Enhanced Landslide Detection using Deep Learning with Recurrent Neural Networks for Temporal Monitoring

S Mahalakshmi K.Valarmathi P A Mathina

Department of Electronics and Department of Electronics and Department of Electronics and
Communication Engineering Communication Engineering Communication Engineering
P.S.R. Engineering College P.S.R. Engineering College P.S.R. Engineering College
Sivakasi, India Sivakasi, India Sivakasi, India
[email protected] [email protected] [email protected]

B Mohan S.Deepa R.Jayasri

Department of Electronics and Department of Electronics and Department of Electronics and
Communication Engineering Communication Engineering Communication Engineering
P.S.R. Engineering College P.S.R. Engineering College P.S.R. Engineering College
Sivakasi, India Sivakasi, India Sivakasi, India
[email protected] [email protected] [email protected]
S.Dhula
Department of Electronics and
Communication Engineering
P.S.R. Engineering College
Sivakasi, India
[email protected]
 Abstract—Landslides are one of the major natural hazards, enabling authorities to take preventive measures like
causing significant impairment to structure and defeat of social evacuating the affected population and deploying emergency
survives. Early and accurate landslide prediction is vital to response teams before a landslide occurs. So, predictive
mitigate such risks. This work introduces an advanced landslide models based on IoT and AI have shown promising results
alert system leveraging a deep learning model, specifically a for disaster preparedness, reducing the time taken to respond
Recurrent Neural Network (RNN). The system is trained on
historical landslide data alongside real-time sensor inputs, to an event, and providing critical information in high-risk
including critical parameters including soil moistness, areas. As these systems evolve, they are likely to become an
temperature, moisture, in addition ground vibration. The RNN integral part of early warning infrastructures that will help
model effectively captures temporal patterns, enabling precise make communities safer and minimize landslide-related
probability calculations for landslide occurrence. The casualties and damages. The challenge lies in refining these
accomplished RNN method is assessed on unseen test data to technologies to ensure low latency, reliability, and
validate its predictive accuracy and generalization capability. scalability, so they can be effectively implemented in diverse
Standards like accuracy, recall, precision, and F1 score are used and rugged terrains where the threat of landslides is always
to measure its performance. The findings of the evaluation present.
indicate that the Recurrent Neural Network demonstrates
superior performance in the detection of landslides, attaining an
exemplary accuracy rate of 89%. The model records a precision II. LITERATURE REVIEW
rate of 82%, an F1-score of 87%, and a recall rate of 94%.
Lu, et al.,[2011][6] illustrate a new image-grounded
Keywords—Landslide prediction, Deep Learning, Recurrent
landslide tracking structure projected in this work which
Neural Network, Sensor data, Predictive accuracy utilizes digital image progressions styles toward descry
landslide passed in the least landslide passed condition to
attain a non-interface landslide tracking scheme by unique
I. INTRODUCTION image. The proposition of this method is grounded on
consuming a ray designer to venture a ray toward the
Landslides are amongst the biggest formidable natural dimension container, which has two acrylic boards in the
threats such occur worldwide, often bringing disastrous dimension container. The dual ray adverts in the image is
results. A landslide may be illustrated as the motion of a induced through the passed landslide. To recognize the
batch of rocks global down an incline, and is usually induced match of ray adverts, and complete a laidback computation,
by the combination of environmental factors, which include the method might contract the landslide relegation with a
massive rainfall, earth-shaking action, volcanic explosion, neutralized position. As stated by the dimension effect, it
then human interventions such as deforestation and might advertise the caution in addition to helping the
unplanned urban development [1]. In landslide-prone areas, impairment.
these events result in substantial economic losses,
infrastructural destruction, and even human loss in the worst Doerksen et al.,[2023][7] recommend to usage of the
scenarios. The unpredictability of landslides, along with their authority of machine learning, deep learning with artificial
sudden onset, makes these phenomena highly difficult to intelligence ways through exposed cause. Space grounded
mitigate. With growing urbanization into vulnerable terrains information to prognosticate landslides in the district position
and extreme weather events becoming increasingly frequent in Nepal at 7, 10, and 14-day time-based judgments through
due to climate change, there is an urgent need for precise and the standardized rush estimations and geomorphic statistics
timely landslide prediction [2]. The traditional landslide as a contribution. Outcomes were given in cooperation with
monitoring techniques have usually been based on geological technical sapience through point significance investigation
survey and visual observation however these techniques are then a robust prophetic competence of landslide vaticination
time consuming, laborious, and rather unsuitable for in Nepal by means of random forest in addition U Net
instantaneous reactions. In addition, the extreme variability models.
of environmental factors involved in landslide initiation Ado et al.,[2022[[8] explains the numerous scholarly
means that a highly localized approach is essential to capture investigations have meticulously examined the claim of
the unique conditions in different geographical areas[3]. machine learning frameworks for the purpose of landslide
Recent advancements in sensor technology, artificial vulnerability mapping, underscoring their efficacy in
intelligence (AI), and big data analytics have unlocked identifying regions with elevated risk. Empirical evidence
recently developed potentiality during enhancing landslide indicates that the identification of landslide causative factors
prediction. Through IoT devices and deep learning models, (LCFs) is contingent upon the availability of data along with
researchers and engineers are now able to monitor and the distinctive attributes of the specific geographical area
explore an extensive parameter in real time[4]. This IoT being analyzed. Research findings illustrate that China
sensor are placed strategically within landslide-prone areas, emerges as the predominant locus of inquiry, with the area
which continue to collect environmental data like under the receiver operating characteristic curve regarded as
temperature, soil moisture, moisture, ground vibrations with the favored metric for evaluation. Recent innovations in
even rainfall intensity. These data points, collected from hybrid, ensemble, and deep learning techniques exhibit
various sources, present an image of the environmental enhanced performance relative to conventional ML
conditions that could cause landslides. Deep learning and methodologies. The escalating scholarly engagement in this
machine learning algorithms, especially RNN can be used to domain signifies a discernible trend towards the adoption of
analyze such complex datasets. RNNs are specifically increasingly sophisticated AI-driven approaches for landslide
designed for time-series analysis and can identify patterns forecasting.
and trends before a landslide. By training these models on
historical landslide events, and correlating them with real- Raut et al.,[2024][9] presents an investigation that
time data, landslide risk can be predicted more accurately primarily emphasizes the comparative assessment of
well in time [5]. These can warn authorities and communities machine learning and statistical methodologies for the
of an impending landslide many days in advance, thereby purpose of landslide susceptibility mapping; however, it
neglects to consider the incorporation of supplementary learning and deep learning practices. The identification of
advanced geospatial techniques or hybrid models that could landslides is pivotal for comprehensive risk evaluation. To
potentially augment predictive accuracy and furnish a more augment model efficacy, scholars have developed integrated
holistic understanding of the dynamics associated with geodatabases, including the Recent Landslide Database
landslides. (RecLD) and the Relict Landslide Database (RelLD).
Convolutional neural networks were ascertained to exhibit
Nair [2024][10] articulates that conventional landslide
the highest accuracy (92.5%) in a instance education
prediction frameworks frequently depend on localized
executed in Lantau, Hong Kong, attributable to their
datasets and substantial computational resources, which
advanced feature extraction proficiency. In evaluating
constrains their applicability on a global scale and culminates
effectiveness, logistic regression, SVM, RF, and boosting
in a notably low accuracy rate of approximately 30%. This
algorithms are ranked just after. These findings illustrate the
underscores the imperative for the development of more
potential of ML and DL frameworks in facilitating landslide
efficacious models that can function on an expansive scale
identification and mitigating associated risks.
devoid of such limitations.
Chaganti et al.,[2023][15] elaborates the purposes of
Miao et al.,[2023][11] explains the effort to enhance the
disaster prevention and risk mitigation, the mapping of
predicted accuracy of landslide susceptibility mapping,
landslide detection is of paramount importance, and
contemporary investigations have employed machine
contemporary research has been focused on the refinement of
learning algorithms alongside remote sensing technologies.
analytical methodologies. Through the utilization of high-
The integration of InSAR deformation data has demonstrated
resolution imaging coupled with remote sensing
a capacity to bolster model efficacy, with collaborative
technologies, investigators have implemented object
training yielding the most favorable results. When evaluating
detection and image segmentation methodologies.
machine learning algorithms, random forest demonstrates
Observations from a study suggest that utilizing a pretrained
greater accuracy against information value, convolutional
VGG 16 neural network alongside a logistic regression
neural networks, and support vector machines. Empirical
classifier leads to the greatest detection rates for floods and
studies indicate that the inclusion of InSAR data significantly
landslides. To enhance risk assessment, the accuracy of
enhances landslide susceptibility mapping while
forecasts is further augmented through the integration of
concurrently reducing the incidence of false positives. These
aerial imagery and field validation. In the realm of effective
findings provide crucial insights for the realm of landslide
disaster management, advancements in deep learning and
risk assessment research and its practical applications.
remote sensing are significantly improving the monitoring of
Casagli et al.,[2023][12] illustrates the purposes of floods and the detection of landslides.
landslide detection, monitoring, and prediction, remote
sensing techniques (RSTs) are extensively utilized due to III. PROPOSED SYSTEM
their provision of critical information across diverse temporal
and spatial dimensions. While terrestrial methodologies such The proposed scheme is perfect for applications such as
as lidar, Doppler radar, and interferometric radar deliver natural language processing (NLP), time series estimating
exceptional precision and real-time surveillance, satellite- then classification or regression tasks involving temporal or
based RSTs afford extensive geographical coverage. sequential patterns since it is meticulously designed to
Empirical studies underscore that each RST possesses handle sequential data quickly. Real-world issues like
distinct advantages and limitations; thus, the integration of landslide detection, which call for the analysis of sequential
these methods is imperative for a comprehensive analysis of geographic and temporal data including rainfall patterns, soil
landslides. The amalgamation of various technologies moisture levels, seismic activity, and other environmental
enhances the efficacy of early warning systems and risk parameters, benefit greatly from its application. The
management strategies. Future investigative efforts should sequential flow of landslides is shown in Fig. 1.
prioritize the application of big data analytics and the The first step in the procedure is gathering raw input
enhancement of communication with at-risk populations to data, which could include satellite photos, sensor numerical
facilitate more effective landslide mitigation. values, or other organized or unstructured landslide
Meena et al.,[2022][13] infer event based landslide prediction data. After that, this data undergoes a thorough
inventories play a pivotal role in aiding hazard and preparation step to make sure it is appropriate for model
susceptibility assessments by elucidating the correlation training. Since landslides occur far less frequently than non-
between triggering phenomena and landslide occurrences. landslide occurrences, class imbalance is a major problem in
Current explorations have involved satellite imagery and landslide datasets. To remedy this, methods including
elevation data to delve into the possibilities of machine creating synthetic samples, under sampling the dominant
learning and deep learning frameworks, particularly U-Net, class, or oversampling the minority class are employed to
for the automated recognition of landslides. An analysis balance the dataset. By doing this, the model is kept from
performed in the Nepali Himalayas illustrates that U-Net becoming skewed toward non-landslide occurrences. To
reached the utmost accuracy when using a 128×128 pixel improve convergence during training, the information is
patch size, surpassing older machine learning models like further standardized by scrabbling the features to a consistent
SVM and Random Forest (RF) . While the incorporation of orbit, usually between 0 and 1 to facilitate better
digital elevation models did not significantly enhance overall convergence during training and improve overall model
accuracy, it did contribute to improved classification performance.
outcomes. U-Net remains a promising yet underexplored The preprocessing stage have a separation of dataset into
approach for landslide identification, necessitating further training and testing sets. A learning rate of 0.005 is used,
research across various topographical contexts. along with the Adam optimizer, 50 epochs, a batch size of
Wang et al.,[2021][14] to strengthen detection 16, and a dropout rate of 0.5. This section guarantees that the
proficiency, recent explorations have engaged machine model is tested on unknown data in landslide detection,
giving a realistic assessment of its generalizability across Fig.1. Sequential flow of landslide
many settings and circumstances. An RNN is constructed
following the preparation of the data. Because RNNs record A. Performance Evaluation Matrix
temporal dependencies across time steps, they are especially Accuracy computes the number of accurately expected
useful for sequential data. This entails seeing trends in the data in general data.
way environmental factors, like soil moisture or rainfall
intensity, change over time in order to detect landslides.
Advanced layers like Long Short-Term Memory (LSTM)
may be added to the model to address issues like vanishing Recall (Sensitivity) computes the number of properly
gradients. This keeps the model's long-term dependencies, expected positive data in general actual positive data.
which are essential for comprehending how landslide risks
gradually evolve.
The system enters the training phase after the RNN Precision calculates the number of suitably expected
architecture is established. The preprocessed data is used to positive data in overall expected positive data.
train the model with the proper optimizer and loss function.
Weighted loss functions can be used to rank infrequent
landslide occurrences for landslide detection. Regularization
techniques like dropout are employed to avoid overfitting, F1 score reviews recall and precision into one number.
particularly when working with the relatively small datasets
typical of landslide investigations, and the learning rate is
modified for maximum performance. This models
performance using the test dataset after training is finished
B. Dataset Description:
and it evaluate the model through the metrics like accuracy,
recall, F1-score and precision. The dataset, sourced from Kaggle consists of 1000
entries, delineated by 7 distinct attributes together with a
Building machine learning models for sequential data binary target variable, landslide, which denotes the
tasks, such as landslide detection, is made comprehensive occurrence (1) or absence (0) of a landslide event. The
and methodical by this organized procedure. The system
attributes include temperature, indicative of the atmospheric
suggests a dependable and scalable explanation for analyzing
environmental data and predicting landslide risks by tackling thermal conditions, and humidity, functioning as an index
important preprocessing issues, making use of RNNs' for the moisture content present within the atmosphere.
advantages, and thoroughly testing the model. In the end, this Pressure signifies the atmospheric pressure, while rain
helps to mitigate disasters and improve public safety. measures the quantity of precipitation recorded. The dataset
additionally incorporates lightavgw/o0, which captures the
average light intensity while excluding zero values, and
lightmax, representing the highest recorded light intensity.
In addition, dampness measures the quantity of water
available in the ground. Collectively, these attributes enable
the prediction of landslides based on the prevailing
environmental conditions.

IV. RESULTS AND DISCUSSION

In this research, the dataset was split into a training
dataset and a testing dataset with 30% of the data reserved
as a test set. Fig. 2 presented in the paper delivers a precise
overview of the data structure and describes the dataset
further. The recurrent neural network was implemented on
Python libraries TensorFlow and Kera’s, and trained and
tested in models very conveniently.
 Fig 3 Confusion matrix
Fig. 2 Data set visualization The Random Forest model also performs well and it
To contextualize the RNN's performance, its accuracy is produces accuracy is 86%, precision is 81%, and recall of
compared with several other classifications in Table I 89%, making it a reliable alternative, though slightly behind
Accuracy comparison methods applied to similar datasets: the RNN in overall performance. The Decision Tree method
shows a similar trend and provides accuracy of 85%,
TABLE I ACCURACY COMPARISON WITH OTHER METHODS precision of 76% then recall of 90%, but it has the lowest F1-
F1- score among the top models at 83%. Finally, the support
Method Accuracy Precision Recall
Score vector machine performs the feeblest and produces an
RNN 89% 82% 94% 87% accuracy is 80%, precision is 75%, and recall is 81%,
Random resulting in the lowest F1-score of 78%. In summary, while
86% 81% 89% 85%
Forest the RNN provides the best overall performance, both Random
Decision Forest and Decision Tree models offer competitive results,
85% 76% 90% 83%
Tree especially in the recall, with SVM lagging in all key metrics.
Support
80% 75% 81% 78%
Vector Machine

The performance of the models reveals that the RNN is

the most effective for landslide detection, achieving the
highest accuracy is 89%. Precision is 82%, and F1-score is
87% shown in Table 1. Its standout feature is a high recall is
94%, indicating its strong ability to identify true landslide
events.

Fig. 4 Loss curves and Accuracy curves

 2020, October. Real-time monitoring and early warning system for
landslide preventing in myanmar. In 2020 IEEE 9th global
conference on consumer electronics (GCCE) (pp. 303-304). IEEE.
[4] Aggarwal, S., Mishra, P.K., Sumakar, K.V.S. and Chaturvedi, P.,
2018, April. Landslide monitoring system implementing IOT using
video camera. In 2018 3rd International conference for
convergence in technology (I2CT) (pp. 1-4). IEEE.
[5] Munasinghe, K. and Karunanayake, P., 2021, April. Recursive
feature elimination for machine learning-based landslide prediction
models. In 2021 International Conference on Artificial Intelligence
in Information and Communication (ICAIIC) (pp. 126-129). IEEE.
[6] Lu, M.C., Tang, T.Y., Tsai, C.P., Wang, W.Y. and Li, I.H., 2011,
June. Image-based landslide monitoring system. In Proceedings
2011 International Conference on System Science and
Engineering (pp. 638-643). IEEE.
[7] Doerksen, K., Gal, Y., Kalaitzis, F., Rossi, C., Petit, D., Li, S. and
Dadson, S., 2023, July. Precipitation-Triggered Landslide
Prediction in Nepal Using Machine Learning and Deep Learning.
In IGARSS 2023-2023 IEEE International Geoscience and Remote
Sensing Symposium (pp. 4962-4965). IEEE.
Fig. 5 Receiver operating characteristic curve [8] Ado, M., Amitab, K., Maji, A.K., Jasińska, E., Gono, R.,
Leonowicz, Z. and Jasiński, M., 2022. Landslide susceptibility
The results represented in the article include the evaluation mapping using machine learning: A literature survey. Remote
metrics of a machine-learning model. Fig. 3 shows the Sensing, 14(13), p.3029.
confusion matrix, indicating which predictions made by the [9] Raut, S., Dutta, D., Bera, D. and Samanta, R., 2024, Assessment of
model were correct and which were not. Fig. 4 shows how the Landslide Susceptibility using Geospatial Techniques: A
Comparative Evaluation of Machine Learning and Statistical
training/validation loss and accuracy curves validate the Models. Geological Journal.
prototype's knowledge capability over epochs. Lastly, the [10] Nair, A., 2024. Geospatial Analysis of Landslide Susceptibility
Receiver Operating Characteristic (ROC) curve is shown in Through Machine Learning In Relation to Environmental
Fig. 5 and its AUC score indicates good classification Indicators Across Global Regions.
performance. All of these Figures together demonstrate the [11] Miao, F., Ruan, Q., Wu, Y., Qian, Z., Kong, Z. and Qin, Z., 2023.
Landslide Dynamic Susceptibility Mapping Base on Machine
operational efficiency of the model. Learning and the PS-InSAR Coupling Model. Remote Sensing,
15(22), p.5427.
[12] Casagli, N., Intrieri, E., Tofani, V., Gigli, G. and Raspini, F., 2023.
Landslide detection, monitoring and prediction with remote-
V. CONCLUSION sensing techniques. Nature Reviews Earth & Environment, 4(1),
In conclusion, the utilization of deep learning and pp.51-64.
[13] Meena, S.R., Soares, L.P., Grohmann, C.H., Van Westen, C.,
machine learning particularly RNNs, significantly enhances Bhuyan, K., Singh, R.P., Floris, M. and Catani, F., 2022. Landslide
landslide prediction accuracy. RNNs excel at processing detection in the Himalayas using machine learning algorithms and
sequential data, making them ideal for analyzing time-series U-Net. Landslides, 19(5), pp.1209-1229.
environmental factors like flood, soil moisture, and [14] Wang, H., Zhang, L., Yin, K., Luo, H. and Li, J., 2021. Landslide
Earthquake occurrences. This high accuracy enables early identification using machine learning. Geoscience Frontiers, 12(1),
pp.351-364.
warnings, allowing local authorities and residents to take [15] Chaganti, P.C.V., Vasireddy, K., Reddy, E.R., Dondeti, R.R. and
timely preventive actions. As the prototype is trained on Syama, S., 2023. Predicting Landslides and Floods with Deep
additional real-time data, its predictive capabilities improve, Learning. In 2023 4th International Conference on Electronics and
leading to more reliable forecasting. The system’s scalability Sustainable Communication Systems (ICESC) (pp. 1259-1265).
ensures its application in various regions and scenarios. IEEE.
Further future work integrating IoT sensors and geospatial
technologies to enhances prediction accuracy and spatial
relevance. Continuous data collection allows for the
refinement of the model, contributing to ongoing risk
reduction. By leveraging advanced technologies, landslide
warning systems become more effective at minimizing loss
and damage. With climate change and urban expansion,
these innovations are crucial for disaster risk reduction. The
combination of RNNs and deep learning promises a safer,
more resilient future for landslide-prone communities.

REFERENCES

[1] Sharma, M. and Laskar, S., 2017, December. Landslide Mapping,

monitoring and modeling techniques: a new approach using
DOFS. In 2017 International Conference on Circuits, Controls,
and Communications (CCUBE) (pp. 21-24). IEEE.
[2] Parks, S.M., 2012. Synthetic Aperture Radar (SAR) and Optical
Imagery Data Fusion: Crop Yield Analysis in Southeast Asia. The
International Archives of the Photogrammetry, Remote Sensing
and Spatial Information Sciences, 39, pp.403-406.
[3] Thein, T.L.L., Sein, M.M., Murata, K.T. and Tungpimolrut, K.,