Characterizing manufactured home parks in North Carolina: A computer vision based approach

Identification of manufactured home parks

No publicly available, comprehensive, granular dataset of MHP at the state or national level exists. The largest source is the MHP database created by the Department of Homeland Security (DHS) (Department of Homeland Security, 2022). These data, while providing precise geographic coordinates of over 53,432 parks in the US, have incomplete data on MHP characteristics. Furthermore, visual analysis of aerial images suggests this database is outdated. Similarly, cadastral datasets can indicate the presence of manufactured structures, or a zoning category or a land use designation that allows MHP. These cadastral datasets are neither comprehensive nor current and are often inconsistent across counties. Private real estate databases contain information about manufactured housing, but are costly.

Among a variety of methods to identify MHP, Ward & Peters (2007) combined geospatial analysis and visual interpretation of aerial/multispectral imagery to identify informal settlements, including MHP in selected US regions. Durst (2016, 2019) used a similar approach to identify Colonias in Texas. Shakib et al. (2024) use crowdsourcing using a mobile application named MapSwipe to activate a volunteer community of OpenStreetMap mappers to identify MHP in Arizona. While these techniques helped fill critical data gaps, the time- and labor-intensive nature of the studies limits their large-scale applicability.

Supervised machine learning can enable large-scale identification of manufactured housing with less manual labor and resources. Durst et al. (2021) used landscape metrics derived from building footprints to train Support Vector Machine models to identify and classify mobile homes in Hidalgo County, Texas. The approach was effective, with 91% accuracy without adding parcel information and 99% accuracy after adding parcel information. Despite widespread availability of building footprints, this approach has generalizability concerns if applied to different landscape and land-use contexts, necessitating separate models and training datasets. Lunga et al. (2018) used a multi-level semantic extraction approach to identify MHU, in which patch-level classification was followed by pixel-level classification. The approach shows promising results (91% classification accuracy on a validation dataset), but detailed methods, scope, and outcomes are not publicized.

The studies demonstrate the potential of large-scale applied machine learning to identify MHCs, while also highlighting potential challenges with the availability of data as well as generalization across geographies (Maxwell et al., 2021). This suggests that a regional-level analysis combining geospatial datasets (such as aerial imagery, building footprints, and parcel records) can be promising in identifying MHP using machine learning. We build on these techniques and use large-scale datasets to identify MHP at a state-wide scale in North Carolina (NC).

Aerial imagery

Between 2020 and 2023, we acquired high-resolution aerial imagery (15 cm) through the North Carolina Orthoimagery Program (NCGIA, 2024). The program captures aerial imagery for one-quarter of the state annually using specialized survey aircraft (see Figure S2 for a county-wise map of imagery collection years), which enforces strict imaging standards for smoke and cloud coverage percentages per square mile and minimum sun angles. All imagery is captured during leaf-off conditions in winter to minimize vegetation interference. 56,020 GeoTIFF tiles (30 TB) of the imagery are stored on Amazon S3. Each 10,000 × 10000 pixel tile has four bands: Red, Green, Blue, and Near-Infrared.

MHP records

We used Mobile Home Park records made publicly available as part of Homeland Infrastructure Foundation Level Datasets (HIFLD) (Homeland Security 2022) to create training data and evaluate our results. The dataset includes 53,412 Mobile Home Parks, residential trailer parks, and recreational vehicle parks in the continental US and Alaska, with information on name, address, geo-coordinates, source, and size. We filtered this dataset to NC. NC has 2971 MHP: 2517 small (<50 units), 369 medium (51–100), and 85 large (>100). However, information about the exact number of units within each park is unavailable.

Parcel records

The parcel records span all 100 counties in the state, containing 5.6 million records (NCOneMap, 2016). The dataset was first created in 2016 and is continually updated. Dataset attributes, including georeferenced parcel polygons, standardized address information, parcel use codes and descriptions, and the number of structures, are defined by the Integrated Cadastral Data Exchange project.

Building footprints

We use the Microsoft US Building Footprints dataset (Microsoft 2018) for training data. This dataset was created using semantic segmentation and polygonization. The dataset contains 129 million footprints in the US, and 4.7 million footprints in NC. We also used Overture Foundation (2023) building footprints data in post-processing.

Home sales records

We used records of homes listed for sale to identify single-wide and double-wide MHU for training data creation. CoreLogic’s Multiple Listings Enterprise Solutions (MLS) dataset contains 157 million US listings spanning decades, constituting 80% of the total (CoreLogic, n.d). There are 3.5 million records in NC for the period between 1994 and 2022. The data includes details on address, property type and type of manufactured home (single wide vs double wide). However, significant data cleanup was necessary to identify manufactured homes listings. We identified 20,452 listings in North Carolina that are likely to be MHU, of which 6.25% are likely repeated or duplicated listings based on the property identifier used in the dataset.

Object detection

We frame the identification of MHU from aerial imagery as an object detection problem. Object detection is an approach in computer vision in which a bounding box and category are assigned to objects of interest in an image, typically using Convolutional Neural Networks (CNN). A CNN is a neural network that uses a kernel filter that convolves over an image to identify locally relevant features (O’shea and Ryan, 2015). Object detection was selected over segmentation or classification approaches for several key reasons. First, object detection provides precise spatial localization through bounding boxes, enabling accurate geospatial mapping essential for our analysis. Unlike semantic segmentation, which classifies pixels but struggles with overlapping objects, object detection can distinguish individual MHU even when they’re closely spaced within parks. Unlike patch-level classification approaches used in previous studies, object detection offers superior scalability across diverse geographic contexts. Individual MHU exhibit more consistent visual characteristics (size, shape, reflectance) than entire parks, making them ideal training targets.

We used a YOLOv8 architecture model pre-trained on the Common Objects in Context (COCO) dataset by retraining it using a custom training dataset (Wang et al., 2023). YOLOv8 is an openly licensed state-of-the-art model that has been used with aerial imagery (Wang et al., 2023). It is a single-stage object detector that detects an object of interest in an image with a single pass, in contrast to classical two-stage object detectors in which regions of interest are first identified and later classified, allowing faster inference, at the expense of accuracy for smaller objects (Redmon et al., 2016). The YOLOv8 single-stage architecture provides computational efficiency crucial for processing tens of thousands of high-resolution imagery tiles. While two-stage detectors might offer marginally better accuracy, YOLO’s speed enables state-wide analysis within reasonable timeframes. Additionally, object detection’s confidence scoring facilitates post-processing workflows, allowing us to filter detections and aggregate individual units into parks systematically. We split images into overlapping tiles to better detect small objects using an open-source tool named Slicing Aided Hyper Inference (SAHI), improving recall performance on small-object detection (Akyon et al., 2022).

The YOLOv8 architecture mainly consists of two parts: the backbone and the head (Jocher et al., 2023). (See Figure S6 for the architecture diagram.) The CSPDarknet backbone includes the SPPF (Spatial Pyramid Pooling Faster) module that extracts features at various scales through max pooling (Sohan et al., 2024). The head processes the backbone features into object presence, location, and classification.

Training data creation

Manual annotation for computer vision is labor-intensive (Man and Chahl, 2022). We created training data programmatically by geospatially linking aerial imagery, building footprints, and parcel datasets, followed by manual validation. We matched geocoordinates of buildings known to be MHU from MLS and HIFLD data with Microsoft building footprints using a Python script. An image slice around the center of the georeferenced bounding box is saved and a text file with category and bounding box in the YOLOv5 data annotation format is saved. The image file and text file saved with the same filename together make one training image-annotation pair.

Since single-wide mobile/manufactured housing is under-represented in the CoreLogic dataset, we created additional training data for the category by using the HIFLD dataset. The HIFLD data only has point locations of the parks. We identified GeoTIFF corresponding to MHP geocoordinates in HIFLD and filtered building footprints that are likely to be MHU based on size (length between 40 and 80 ft and width between 12 and 18 feet). The programmatically created training dataset had issues such as miscategorization, missing annotations, misalignment of bounding boxes, and missing objects. Substantial human effort was dedicated to ensuring data quality by visually verifying the objects within the annotation boundaries (see Figure S4). The Roboflow web portal was used to edit incorrect annotations.

We included Null images in the training dataset. These images have objects that look like but are not mobile/manufactured housing from an aerial view. Examples of such images include, roads, warehouses, subdivision homes, trees, and fields. These images accompanied by no annotated bounding boxes, are useful to train the model to distinguish those from MHU. From the parcel attributes, we identified parcels from different regions in the state that have the potential to contain these different objects. For each of the parcels, we sliced the smallest square enclosing the parcel from the corresponding aerial imagery.

The final training dataset consists of 4674 images, including 3134 instances of single wide MHU, 1613 instances of double wide MHU, and 126 null instances. The dataset was split into Train, Validation, and Test sets, roughly at 70:20:10 proportions. A detailed breakdown of the training dataset is shown in Table S1.

Training used an NVIDIA L4 24GB GPU for 223 epochs (0.99 h). The best results on the validation set were obtained at Epoch 173. We evaluated the performance of the trained model on the held-out test set, representing the in-distribution performance of the model.

Inference and post-processing

To improve recall on the 10,000 × 10,000 pixel imagery, we used a tiled inference approach with Slicing Aided Hyper Inference (SAHI) on top of the YOLOv8 model. Imagery was split into tiles of preset size (360 pixels × 360 pixels) and percentage overlap (30%) before inference. Non-maximum suppression removed redundant detections during aggregation, favoring high confidence overlaps. A default of 0.7 is used as Intersection over Union (IoU) threshold for non-maximum suppression, and a confidence threshold of 0.3 for object presence. Each image was processed in a single-CPU environment, representing 56,020 jobs in the computational cluster. We geolocated the bounding boxes detected by the YOLOv8 model by using the coordinate system of the underlying inference image.

We followed the following post-processing steps to remove potential false positives. First, we removed any inference smaller or larger than realistic sizes for MHU. Based on the minimum and maximum dimensions of single and double-wide homes, we calculated the diagonal length of permissible sizes of minimum-rotated bounding boxes for each category (i.e., 37 to 96 ft for single-wide and 41 to 82 ft for double-wide). This step removes objects such as large warehouses that might look stylistically like MHU because of comparable reflectivity but are substantially larger. We removed inferences outside parcel geometries, eliminating the objects detected in and around roads. Roads have spectral characteristics similar to MHU, leading to incorrect detections.

Finally, we visually inspected the parcels with multiple detections and validated that they contained at least 3 MHU. This threshold was chosen based on the observation that many parcels with three or fewer detections were primarily used for agricultural and industrial purposes, and not as a manufactured housing park/subdivision in which most of the parcel is used by housing units. We selected parcels with at least one detection and at least one Foundation (2023) building footprint for validation to select parcels with a high chance of being an MHP without having to manually validate all parcels with detections. Manual validation entails slicing out aerial imagery corresponding to the parcel geometry as an image and deciding whether the image represents an MHP. The validation was performed by two of the authors. To address inter-rater reliability, four counties were first validated by both authors, with an Intersection over Union (IoU) score of 0.72.

Object detection performance

We obtained an in-distribution performance of over 93% recall (the number of detections out of the total count in ground truth) on both single-wide and double-wide MHU on the test dataset. Similarly, we obtained over 93% bounding-box precision (overlap between the bounding boxes of detection and ground truth) for both classes. Minimum Average Precision at 50% (correct categorization with at least 50% bounding box overlap with ground truth) was 96% for single wide and 94% for double wide MHU (see Table S2). Comparable performance in both classes across the metrics suggests a balanced training set composition.

We observed differences between in-distribution and out-of-distribution performance. While we observed comparable in-distribution precision and recall across the two classes of MHU, out-of-distribution performance suggests the model performed better with single wide MHU than double wide MHU. False positives, such as larger warehouses or other types of buildings, are prevalent in the out-of-distribution inferences (see Figure 1). The model sometimes falsely detected apartment buildings, industrial containers, tennis/pickleball courts, forests, parking lots, roads/railway tracks, utility-scale solar, and sheds, with MHU (see Figure 1). These objects share visual characteristics such as color, shape, and other features, resulting in false-positive detections.OPEN IN VIEWER

Following the post-processing steps, we found that nearly 79% of detections were outside the range of expected size (refer to Table S3). Filtering by expected size removed smaller objects such as small solar arrays, small sheds, boats, and cars and larger objects such as large sheds, elongated water bodies, and different housing types such as apartments and formal subdivision homes (see Figure S5). Matching with parcels removed another 6.77%, removing roads and vehicles falsely detected as MHU. After manually validating the imagery respective to matched parcels with multiple detections, about 15% of the remaining MHU belong to MHP.

Geographic distribution of MHP

We identified 8460 MHP in North Carolina, including 6653 isolated parcels, 592 paired clusters, and 172 groups of 3+ contiguous parcels. Since it is not possible to determine if the contiguous parcels are functionally different MHP without extensive validation, we treat each parcel as a separate MHP. Counties average 84.6 MHPs (median: 66). Southeastern, central, and southwestern counties in the State have higher numbers of MHP (see Figure 2(a)). The five counties with the highest number of MHP are Buncombe, Wake, Robeson, Onslow, and Davidson. Five counties contain fewer than 10 MHP.OPEN IN VIEWER

Large MHP are concentrated in the south-central (Lincoln and Mecklenburg) and central (Wake and Wayne) and eastern (Onslow) counties. Two counties with these concentrations have large military installations nearby and significant seasonal work in the agricultural and farming industries. Mecklenburg (home to Charlotte) and Wake (home to Raleigh) are the two biggest counties in North Carolina with high housing costs. Urban areas contain 33.7% of MHPs (predominantly large and medium-sized parks). We use the urban area 2020 definition from the US Census to distinguish urban and rural MHP. A large proportion of MHP are at the fringes of urban areas, both outside and inside the borders of the clusters (see Figure 2(b)). MHP are abundant in the fringes of large urban areas in Charlotte and Greensboro, two large cities in North Carolina. In smaller urban areas (such as Fayetteville and Asheville), MHP are also prevalent in central as well as fringes of the urban areas.

The prevalence of small MHP in rural areas is striking. Over 67% of the MHP with less than 50 units are found in rural areas and they are prevalent across the state. These small MHP make up 95% of the MHP in the state. Counties with the highest number of small MHP are Buncombe, Wake, and Robeson. In Buncombe County, 74% of small MHP are urban, whereas in Robeson County, 86% are rural.