The document describes a methodology for forecasting dengue vector populations one week ahead using multi-source earth observation data and recurrent neural networks. The methodology clusters mosquito population ground truth data using k-means clustering. It then trains an encoder-decoder LSTM neural network on time series data of earth observation features means for each cluster. The model is tested against random forest and k-nearest neighbor models on data from two locations in Brazil. Results show the LSTM model more accurately follows the highest and lowest observed mosquito population values compared to the other models.

1. DENGUE VECTOR POPULATION FORECASTING USING
MULTI-SOURCE EARTH OBSERVATION PRODUCTS
AND RECURRENT NEURAL NETWORKS (RNN).

2. Mudele, O., Frery, A. C., Zanandrez, L. F., Eiras, A. E., & Gamba, P.
(2021). Dengue Vector Population Forecasting Using Multisource
Earth Observation Products and Recurrent Neural Networks. IEEE
Journal of Selected Topics in Applied Earth Observations and Remote
Sensing, 14, 4390-4404.
Reference literature

3. 1 INTRODUCTION
2 RECURRENT NEURAL NETWORKS
3 METHODOLOGY
4 RESULTS

4. • Haphazard urban expansion and industrialization affect
obtainable quality of life.
• Landscape epidemiology explores the relationship between
human health and urban environments.
• Spaceborne/Earth observation (EO) data provide perspectives
on urban area changes and their effects.
Figure: A tale of two different stories in the same city - Rio De Janeiro, Brazil (Source: Google images)
Urbanisation and Epidemiology

5. • Global multiband Earth observation between 0.4 µm—15 µm
wavelengths.
• Applications include urban targets detection, land cover
mapping, urban ecology analysis (e.g epidemiological processes)
• Multispectral: sparsely sampled bands (up to about 12);
Hyperspectral: densely sampled bands (up to about 220).
• Advantages of multispectral optical EO data include:
• less number of bands compared to hyperspectral —
meaning lower compute overhead;
• Possibility to extract earth surface information using
easy-to-compute normalised indices and thermal
infrared bands.
Urban remote sensing with optical EO data

6. • Improved EO data input (spatial and temporal) resolutions.
• Improved bespoke modeling and features extraction methods.
Modeling epidemiological risks with multispectral data

7. • Sample vector: Ae. aegypti mosquito species.
• Diseases carried: Zika, Dengue, Chikungunya viruses.
• Presence: over 100 countries.
• Informative environmental effects: Vegetation condition,
temperature, humidity, precipitation.
• Environmental effects can be extracted across the world from
EO data to derive models.
Specific application: Mosquito diseases risks modeling

8. Mission Available No of Spatial Temporal Thermal Free?
Since bands resolution (m) resolution (days) Infrared?
Landsat July 1972 8 30 16 Yes Yes
SPOT a Feb. 1986 6 (max) 1—4 No No
AVHRR b May 1998 6 1100 1 Yes Yes
MODIS c Dec. 1999 34 250 (max) 1 Yes Yes
GPM/TRMM d Nov. 1997 34 0.1° hourly - Yes
a
SPOT: Systeme Pour l’Observation de la Terre (French mission).
b
AVHRR: Advanced Very High Resolution Radiometer
(by US government’s National Oceanic and Atmospheric Administration (NOAA).
c
MODIS: Moderate Resolution Imaging Spectroradiometer (by NASA).
d
GPM/TRMM: Global precipitation mission Tropical Rain Measurement Mission (by NASA and JAXA)
• There is a spatial - temporal resolution trade-off.
• Studies apply different data based on their properties — spatial
and/or temporal modeling.
Spaceborne/EO data applied so far in this domain

9. Index Formula
Normalized Difference Vegetation Index (NDVI) NIR−Red
NIR+Red
Normalized Difference Water Index - (NDWI) NIR−SWIR
NIR+SWIR
Blue: Blue band (≈ 490 nm)
Red: Red band (≈ 700 nm)
NIR: Near infrared band (≈ 850 nm)
SWIR: Shortwave infrared band (≈ 1500 nm—2200 nm)
Spectral indices that are of interest

10. Study Sensor EO features Limitation(s)
Espinosa et al., 2016 SPOT 5 Vegetation map
Data: SPOT data are not free.
Hence, not study reproducible.
Water bodies
Spectral feature
Urban distribution
NBR
Scavuzzo et al., 2018
German et al., 2018
MODIS and NDVI
- Method:
performance—explainability
trade-off between statistical and
machine learning models.
TRMM/GPM NDWI
Daytime LST
Night-time LST
Precipitation
NBR: Normalised Burn Ratio – Temperature proxy.
NDVI: Normalised Difference Vegetation Index.
NDWI: Normalised Difference Water Index.
LST: Land surface temperature (obtained from thermal infrared bands).
Featured studies that apply EO data for
Ae. aegypti vector/diseases modeling

11. • Spatial models:
• Free high resolution EO data.
• Temporal models:
• Nowcasting — information arrives when it is already too
late.
• Spatio-temporal modeling.
Needed contributions

12. To propose a framework for using EO data for Dengue vector
population one-week-ahead forecasting in an urban area using
RNN.
Objective

13. • RNN uses recurrent connection to capture sequential
information. Makes it fit for time series prediction.
• Given X = (x1, x2, . . . , xT ) with xt ∈ Ru time series, a simple
RNN is defined:
ht = f (ht−1, xt), (1)
ht ∈ Rv : hidden state at time t.
• Due to vanishing gradients in RNN, long short term memory
(LSTM) variant has been developed.
RECURRENT NEURAL NETWORKS (RNN)

14. • Uses gating mechanisms to solve vanishing gradient.
• ht is obtained based on input xt as follows:
ft = σ(Wf [ht−1; xt] + bf ),
it = σ(Wi [ht−1; xt] + bi ),
ot = σ(Wo[ht−1; xt] + bo),
st = ft

15. st−1 + it

16. tanh(Ws[ht−1; xt] + bs),
ht = ot

17. tanh(st),
(2)
• W: weights, b: biases, σ: sigmoid,

18. : Hadamard product.
Background: LSTM

19. Methodology overview

20. • Cluster ground truth data by K-means approach.
• Given Ae. aegypti population across m locations:
Y = {y(1), y(2), . . . , y(m)}, y(i) ∈ RP;
• Partition Y into k clusters with centers
{c(1) . . . , c(k)}, c(i) ∈ RP;
• For n EO features means over k clusters as
X = {x(1), x(2), . . . , x(k×n)}, x(i) ∈ RP across time period P.
• Resulting model is defined a NARX model:
b
ct = F([ct−T , . . . , ct−1]; [xt−T , . . . , xt−1]), (3)
Methodology — Model

21. • An encoder-decoder LSTM architecture has been used its
success in time series forecasting.
• Encoder: an LSTM that maps the input into a learned
representation ht ∈ Rv
• Decoder:
• an LSTM that maps ht to decoder output; dt
• and a fully connected layer with ReLU activation which
takes dt as input and produces b
ct.
Methodology — LSTM architecture

22. Location Batch Date range Total Traps Differentiating
weeks condition
Vila Velha 2017 10/04/2017 - 31/12/2017 36 193 -
2018 02/01/2018 - 05/10/2018 40 325 Vector
control
Serra 2017 27/04/2017 - 30/12/2017 38 567 -
2018 05/01/2018 - 05/10/2018 40 95 Vector
control
• Two different locations, two batches (ground conditions) per
locations.
• 50% points per cluster split for training/test. 20% training time
points selected for validation.
Test areas and ground truth data.

23. • Find optimal number of mosquito data clusters (k) and obtain
clusters using k-means.
• k is chosen by elbow method to reduce distortion (J):
J =
k
X
j=1
m
X
i=1
x
(j)
i − c(j)
2
, (4)
• Find optimal lag T ∈ {3, 6, 9}.
• Find optimal rep.size, v ∈ {16, 32, 64, 128}.
• Compare model to RF and KNN equivalents.
• Metric: Mean absolute error (MAE).
Experimental procedure

24. Figure: Mean temporal distribution for clusters obtained
1 6 11 16 21 26 31 36
Week
0
1
2
3
Mosquito
population 1A
2A
3A
4A
5A
6A
(a) Vila Velha, 2017
1 6 11 16 21 26 31 36
Week
0
2
4
6
8
Mosquito
population
1B
2B
3B
4B
5B
6B
(b) Vila Velha, 2018
1 6 11 16 21 26 31 36
Week
0.0
0.5
1.0
1.5
2.0
Mosquito
population
1A
2A
3A
4A
5A
(c) Serra, 2017
1 6 11 16 21 26 31 36 41 46
Week
0
2
4
6
Mosquito
population
1B
2B
3B
4B
5B
(d) Serra, 2018

25. Vila Velha Serra
T ⇒ 3 6 9 3 6 9
2017
Training 0.3117 0.3392 0.4926 0.2254 0.1509 0.2451
Validation 0.4627 0.1810 0.3745 0.1985 0.3275 0.2729
Test 0.6120 0.6450 0.7565 0.4048 0.4703 0.5889
2018
Training 0.1407 0.2067 0.2762 0.2738 0.7999 0.2642
Validation 0.1998 0.4516 0.3395 0.2407 0.8574 0.3151
Test 0.3600 0.4624 0.4602 0.4418 0.9028 0.5372
Gomes et al., 2012 obtained its best dengue vector model in a
Brazilian city with T = 4 lagged effect of temperature and
precipitation. Hence, the result here (T = 3) is in line.
RESULTS: how many weeks of EO variables lagged effects?

26. Vila Velha Serra
v Year ⇒ 2017 2018 2017 2018
16 Training 0.3117 0.1407 0.2254 0.2738
Validation 0.4627 0.1998 0.1985 0.2407
Test 0.6120 0.3600 0.4048 0.4418
32 Training 0.2637 0.1501 0.1880 0.2652
Validation 0.4774 0.1975 0.2456 0.1817
Test 0.6274 0.4126 0.4231 0.3986
64 Training 0.2841 0.1781 0.1630 0.2459
Validation 0.4765 0.2231 0.1472 0.3632
Test 0.6203 0.3816 0.3984 0.4329
128 Training 0.2802 0.1808 0.2266 0.6732
Validation 0.3880 0.3189 0.2244 0.8454
Test 0.5767 0.4038 0.4440 0.5794
Lower representation dimension is obtained when encoding lower
variability.
RESULTS: Evaluating representation feature size

27. Vila Velha Serra
Model Year ⇒ 2017 2018 2017 2018
LSTM Training 0.2802 0.1407 0.1630 0.2652
Validation 0.3880 0.1998 0.1472 0.1817
Test 0.5767 0.3600 0.3984 0.3986
RF Training 0.1660 0.1300 0.1424 0.1743
Validation 0.8822 0.3920 0.3940 0.4709
Test 0.6348 0.4045 0.4502 0.5057
KNN Training 0.4237 0.3167 0.3067 0.3643
Validation 0.9755 0.5010 0.4165 0.6127
Test 0.6676 0.3869 0.4906 0.5000
Results: LSTM vs. RF vs. KNN

28. 0.20 0.40 0.59
0.40
0.59
Predicted
Cluster 1A
0.00 0.13 0.25 0.38
0.13
0.25
0.38
Cluster 2A
0.00 0.80 1.59 2.39
0.80
1.59
2.39
Cluster 3A
0.00 0.33 0.65 0.98
Observed
0.33
0.65
0.98
Predicted
Cluster 4A
0.00 0.48 0.96 1.44
Observed
0.00
0.48
0.96
1.44
Cluster 5A
0.00 1.39 2.79 4.18
Observed
0.00
1.39
2.79
4.18
Cluster 6A
KNN
LSTM
RF
Figure: 2017
0.65 1.30 1.94
0.65
1.30
1.94
Predicted
Cluster 1B
0.21 0.43 0.64
0.21
0.43
0.64
Cluster 2B
0.38 0.77 1.15
0.38
0.77
1.15
Cluster 3B
0.43 0.85 1.28
Observed
0.43
0.85
1.28
Predicted
Cluster 4B
0.00 1.56 3.12 4.68
Observed
1.56
3.12
4.68
Cluster 5B
0.19 0.28
Observed
0.19
0.28
Cluster 6B
KNN
LSTM
RF
Figure: 2018
• LSTM follows the highest and lowest observed values better.
Vila Velha: LSTM vs RF on test data

29. 0.14 0.28 0.43
0.14
0.28
Predicted
Cluster 1A
0.46 0.91 1.37
0.46
0.91
Cluster 2A
0.87 1.74 2.61
Observed
0.87
1.74
2.61
Cluster 3A
0.00 0.47 0.93 1.40
Observed
0.00
0.47
0.93
1.40
Predicted
Cluster 4A
0.34 0.68 1.01
Observed
0.34
0.68
1.01
Cluster 5A
KNN
LSTM
RF
Figure: 2017
0.00 0.65 1.29 1.94
0.00
0.65
1.29
1.94
Predicted
Cluster 1B
0.41 0.82 1.23
0.41
0.82
1.23
Cluster 2B
0.12 0.23 0.35
Observed
0.12
0.23
0.35
Cluster 3B
0.00 2.13 4.27 6.40
Observed
0.00
2.13
4.27
6.40
Predicted
Cluster 4B
1.07 2.14 3.21
Observed
1.07
2.14
3.21
Cluster 5B
KNN
LSTM
RF
Figure: 2018
• LSTM follows the highest and lowest observed values better.
Serra: LSTM vs RF on test data

30. Asides forecasting risks (e.g. early warning), the following
applications are reachable:
• Spatio-temporal gap filling, especially when a trap location with
missing data had previously been classified into a cluster.
• Data collection and collation resources optimization (e.g
rotation across zones).
• The method can serve operational vector control programs in
spatio-temporal gap filling and man-power optimisation.
• RNN’s powerful prediction capability enables more robust
modeling.
Possible applications

31. • The proposed EO data-based sub-municipal (spatio-temporal)
one-step-ahead forecasting framework shows robust performance
for the task at hand.
• The method can serve operational vector control programs in
spatio-temporal gap filling and man-power optimisation.
• RNN’s powerful prediction capability enables more robust
modeling.
• LIMITATION: The model is a blackbox. Next iterations can
consider explainability approaches to improve the utility of the
proposed methodology.
Conclusions

32. THANK YOU FOR LISTENING
...
Questions?