Informal settlements have become a pressing issue across Latin America, reflecting deep-rooted socioeconomic inequalities and rapid urbanization. In 2014, approximately one-quarter of the urban population in Latin American cities lived in informal housing. By 2017, this number had increased to nearly one-third, and the trend continues to rise.
Each country’s informal settlements have unique names that reflect local history, geography, and economic conditions. Whether referred to as favelas in Brazil, villas miseria in Argentina, or barrios populares in Colombia, these communities share common characteristics: lack of proper infrastructure, limited access to basic services, and informal land occupation. Their expansion underscores the urgent need for sustainable urban policies, affordable housing solutions, and inclusive development strategies to address the challenges of growing urban inequality. This project will specifically focus on Chile, examining the development of informal settlements, their impact on urban life, and the policies surrounding them.
In Chile, informal settlements—commonly known as tomas and campamentos—are concentrated on the outskirts of urban areas. As people migrate to cities in search of job opportunities, they often face prohibitively high housing costs, forcing them to occupy land illegally.
Currently, there are over 1,000 informal camps across Chile, housing approximately 72,000 families, with an average of three people per home. These settlements begin as tomas, where land is seized without legal authorization. Over time, some of these evolve into campamentos, which, while still informal, are officially registered by the Ministry of Housing and Urbanism (MOH). Unlike tomas, which can face eviction with severe legal consequences—including prison sentences since 2023—campamentos receive certain legal protections in eviction processes.
Residents often construct their own homes using recycled, found, or low-cost materials. The most common housing type is the mediagua, a small prefabricated wooden structure, as shown in the image. However, these homes often lack basic infrastructure, such as proper water, sewage, and electricity connections, leading some residents to resort to illegal hookups.
In addition to poor living conditions, these settlements are frequently located in high-risk areas prone to natural disasters, such as floods, landslides, and earthquakes. The lack of proper urban planning also results in narrow, unpaved roads and unsafe terrain, making it difficult for emergency vehicles to reach these communities in times of crisis, further exacerbating their vulnerability.
Some settlements remain unregistered due to administrative barriers or residents’ refusal, often linked to crime, distrust, or indigenous autonomy. Without recognition, these communities lack access to basic services, legal protections, and government support, deepening their marginalization.
Natural hazards are a key factor in decision-making. Many settlements are in high-risk zones—floodplains, steep slopes, or earthquake-prone areas—where precarious housing increases vulnerability. If residents remain, their exposure to disasters worsens over time, making intervention essential. Evaluating risk levels helps determine whether to formalize, relocate, or evict and rehome affected communities. Without inclusion, both physical and social vulnerabilities grow.
In Chile, a long and thin country stretching over 4,300 kilometers from north to south, the north is characterized by arid desert landscapes, shifting to a temperate Mediterranean climate in the central region, and finally to the wet and chilly regions of the south. The eastern boundary of Chile is defined by the Andes mountain range, while the west features lower coastal hills with steep slopes and ravines. The varying topography and geological complexity contribute to natural risks such as mass movements/landslides, floods, tsunamis, volcanic hazards, and others unique to each region. These natural risks shape urban planning and influence the diverse formations of campamentos, as well as the challenges faced by residents in different locations.
The potential for exclusion becomes a major issue, as marginalized communities often lack the resources to relocate or mitigate these risks, leading them to settle in unsafe environments. Addressing these risks requires not only better infrastructure but also ensuring that communities have the support and means to move away from high-risk zones. If people remain in these areas, their vulnerability is exacerbated, as they are repeatedly exposed to natural disasters.
The morphology of campamentos varies across Chile’s natural regions due to differing environmental conditions, as seen in aerial photographs. These diverse natural factors, such as topography, climate, and proximity to natural hazards, shape the formation of campamentos, influencing both their layout and the materials used for construction.
The rise of informal settlements in Chile has been driven by multiple crises over the past decade. In 2010, a devastating 8.8-magnitude earthquake affected 80% of the population, leaving 9% homeless and accelerating the expansion of informal camps. More recently, the COVID-19 pandemic significantly worsened the housing crisis, as widespread job losses and wage reductions made rent unaffordable for many. Another key factor is immigration, with approximately 40% of residents in these settlements being migrants, particularly from Venezuela and Bolivia, entering Chile through its northern border. These trends have led us to focus on Tarapacá, a region experiencing significant growth in informal camps.
Within Tarapacá, we have chosen to examine Alto Hospicio, an area with some of the largest informal settlements in the region. With thousands of homes spread across 111 hectares, Alto Hospicio represents a critical case for understanding the challenges of informal housing, infrastructure gaps, and the broader socioeconomic factors driving settlement expansion in northern Chile.
METHODOLOGY
To detect informal settlements, we began by locating them using existing data on their location and size and ensured that they were visible on satellite imagery. We collected images from these areas, annotated them, and used them to train different models: for object detection and segmentation. Object detection prioritizes the number of objects, allowing instances to be counted, while segmentation identifies whether an area contains the object and where, rather than identifying particular instances. We used the resulting models to detect camps using the Deepness plug-in on QGIS.
During the segmentation process, we successfully trained the model and generated an ONNX file. However, we encountered difficulties when inputting the file into QGIS using the Deepness plugin. The model’s output exceeds the tile size used for processing, which causes the script to fail in calculating the necessary offset to handle the image in smaller chunks. These issues will be the next focus of our work.
To train the model, we selected informal settlements that were already identified by the Chilean Ministry of Housing in their 2023 registry. We downloaded image tiles of these informal settlements and their surroundings, and annotated these manually in Roboflow, identifying the particular shapes of each block of informal settlement or marking null when there were no informal settlements in the image.
TRAINING THE MODEL
We trained various models and gradually improved the detection of informal settlement blocks. First, we trained the model on 267 annotated images, then used a model with 440 images, and finally 674 images.
We adjusting object identification parameters on Deepness in QGIS to see which settings most accurately identified informal settlements. We optimized settings based on two different parameters: the percent of detected objects within the official camp territories and the percent of the official camp territories covered by detected objects.
TESTING OBJECT DETECTION IN DIFFERENT AREAS
ALTO HOSPICIO
In Alto Hospicio, we conducted 16 different tests and settled on the following best parameters:
Resolution: 60
Tile Overlap: 15%
Confidence: 0.5
IoU: 0.3
Which gave us 76% of detected objects within registered camps and 71% coverage of detected camp areas.
CALAMA
We then checked our model in Calama, another city in the desert. We conducted 5 different tests and settled on the following best parameters:
Resolution: 60
Tile Overlap: 15%
Confidence: 0.2
IoU: 0.3
Which gave us 39% of detected objects within registered camps and 58% coverage of detected camp areas. While the model was more prone to false positives, it was still adept at detecting blocks within the biggest informal settlement in Calama.
ARICA
Finally, we conducted 5 tests in Arica, another northern desert city. Though the parameters that seemed to most accurately identify formal settlements were:
Resolution: 60
Tile Overlap: 15%
Confidence: 0.2
IoU: 0.3
Which gave us 3% of detected objects within registered camps and 19.6% coverage of detected camp areas.
However, the “false positives” this model was outputting raised important questions about the validity of the official register, as the “false positives” fit the informal settlement block typology well. These detected objects seem to be informal settlements that exist outside the official register.
CONCLUSIONS
The methodology demonstrates how machine learning and satellite imagery can help identify informal settlements, including those not officially recognized. Traditional methods often miss settlements in less visible areas, but this approach might provide a more comprehensive way to detect and map these communities, offering valuable insights for urban planning.
As a next step we will overlay the identified informal settlements with data on different types of risks. By combining this information with the settlement locations, we can assess the level of vulnerability these communities face. This risk detection can help prioritize interventions, ensuring that resources are directed toward the most vulnerable areas, and contributing to more informed, risk-aware urban planning.
RISKS DETECTION
