Enhanced Detection of Human-Driven Forest Alterations using Echo State Networks

Abstract

Forest monitoring is crucial for understanding ecosystem dynamics, detecting changes, and implementing effective conservation strategies. In this work, we propose a novel approach for automated detection of human-induced changes in woodlands using Echo State Networks (ESNs) and satellite imagery. Using ESNs offers a promising solution for analyzing time-series data and identifying deviations indicative of forest alterations, particularly those caused by human activities such as deforestation and logging. The proposed experimental setup leverages satellite imagery to capture temporal variations in the Normalized Difference Vegetation Index (NDVI) and involves the training and evaluation of ESN models using extensive datasets from Chile's central region, encompassing diverse woodland environments and human-induced disturbances. Our initial experiments demonstrate the effectiveness of ESNs in predicting NDVI values and detecting deviations indicative of human-related changes in woodlands, even in the presence of climate-induced changes like drought and browning. Our work contributes to forest monitoring by offering a scalable and efficient solution for automated change detection in woodland environments. Integrating ESNs with satellite imagery analysis provides valuable insights into human impacts on forest ecosystems, facilitating informed decision-making for sustainable land management and biodiversity conservation.

Full text

Enhanced Detection of Human-Driven Forest
Alterations using Echo State Networks
Tomás Couso Coddou
Department of Computer Science
Pontificia Universidad Católica de Chile
Santiago, Chile
tcouso@uc.cl
Paula Aguirre
Institute of Mathematical Engineering and Computation
Pontificia Universidad Católica de Chile
Santiago, Chile
paaguirr@uc.cl
Rodrigo A. Carrasco
Institute of Mathematical Engineering and Computation & School of Engineering
Pontificia Universidad Católica de Chile
Santiago, Chile
rcarrass@uc.cl
Javier Lopatin
Faculty of Engineering and Sciences
Universidad Adolfo Ibáñez
Santiago, Chile
javier.lopatin@uai.cl
Abstract
Forest monitoring is crucial for understanding ecosystem dynamics, detecting
changes, and implementing effective conservation strategies. In this work, we
propose a novel approach for automated detection of human-induced changes in
woodlands using Echo State Networks (ESNs) and satellite imagery. Using ESNs
offers a promising solution for analyzing time-series data and identifying deviations
indicative of forest alterations, particularly those caused by human activities such as
deforestation and logging. The proposed experimental setup leverages satellite im-
agery to capture temporal variations in the Normalized Difference Vegetation Index
(NDVI) and involves the training and evaluation of ESN models using extensive
datasets from Chile’s central region, encompassing diverse woodland environments
and human-induced disturbances. Our initial experiments demonstrate the effec-
tiveness of ESNs in predicting NDVI values and detecting deviations indicative
of human-related changes in woodlands, even in the presence of climate-induced
changes like drought and browning. Our work contributes to forest monitoring
by offering a scalable and efficient solution for automated change detection in
woodland environments. Integrating ESNs with satellite imagery analysis provides
valuable insights into human impacts on forest ecosystems, facilitating informed
decision-making for sustainable land management and biodiversity conservation.
Tackling Climate Change with Machine Learning: workshop at NeurIPS 2024. Link

1 Introduction
Monitoring forest ecosystems is crucial for addressing global change and mitigating ecosystem
degradation, as forests provide essential services such as water and carbon regulation, nutrient
cycling, air purification, biodiversity conservation, climate moderation, and recreational opportunities
(
16
;
19
;
20
;
2
;
4
;
7
;
32
). Deforestation, driven by urbanization and agricultural expansion, leads to
the destruction and fragmentation of habitats, disrupting vital ecological functions and threatening
biodiversity and sustainable land management (
29
;
5
;
18
;
45
). While the Amazon rainforest has
garnered significant attention, forests worldwide hold immense ecological, economic, and societal
value (35; 27; 10).
Natural disturbances such as fires and droughts also contribute to forest degradation, making it
challenging to differentiate between deforestation and other forms of degradation (
9
;
26
;
36
;
33
).
Effective landscape management and policy-making require robust monitoring systems that can
accurately detect these changes on a large scale, for which remote sensing provides a powerful tool.
The time series analysis of optical satellite imagery, combined with machine learning and statistical
techniques, has proven effective in detecting landscape changes (
38
;
25
;
17
;
44
). Notable examples
include the Continuous Change Detection and Classification algorithm (CCDC) and Breaks for
Additive Season and Trend (BFAST) (
46
;
43
;
24
). Deep learning methods have further enhanced our
ability to analyze complex, multi-dimensional remote sensing data, with CNNs and RNNs particularly
successful in this domain (39; 34; 6).
However, the scalability of CNNs and RNNs remains challenging, particularly regarding compu-
tational demands and data requirements (
15
). Recurrent neural networks (RNNs), despite their
ability to model long-term temporal dependencies, face difficulties when adapting to changes in data
distribution, a critical factor in dynamic environments like forests (
23
;
37
;
11
). In this context, Echo
State Networks (ESNs), a type of RNN, offer significant advantages. ESNs can efficiently adapt to
new data and predict chaotic time dynamics. Their training process is faster and less prone to the
gradient problem compared to other RNNs, as only the readout layer is trained (
13
;
14
;
21
;
22
;
28
).
ESNs have been successfully applied in various domains, such as fault detection in industrial systems,
yet their application in forest change detection remains underexplored (3; 30).
Our research addresses this gap by developing an ESN-based system for detecting large-scale forest
changes, explicitly focusing on central Chile. This region has experienced severe droughts and fires
over the past decade, leading to significant tree mortality in a highly heterogeneous landscape of
forestry and agricultural mosaics, making it an ideal testing ground for this approach (9; 26).
2 Methods
2.1 Data selection, processing, and training
The study area covers approximately 78,000 km
2
in central Chile (latitudes from -36.5°S to -32.2º S),
a region recognized as one of the global biodiversity hotspots (
27
). Local forests are confronted with
anthropogenic pressures such as urbanization, agricultural encroachment, and the introduction of non-
native plant species. In the last decade, the region has experienced extreme drought conditions that
cause widespread "browning" phenomena (
9
;
26
), and intense wildfire activity, further exacerbating
the stress on these vital ecosystems (36).
Our analysis is based on a Landsat multispectral dataset comprising 8,804 scenes captured between
January 2000 and June 2022 with the TM, ETM
+
, and OLI/TIRS sensors aboard Landsat 5, 7, 8,
and 9 satellites (
41
). For each pixel in the study region, we calculated the normalized difference
vegetation index (NDVI) (
31
) using the red and near-infrared bands in each collection and obtained
NDVI time series across all observed dates. To address the significant noise and missing values in
the raw NDVI signal, we developed a three-step “denoising algorithm" consisting of i)resampling
of the NDVI signal to bi-weekly intervals, ii) application of a standard deviation filter with linear
interpolation of missing values, and iii)a Holt-Winters exponential smoothing (12).
For training of the change-detection algorithm, we identified areas that likely experienced different
forms of deforestation between 2016 and 2022 using records from the Global Forest Watch (GFW)
dataset (
1
). We relied on the visual interpretation of Google Earth Pro imagery for detailed demar-
cation of degraded polygons. In total, we selected 382 sites. Of these, 142 sites were linked to
2

MAPE MAE MSE R²
non-feedback ESN 0.0671 0.0313 0.0021 0.9115
feedback ESN 0.0732 0.0379 0.0028 0.8462
Table 1: Regression Metrics for non-feedback and feedback ESNs.
human-driven deforestation, while 147 sites corresponded to other changes: 91 areas were affected by
the severe drought and vegetation “browning" in the summer of 2019–2020 (
26
), and 56 sites affected
by fires. Additionally, we selected 93 sites with stable time series where no changes occurred. We
used pixel-based data for our analyses, resulting in 23,053 individual pixel-based time series across
all sites. Further details on the dataset can be found in (8) and in A.3.
2.2 Echo state networks
We trained an Echo State Network (ESN) to predict the expected evolution of the NDVI signal for
each pixel. An ESN is a recurrent neural network (RNN) characterized by a reservoir of sparse,
randomly initialized, fixed weights (
21
), in which only the weights to output units are modified
for achieving the desired learning task(
13
). An additional description of an ESN can be found
in Appendix A.1. We assume that the generative process underlying the NDVI signal is dynamic
since vegetation phenological cycles may vary over time. In such scenarios, an online learning
rule like Recursive Least’s Squares (RLS) is suitable (
21
). The ESN training instances consisted
of feature-target pairs derived from the NDVI signal. We defined a feature vector of 104 NDVI
values (approximately two years of data), with the values of week 105 designated as the target values.
This window was then shifted sequentially across the entire time series for each pixel, resulting in
104-dimensional NDVI feature vectors. This process, which we refer to as “signal featurization",
involves taking the previous two years for each signal value, so the first target value fed into the net
corresponds to the beginning of the third year of available NDVI signal. For training the ESN, we
randomly selected 30% of the stable and drought polygons, setting aside the remaining instances to
validate the fault detection procedure. Further details on the training procedure are given in A.2.
2.3 Change detection algorithm and model validation
To detect human-driven forest alterations, we adapted the fault detection scheme from (
3
) to analyze
changes in the NDVI time series for each pixel. First, the raw NDVI signal is de-noised and
transformed into feature-target pairs of 104 features. Then, the series is divided into two parts: the
training section, comprising NDVI values from the beginning of the satellite observations up to a
year before the final datum, and the hidden section, which consists of the last value of the signal.
The ESN is fine-tuned in the training section and used to predict the hidden section. We refer to this
prediction as the signal’s reference lower bound. Then, the hidden section of the signal is compared
to the lower bound using time-shifted predictions. This is achieved by taking a unit from the second
section, feeding it into the fine-tuned ESN, and predicting the remaining values. A pixel is classified
as changed if
N
consecutive forecasted values of the time-shifted signal are lower than the lower
bound multiplied by a constant integer
k
. If no change is detected, the signal is shifted forward by
one unit, and the procedure is repeated until a change is detected or no remaining signal is left. We
used a voting mechanism to detect changes in entire polygons: a polygon is labeled as changed if the
percentage of pixels with a positive flag is greater than or equal to the threshold parameter th.
Appendix A.3 provides detailed information about the ESN dataset used. The data reserved for testing
the change detection algorithm encompassed the remaining 70% of the stable and drought polygons
(non change category), and the complete selection of the fire and drought polygons (change category).
For each pixel, we used the change detection algorithm with thresholds
th ∈ {0.25,0.5,0.75}
,
Number of consecutive failures
N
ranging from
3
to
18
and lower bound coefficient
k
ranging from
0.8
to
1
. The tracked metrics were the Accuracy, the F1-score, the Precision, and the Recall. The
details of this analysis can be found in appendix A.3.
3

3 Results
Table 1 shows the forecasting metrics obtained for both models. Results show that the non-feedback
model achieved superior performance, with all error metrics lower than the feedback model and with
a higher coefficient of determination.
Figure 1 shows two example polygons before and after the change event, with the algorithm’s
detection tagged. Appendix A.4 details the tuning parameters used for the detection.
Before event After event Detections
Figure 1: Example of fire (top row) and logging (bottom row) polygons. Green dots indicate true
positives, while red dots indicate false positives for individual pixels.
Table 2 shows the classification metrics, where an accuracy of 0.708 was achieved. These metrics
highlight the model’s capacities for capturing true positives, with a recall of 0.823. It is also evident
that there is a tendency to incur false positives, with a precision of 0.619. Appendix A.4 has further
details on the classification metrics.
To further distinguish the results according to the type of disturbance, we analyzed the classification
metrics on four subsets of the polygons according to the type of non-change category (stable or
drought) and change category (logging or fire). Table 3 presents the detailed classification metrics,
and reveals an uneven performance across the subsets of polygons. The detection of logging and
fire events over stable polygons reached an accuracy of 0.926 and 0.871, respectively, but lower
detection scoers are obtained for polygons affected by drought(0.609 for logging and 0.478 for fire).
It is noticeable how precision scores rise considerably without the influence of drought-related false
positives, with a score of 0.979 for stable polygons and 0.950 for fire.
4 Discussion
The results presented here demonstrate the capabilities of ESNs for capturing forest dynamics and
detecting change events in forests. When detecting over stable polygons, ESNs had an overall
accuracy of 0.708, 0.926 for fires, and 0.871 for logging events. For polygons affected by drought,
Precision Recall F1-Score Support
Non change 0.825 0.623 0.710 106
Change 0.619 0.823 0.707 79
Accuracy 0.708 0.708
Macro Avg 0.722 0.723 0.708 185
Weighted Avg 0.737 0.708 0.708 185
Table 2: Classification Metrics for the global model.
4

Non-Change Change Accuracy F1-Score Recall Precision
stable logging 0.926 0.929 0.885 0.979
stable fire 0.871 0.809 0.704 0.950
drought logging 0.609 0.672 0.885 0.541
drought fire 0.478 0.447 0.704 0.328
Table 3: Metrics for different cases of vegetation
the accuracy in change detection was comparatively lower, with a score of 0.609 for logging and
0.478 for fire events.
Our current work focuses on reducing the rate of false positives for drought polygons. One promising
venue for this matter is adjusting the change detection criteria. Hence, it considers the prediction
error of each particular pixel when determining the extent to which different predicted versus real
NDVI values consist of a change in the land cover. Possible ways to implement this are in the change
detection criteria defined by (
46
). Another improvement consists of evaluating the model to detect the
precise dates of the events. Doing so frames the problem as one of Change Point Detection (CPD),
where the objective is to determine time points where a time series changes its state(
42
). Suitable
metrics that can be used for CPD are shown in (42).
Acknowledgments and Disclosure of Funding
Funding
We would like to acknowledge the financial support provided by FONDEF ID21I10102 and FONDE-
CYT 1231245 Grants from ANID. Paula Aguirre also acknowledges funding from CENIA. Tomás
Couso Coddou received funding from CENIA to attend the conference.
Competing interests
There are no competing interests to disclose.
References
[1] Global forest watch, 1997.
[2]
BOTTALICO, F., TRAVAGLINI, D., CHIRICI, G., GAR FÌ, V., GIANNETTI, F., DEMARCO,
A., FARES, S., MARC HE TT I, M., NOCENTINI, S., PAOL ET TI, E., ET AL. A spatially-explicit
method to assess the dry deposition of air pollution by urban forests in the city of florence, italy.
Urban Forestry & Urban Greening 27 (2017), 221–234.
[3]
CHO, A. D., CARRASCO, R. A., RUZ, G. A., AND ORTIZ, J. L. Slow degradation fault
detection in a harsh environment. IEEE Access 8 (2020), 175904–175920.
[4]
DE OLIVEIRA ROQUE, F., MENEZES, J. F., NO RTHFIELD, T., OCHOA-QUI NT ERO, J. M.,
CAMPBELL, M. J., AND LAURA NC E, W. F. Warning signals of biodiversity collapse across
gradients of tropical forest loss. Scientific Reports 8, 1 (2018), 1–7.
[5]
DEFRIES, R. S., RUDEL, T., UR IARTE, M., AND HANSEN, M. Deforestation driven by urban
population growth and agricultural trade in the twenty-first century. Nature Geoscience 3, 3
(2010), 178–181.
[6]
DU, B., YUA N, Z., BO, Y., AND ZHA NG , Y. A combined deep learning and prior knowledge
constraint approach for large-scale forest disturbance detection using time series remote sensing
data. Remote Sensing 15, 12 (2023).
[7]
FIGUEROA, E., AN D PASTEN, R. The economic value of forests in supplying local climate
regulation. Australian Journal of Agricultural and Resource Economics 59, 3 (2015), 446–457.
5

[8]
FUENTES, I., LO PATIN, J., GALLEGUILLOS, M., CEBALLO S-COMISSO, A., EYHERAMENDY,
S., AND CARRASCO, R. Is the change deforestation? using time-series analysis of satellite data
to disentangle deforestation from other forest degradation causes. Remote Sensing Applications:
Society and Environment (2024), 101210.
[9]
GARREAUD , R. D., BOISIER, J. P., RONDANELLI, R., MONTECINOS, A., SEP ÚLVEDA ,
H. H., AND VELOS O-AGUILA, D. The central chile mega drought (2010–2018): a climate
dynamics perspective. International Journal of Climatology 40, 1 (2020), 421–439.
[10]
HAMILTON, S. E., AND FRIESS, D. A. Global carbon stocks and potential emissions due to
mangrove deforestation from 2000 to 2012. Nature Climate Change 8, 3 (2018), 240–244.
[11]
HUYEN, C. Designing Machine Learning Systems: An Iterative Process for Production-ready
Applications. O’Reilly Media, Incorporated, 2022.
[12]
HYNDMAN, R., AND ATHANASOPOULOS, G. Forecasting: Principles and Practice, 2nd ed.
OTexts, Melbourne, Australia, 2018.
https://otexts.com/fpp2/
Accessed on 2024-05-01.
[13]
JAEGER, H. The “echo state” approach to analysing and training recurrent neural networks.
GMD-Report 148, German National Research Institute for Computer Science (01 2001).
[14]
JAEGER, H., A ND HA AS , H. Harnessing nonlinearity: Predicting chaotic systems and saving
energy in wireless communication. Science 304, 5667 (2004), 78–80.
[15]
JEON, W., KO, G., LEE, J., LEE , H., HA, D., AND RO, W. W. Chapter six - deep learning
with gpus. In Hardware Accelerator Systems for Artificial Intelligence and Machine Learning,
S. Kim and G. C. Deka, Eds., vol. 122 of Advances in Computers. Elsevier, 2021, pp. 167–215.
[16]
KOFINAS, G. P. Adaptive co-management in social-ecological governance. In Principles of
ecosystem stewardship. Springer, 2009, pp. 77–101.
[17]
LARY, D. J., ALAVI, A. H., GANDOMI, A. H., AND WALKE R, A. L. Machine learning in
geosciences and remote sensing. Geoscience Frontiers 7, 1 (2016), 3–10.
[18]
LEBLOIS, A., DAMETTE, O., AND WOLFERSBERGER, J. What has driven deforestation
in developing countries since the 2000s? evidence from new remote-sensing data. World
Development 92 (2017), 82–102.
[19]
LÖF, M., MADSEN, P., METSLAID, M., WITZELL, J., A ND JACOBS, D. F. Restoring forests:
regeneration and ecosystem function for the future. New Forests 50, 2 (2019), 139–151.
[20]
LONGO, M., SA ATCHI, S., KELLER, M., BOWMAN, K., FERRAZ, A., MOORCROFT, P. R.,
MORTON, D. C., BONAL, D., BRANDO, P., BURBAN , B., ET AL. Impacts of degradation
on water, energy, and carbon cycling of the amazon tropical forests. Journal of Geophysical
Research: Biogeosciences 125, 8 (2020), e2020JG005677.
[21]
LUKOŠEVI ˇ
CIUS, M. A Practical Guide to Applying Echo State Networks. Springer Berlin
Heidelberg, Berlin, Heidelberg, 2012, pp. 659–686.
[22]
LUKO ŠEVI ˇ
CIUS, M., AND JAEGER, H. Reservoir computing approaches to recurrent neural
network training. Computer Science Review 3, 3 (2009), 127–149.
[23]
LYU, H., LU, H., AND MOU, L. Learning a transferable change rule from a recurrent neural
network for land cover change detection. Remote Sensing 8, 6 (2016).
[24]
MASILI
¯
UNA S, D., TSENDBA ZAR, N.-E., HEROLD, M., AN D VER BESSELT, J. BFAST lite: A
lightweight break detection method for time series analysis. Remote Sens. (Basel) 13, 16 (Aug.
2021), 3308.
[25]
MCALLISTER, E., PAYO, A., NOVEL LI NO , A., DOLPHIN, T., AND MEDINA-LOPEZ, E.
Multispectral satellite imagery and machine learning for the extraction of shoreline indicators.
Coastal Engineering (2022), 104102.
6

[26]
MIRANDA, A., LARA, A., ALTAMIRANO, A., DIBELLA, C., GONZÁLEZ, M. E., AN D
CAMARERO, J. J. Forest browning trends in response to drought in a highly threatened
mediterranean landscape of south america. Ecological Indicators 115 (2020), 106401.
[27]
MYERS, N., MITTERMEIER, R. A., MITTERMEIER, C. G., DAFONSECA, G. A., AND KENT,
J. Biodiversity hotspots for conservation priorities. Nature 403, 6772 (2000), 853–858.
[28]
PASC AN U, R., MIKOL OV, T., AND BENGIO, Y. On the difficulty of training recurrent neural
networks, 2013.
[29] RITCHIE, H., AND ROSER, M. Forests and deforestation. Our World in Data (2021).
[30]
RODRÍGUEZ-OSSORIO , J. R., MOR ÁN , A., ALONSO, S., PÉR EZ , D., DÍAZ, I., AND
DOMÍNGUEZ, M. Echo state networks for anomaly detection in industrial systems. IFAC-
PapersOnLine 56, 2 (2023), 1472–1477. 22nd IFAC World Congress.
[31]
ROUSE JR, J., HAA S, R., SCHELL, J., AND DEERI NG , D. Paper a 20. In Third Earth Resources
Technology Satellite-1 Symposium: The Proceedings of a Symposium Held by Goddard Space
Flight Center at Washington, DC on (1973), vol. 351, p. 309.
[32]
SÁNCHEZ, J. J., MARCOS-MARTINEZ, R., SRI VASTAVA , L., AND SOONSAWAD, N. Valuing
the impacts of forest disturbances on ecosystem services: An examination of recreation and
climate regulation services in us national forests. Trees, Forests and People 5 (2021), 100123.
[33]
SEBA LD, J., SENF, C., AND SEIDL, R. Human or natural? landscape context improves the
attribution of forest disturbances mapped from landsat in central europe. Remote Sensing of
Environment 262 (2021), 112502.
[34]
SHEN, R., HUANG, A., LI, B., AND GUO, J. Construction of a drought monitoring model
using deep learning based on multi-source remote sensing data. International Journal of Applied
Earth Observation and Geoinformation 79 (2019), 48–57.
[35]
SILVA JUNIOR, C. H., PE SS ÔA , A. C., CARVALHO , N. S., REIS, J. B., ANDERSON, L. O.,
AND ARAGÃO, L. E. The brazilian amazon deforestation rate in 2020 is the greatest of the
decade. Nature Ecology & Evolution 5, 2 (2021), 144–145.
[36]
SMITH-RAMÍREZ, C., CASTILLO-MANDUJANO, J., BE CERRA, P., SANDOVAL, N.,
FUENTES, R., AL LENDE, R., AND ACUÑA, M. P. Combining remote sensing and field
data to assess recovery of the chilean mediterranean vegetation after fire: Effect of time elapsed
and burn severity. Forest Ecology and Management 503 (2022), 119800.
[37]
SONG, A., CHOI, J., HA N, Y., AN D KIM, Y. Change detection in hyperspectral images using
recurrent 3d fully convolutional networks. Remote Sensing 10, 11 (2018).
[38]
THAKUR, S., MONDAL, I., GHOSH, P., DA S, P., AND DE, T. A review of the application of
multispectral remote sensing in the study of mangrove ecosystems with special emphasis on
image processing techniques. Spatial Information Research 28, 1 (2020), 39–51.
[39]
TORRES, D. L., TURNES, J. N., SOTO VE GA, P. J., FEITO SA, R. Q., SILVA, D. E., MAR-
CATO JUNIOR, J., AND ALMEIDA, C. Deforestation detection with fully convolutional networks
in the amazon forest from landsat-8 and sentinel-2 images. Remote Sensing 13, 24 (2021).
[40]
TROU VAIN , N., PE DRELLI, L., DINH, T. T., AND HINAUT, X. ReservoirPy: an Efficient and
User-Friendly Library to Design Echo State Networks. In ICANN 2020 - 29th International
Conference on Artificial Neural Networks (Bratislava, Slovakia, Sept. 2020).
[41]
USGS. Landsat 8-9 collection 2 (c2), level 2 science product (l2sp) guide, 2022. Accessed on
2023-11-13.
[42]
VAN DEN BURG, G. J. J., A ND WILLIAMS, C. K. I. An evaluation of change point detection
algorithms, 2022.
7

[43]
WU, L., LI, Z., LIU , X., ZHU, L., TANG , Y., ZHANG, B., XU, B., LI U, M., MENG, Y., AND
LIU, B. Multi-type forest change detection using BFAST and monthly landsat time series for
monitoring spatiotemporal dynamics of forests in subtropical wetland. Remote Sens. (Basel) 12,
2 (Jan. 2020), 341.
[44]
YIN, H., PR IS HC HEPOV, A. V., KU EMMERLE, T., BLEYHL, B., BUCHNER, J., AND RADE-
LOFF, V. C. Mapping agricultural land abandonment from spatial and temporal segmentation
of landsat time series. Remote sensing of environment 210 (2018), 12–24.
[45]
ZEMP, D., SCHLEUSSNER, C.-F., BARBOS A, H. D. M. J., AND RAMMIG, A. Deforestation
effects on amazon forest resilience. Geophysical Research Letters 44, 12 (2017), 6182–6190.
[46]
ZHU, Z., AND WOODCOCK, C. E. Continuous change detection and classification of land
cover using all available landsat data. Remote Sens. Environ. 144 (2014), 152–171.
8

A Appendix
A.1 Echo state networks
The reservoir activation vector
x(n)
is updated according to Equations 1 and 2. At each time step
n
, the interim activation vector
ex(n)
is calculated by summing the input vector activation
Winu(n)
and the prior activation
W x(n−1)
, then applying an activation function to the result. Optionally, a
feedback layer
Wfb
may also be incorporated to include the previous output vector in the state vector
computation. The new activation vector
x(n)
is subsequently determined through a leaky integration
of
ex(n)
and
x(n)
, with a leaking rate of
α
. Finally, the output vector
y(n)
is calculated through the
readout layer in Eq. 3.
ex(n) = tanh(Winu(n) + W x(n−1) + Wfb y(n−1)) (1)
x(n) = (1 −α)x(n−1) + αex(n)(2)
y(n) = Wout[u(n); x(n)] (3)
A.2 Training of the echo state network
For training the ESN, we randomly selected 30% of the stable and drought polygons, setting aside
the remaining instances for the fault detection procedure. The training dataset of the ESN used the
first ten years of instances of each pixel (520 feature-target pairs), roughly 50% of the ESN dataset,
and the remaining ten years of values were set aside for testing. We implemented the ESN using the
ResevoirPy python library (
40
). We set the reservoir size to 500 units, the spectral radius 0.9, and
the leaking rate to 0.5. We trained two ESN variations in the NDVI forecasting task: a plain-vanilla
ESN without feedback connections and an ESN with feedback connections. For validating the trained
models, we computed the mean absolute percentage error (MAPE), the mean absolute error (MAE),
the mean squared error (MSE), and the coefficient of determination (R2).
A.3 Data and validation
Table 4 shows the ESN dataset in detail.
Change Type Number of Polygons Number of Pixels
Stable 24 780
Drought 27 2405
Table 4: Echo state network training dataset
Table 5 shows the change in the detection task of the dataset in detail, with corresponding labels for
each category for the change detection task.
Change Type Number of Polygons Number of Pixels Label
Stable 43 1119 0
Drought 63 4234 0
Logging 52 2064 1
Fire 27 4489 1
Table 5: Change detection dataset, corresponding classification task labels for each category.
A.4 Supplemental Results
Figure 2 shows a grid of F1-scores for our change detection algorithm for all the parameter configura-
tions tested. It can be seen that the best parameters follow a diagonal pattern for all voting thresholds,
where lower values of
k
favored lower values of N, and higher values of
k
favored higher values of
N
.
The optimal model used the values
th = 0.75
,
N= 3
, and
k= 0.81
, which achieved an F1-score of
0.708.
9

Figure 2: F1-scores for each parameter configuration tested. Darker sections of the grids indicate a
higher F1 score, meaning a better balance between Type I and Type II errors.
Table 6 shows the confusion matrix to classifications of the change detection procedure under the
optimal parameters. The results are detailed for each possible set of polygons according to the Non-
Change and Change categories. It can be seen that the algorithm achieved a high overall performance,
with false negatives of only 7.5%. There is a higher tendency for false positives, with a value of
21.6%, which can be explained mainly by mistakes detecting changes in polygons with drought.
Predicted: Non-Change Predicted: Change
Actual: Non-Change Stable 42 (22.7%) 1 (0.5%)
Drought 24 (13.0%) 39 (21.1%)
Actual: Change Logging 6 (3.2%) 46 (24.9%)
Fire 8 (4.3%) 19 (10.3%)
Table 6: Confusion matrix for native vegetation detailed by polygon type.
10