Reviewing large satellite imagery datasets can be time-consuming, but pairing geospatial processing with artificial intelligence (AI) improves the speed and consistency of image analysis. Satellite imagery provides a broad view of landscapes, infrastructure, and environmental conditions, while AI image recognition tools can automatically screen image datasets for features of interest.
This article focuses on how satellite imagery and similar datasets (e.g., drone data) can be prepared for AI-based analysis, drawing on our recent experience identifying surface features related to abandoned mine lands (AMLs) in Wyoming.
Why Abandoned Mine Lands Can Be Difficult to Identify in Large Imagery Datasets
Satellite imagery is often organized into datasets that cover large areas and include varied landforms, image resolutions, and data formats. Reviewing these datasets manually is difficult when the objective is to identify features the size of a house or smaller.
Several common challenges can affect image-analysis workflows, including these examples:
- Inconsistent image dimensions
- Variable spatial resolution
- Differences in source formats
- Difficulty connecting image-based results to physical coordinates
These challenges are compounded because important features may be small, scattered, difficult to recognize, or located in unexpected areas. Furthermore, the images may be inconsistent due to changes in lighting, seasonal or interannual vegetation differences, and variations in feature colors and shape. Without a programmatic workflow, processing and interpreting the imagery becomes less efficient, more difficult to reproduce or validate, and increasingly resource intensive.
Earlier automated techniques have failed to address these issues because they rely on fixed rules, thresholds, or simple image statistics that only work under narrow conditions. Consequently, manual review becomes practical only over smaller spatial extents, which can bias feature identification toward locations where features are already expected and limit the feasibility of comprehensive regional evaluations, such as an assessment of the entire state of Wyoming.
AI image recognition has improved the ability of automated techniques to detect features under a larger range of conditions, but the imagery data must still be prepared, standardized, and georeferenced to real-world coordinates in a programmatic and repeatable way.
Preparing Satellite Imagery for AI-Based Abandoned Mine Lands Analysis
Before satellite imagery can be used in AI image identification models, it needs to be converted from geospatial data formats into standardized image files. Using an AI image recognition model configured and trained by Trihydro to identify abandoned mine lands, imagery from sources such as WMS tiles and GeoTIFF files was processed into image sets with consistent pixel dimensions and consistent geographic coverage. Each processed image contained the same number of pixels along the x- and y-axes and used a consistent relationship between pixel dimensions and ground distance.
This standardization reduces variations unrelated to the target feature, allowing the AI image model to focus on visual patterns in abandoned mine lands rather than erroneous differences caused by scale and size.
Building an AI Training Dataset for Abandoned Mine Lands Detection
AI image recognition tools require training datasets to teach the model what features to recognize. For geospatial applications, training datasets typically include examples that contain target features as well as examples where those features are absent. Including both types of imagery helps models distinguish meaningful patterns from surrounding landscape conditions.
Training imagery should represent the range of conditions likely to be encountered during deployment, including differences in terrain, vegetation, lighting, seasonal conditions, image resolution, and feature appearance. This type of diverse training dataset helps improve a model’s ability to recognize features when conditions differ from those present in the original training images.
Improving Model Performance with Windowing Techniques
Windowing techniques can then be used to make the training data more realistic. Windowing methods create a smaller, usable selection of spatial data. Rather than placing an adit or mining test pit in the center of every image, the image processing workflow is designed so these features appear in different parts of the image frame. This variation reduces the risk that the model learns the image position rather than the feature. Additionally, the model can also be trained to identify abandoned mine land features split between two images. Using the range in geographic size of known abandoned mining features, the windowing techniques incorporate a small systematic overlap between windowed images.
Applying AI Image Recognition to Abandoned Mine Lands Mapping
After model training, AI image recognition can identify features within the satellite imagery. The AI image recognition model, a region-based convolutional neural network (Faster R-CNN), detects objects or features within an image rather than classifying the image as a whole, and is especially useful for smaller features that may occupy only a fraction of the image. The model also retains the coordinate data, which helps guide the mapping, review, and validation of flagged features in GIS software and through fieldwork. This connection bridges the gap between computer vision outputs and field-verifiable locations.
Applying Satellite Imagery and Geospatial Processing Beyond Abandoned Mine Lands
Combining geospatial processing with AI image recognition can help project teams evaluate large imagery datasets more efficiently. Although this workflow was applied to satellite imagery, it could also be adapted for high-resolution drone imagery, allowing teams to identify much smaller land features. Examples could include wildlife habitat features such as bird nests, reclamation vegetation gaps, small-scale erosion features, surface water and drainage patterns, and other site-specific conditions that may require follow-up review or field verification. Importantly, the model can be used on incredibly large scales: Trihydro’s AI image recognition model was designed to survey over 97,813 square miles – i.e., the entire state of Wyoming.
AI should not replace professional judgment and field verification but provides new ways to evaluate datasets previously considered too large, time-consuming, or costly for human-only analysis. This customized, client-focused workflow allows the user to narrow search areas, identify potential patterns, and improve the repeatability of image analysis, offering an emerging technology for environmental and geospatial challenges.
Common Questions About Using AI and Satellite Imagery for Abandoned Mine Lands Mapping
Applying AI to satellite imagery presents new opportunities for identifying potential Abandoned Mine Lands (AML) features across large geographic areas. The questions below address common considerations regarding image preparation, model training, feature detection, and AI’s role in supporting AML mapping and assessment efforts.
No. AI image recognition can help identify potential Abandoned Mine Lands (AML) features across large geographic areas, but field verification remains essential to the assessment process. Satellite imagery and image recognition models can screen large datasets and flag areas for further review, helping focus field efforts where they are most needed. However, professional judgment and site visits are still required to confirm feature types, evaluate conditions, and support decision-making.
AML features are often small, dispersed, and visually varied. Factors such as vegetation cover, lighting conditions, seasonal changes, image resolution, and topography can make features difficult to distinguish from surrounding landscapes. In addition, large imagery datasets may contain inconsistent formats, scales, and resolutions. Effective AI applications depend on careful image preparation, standardization, and training datasets that represent the range of conditions likely to be encountered in the field.
Traditional manual image review becomes increasingly challenging as project areas grow. By combining geospatial processing with AI image recognition, organizations can evaluate imagery across large regions while maintaining the geographic coordinates needed for GIS analysis and field verification. This approach can help identify potential features of interest, support more consistent image review, and make statewide or regional AML inventory efforts more practical.
