R.A., Wagner, W., McCabe, M.F., Van Ripley, B., Venables, W., Ripley, M.B., 2016. Package ‘nnet’. R package version 7, 3–12.
Dijk, A.I.J.M., 2012. Trend-preserving blending of passive and active microwave soil Robinson, T., et al., 2014. Mapping the Global Distribution of Livestock. PloS one.
moisture retrievals. Remote Sens. Environ. 123, 280–297. https://doi.org/10.1371/journal.pone.0096084.
Maeno, K., Tanaka, S., 2009. Is juvenile hormone involved in the maternal regulation of Schliep K., Hechenbichler K., Lizee A. (2016). kknn: weighted k-nearest neighbors.
egg size and progeny characteristics in the desert locust? J. Insect Physiol. 55 (11), https://cran.r-project.org/web/packages/kknn/index.html. (Accessed at 04/04/
1021–1028. 2019).
Maeno, K.O., Ely, S.O., Nakamura, S., Abdellaoui, K., Cissé, S., Jaavar, M.E.H., Babah, Showler, A.T., 2008. dEsert Locust, Schistocerca gregaria forskål (orthoptera: acrididae)
M.A.O., 2016. Daily microhabitat shifting of solitarious-phase Desert locust adults: plagues. In: Encyclopedia of Entomology. Springer, Dordrecht, pp. 1181–1186.
implications for meaningful population monitoring. SpringerPlus 5 (1), 107. Simpson, S.J., McCaffery, A.R., Hägele, B.F., 1999. A behavioral analysis of phase change
Magor, J., et al., 2008. Preventive control and dEsert Locust plagues. Crop Prot. https:// in the desert locust. Biol. Rev. 74, 461–480.
doi.org/10.1016/j.cropro.2008.08.006. Simpson, S.J., Sword, G.A., Lo, N., 2011. Polyphenism in insects. Curr. Biol. 21 (18),
Melesse, A., et al., 2007. Remote sensing sensors and applications in environmental re- R738–R749.
sources mapping and modelling. Sensors. https://doi.org/10.3390/s7123209. Symmons, P., 2001. Desert Locust Guidelines: Biology and Behaviour. FAO, Rome.
Meynard, C.N., Gay, P.E., Lecoq, M., Foucart, A., Piou, C., Chapuis, M.P., 2017. Climate- https://www.researchgate.net/profile/Keith_Cressman/publication/265217762_
driven geographic distribution of the desert locust during recession periods: Desert_Locust_Guidelines_1_Biology_and_Behaviour/links/
Subspecies’ niche differentiation and relative risks under scenarios of climate change. 563b3ea808ae45b5d285c7fe.pdf.
Glob. Change Biol. 23 (11), 4739–4749. Symmons, P.M., Cressman, K., 2001. Desert locust guidelines: biology and behaviour.
Monserud, R.A., Leemans, R., 1992. Comparing global vegetation maps with the Kappa FAO, Rome.
statistic. Ecol. Model. 62 (4), 275–293. Tucker, C.J., Hielkema, J.U., Roffey, J., 1985. The potential of satellite remote sensing of
Nevo, D., 1996. The desert locust, Schistocerca gregaria, and its control in the land of ecological conditions for survey and forecasting desert-locust activity. Int. J. Remote
Israel and the Near East in antiquity, with some reflections on its appearance in Israel Sens. 6 (1), 127–138.
in modern times. Phytoparasitica 24 (1), 7–32. Thuiller, W., Lafourcade, B., & Araujo, M. (2009). ModOperating manual for BIOMOD.
Ott, S., et al., 2010. Gregarious desert locusts have substantially larger brains with altered Thuiller W, Lafourcade B (2010) BIOMOD: species/climate modelling functions. R
proportions compared with the solitarious phase. Proc. R. Soc. B: Biol. Sci. https:// package version, 1-1. https://pdfs.semanticscholar.org/5522/
doi.org/10.1098/rspb.2010.0694. 2d3a8501206807a9a0a32a2b7e62260f53af.pdf (Accessed at 09/05/2019).
Pener, M.P., Simpson, S.J., 2009. Locust phase polyphenism: an update. Adv. Insect Wagner, W., Dorigo, W., de Jeu, R., Fernandez, D., Benveniste, J., Haas, E., Ertl, M., 2012.
Physiol. 36, 1–272. Fusion of active and passive microwave observations to create an essential climate
Phillips, S.J., Dudík, M., Elith, J., Graham, C.H., Lehmann, A., Leathwick, J., Ferrier, S., variable data record on soil moisture. ISPRS Annals of the Photogrammetry. Remote
2009. Sample selection bias and presence-only distribution models: implications for Sensing Spatial Inform. Sci. (ISPRS Annals) 7, 315–321.
background and pseudo-absence data. Ecol. Appl. 19 (1), 181–197. Waldner, F., Ebbe, M., Cressman, K., Defourny, P., 2015. Operational monitoring of the
Piou, C., et al., 2019. Soil moisture from remote sensing to forecast desert locust presence. Desert Locust habitat with Earth Observation: An assessment. ISPRS Int. J. Geo-Inf. 4
J. Appl. Ecol. https://doi.org/10.1111/1365-2664.13323. (4), 2379–2400.
Piou, C., Bacar, M.E.H.J., Ebbe, M.A.O.B., Chihrane, J., Ghaout, S., Cisse, S., Halima, T.B., Wang, X., Xie, H., Guan, H., Zhou, X., 2007. Different responses of MODIS-derived NDVI
2017. Mapping the spatiotemporal distributions of the desert locust in Mauritania to root-zone soil moisture in semi-arid and humid regions. J. Hydrol. 340 (1–2),
and Morocco to improve preventive management. Basic Appl. Ecol. 25, 37–47. 12–24.
Piou, C., Gay, P.E., Benahi, A.S., Babah Ebbe, M.A.O., Chihrane, J., Ghaout, S., Merlin, O., Zaniewski, A.E., Lehmann, A., Overton, J.M., 2002. Predicting species spatial distribu-
2018. Soil moisture from remote sensing to forecast desert locust presence. J. Appl. tions using presence-only data: a case study of native New Zealand ferns. Ecol. Model.
Ecol. https://doi.org/10.1111/1365-2664.13323. 157 (2–3), 261–280.