BLOG/UPDATES

From Multi-Class to Binary Collapse: Garbage Detection in Satellite Imagery around Zoopigi

A new model has been trained to classify satellite image patches as either garbage or not_garbage. This differs from previous models by combining a ResNet50 CNN with a multi-class classification approach followed by a final binary collapse.

The model is trained on a labelled multi-class dataset containing a range of common landscape features, such as grass, trees, roads/tracks, and garbage. It then classifies previously unseen satellite image patches according to their best match among these feature classes. Finally, the multi-class predictions are collapsed into a binary result: patches classified as garbage retain the garbage label, while predictions belonging to all other classes are combined into the not_garbage class.

Testing was performed on a realistically weighted dataset, with garbage/not_garbage in an approximately 1 to 3 ratio. This has resulted in a significant improvement in performance over previous models.

A substantial improvement is observed in overall accuracy. The increase in precision is particularly encouraging, even though it comes at some cost to recall. Poor precision has been one of the main limitations of previous models, with false positives sometimes approaching the number of true positives. In a real-world application, where detected sites might need to be physically inspected and potentially cleaned up, this would result in too many unnecessary site visits for the system to be genuinely useful. The new approach reduces the number of false positives to a much more manageable level. Although the trade-off is that some genuine garbage sites will inevitably be missed, the resulting predictions are considerably more practical for real-world use.

The generalisation of the model has been assessed by collecting and applying it to a set of satellite images collected from an area in Cyprus in the Pitsilia region. This consisted of 18 square km that make up the Zoopigi community boundary. A total of 23,328 image patches were analysed by the model which detected garbage in 987 of these, representing 4.23% of those examined. The performance of the model on this region was assessed by performing three lots of random testing. The results were actually better than for the evaluation data, as illustrated in the following tables:

The generated google maps can be found in the link Zoopigi garbage.

I know that datasets of the sort I have developed during study of illegal garbage dumps, labelled and unlabelled, are difficult to find. If these can be of use to anyone elase than I am happy to share. Just look up by contact details and send me a message.

Urban Expansion in Spitali: A Satellite-Based Analysis (2015–2025)

Cyprus is undergoing a rapid expansion in population, primarily due to geopolitical issues in Eastern Europe and the Middle East. Much of this growth has been centred around the town of Limassol, which has grown by between 8% and 10% since 2020.

Associated with this has been a building boom. The village of Spitali, which lies at the northern tip of the Limassol conurbation, has seen extensive expansion of its housing network. To quantify this growth, I used historical satellite imagery together with a CNN-based neural network model (previously trained on buildings elsewhere in Cyprus) to estimate changes in the landscape for the Spitali village environs, specifically detecting the extent of buildings.

The area sampled was approximately 1.5 km by 1 km. Each image patch analysed represented an area of around 150 × 150 square metres. Random sampling indicated that the model achieved an overall accuracy of 93%, with an F-score of 0.89 when applied to the Spitali datasets. This represents strong performance given the training data and application domain.

Two datasets were used: one generated from January 2015 and another ten years later, from June 2025 (the most recent available). The results showed that in 2015 there were 51 image patches that included buildings. By 2025, this had increased to 66, representing a 29% increase. Building has continued apace since the latest high-resolution satellite images became available. I intend to collect updated images as they become available and rerun the models to monitor the rate of change.

Graph Neural Networks

Prototyped some graph neural networks, with a ResNet50 backbone, using pytorch. Graph neural networks model data as nodes and edges to learn from relationships and structure rather than isolated features. In image classification, they represent pixels, regions, or objects as graphs, enabling context-aware recognition that captures spatial and semantic relationships beyond standard CNNs. Promise shown; on the garbage detection domain it exhibited similar performance to that of basic ResNet50.