The escalating global waste crisis necessitates efficient and cost-effective methods for monitoring and managing dumpsites. Traditional methods, relying heavily on manual labor, are time-consuming, expensive, and often insufficient to cover large areas. The World Bank highlights the significant economic burden of collecting illegal dumpsites, emphasizing the need for improved monitoring strategies. This research addresses this challenge by proposing a novel deep-learning approach to automatically detect dumpsites from high-resolution satellite imagery. The rapid increase in global waste generation, coupled with the global push for carbon neutrality, underscores the urgency of addressing this issue. Greenhouse gas emissions from waste contribute significantly to climate change, further motivating the need for effective waste management practices. The uncontrolled growth of dumpsites, both legal and illegal, poses serious environmental and public health risks, including the spread of infectious diseases and harm to wildlife. Illegal dumping, often conducted by nearby residents, makes tracking these sites challenging. Timely and accurate information on dumpsite locations is crucial for effective policymaking and waste management strategies. This research aims to provide a robust and scalable solution to accurately identify and classify dumpsites on a global scale, thereby informing research on human behavior, geoscience, and environmental protection.

Existing research on dumpsite detection often employs manual or semi-manual methods using satellite imagery, which is still labor intensive and expensive. While some studies use Unmanned Aerial Vehicles (UAVs), this approach is not scalable for global monitoring. Previous works have integrated deep learning with earth observation for dumpsite detection, but these efforts are generally limited to specific regions or countries, and lack publicly available datasets. BigEarthNet contains a small number of unclassified dumpsites but is primarily used for general land use classification. The lack of publicly available, classified global dumpsite datasets has hampered broader research efforts. This study addresses this gap by creating a comprehensive, globally representative dataset and developing a novel deep learning model specifically designed for dumpsite detection.

This research employed a four-fold approach: 1) Dataset Creation: A global dumpsite dataset was constructed by manually labeling approximately 2500 dumpsites across nearly 4800 square kilometers of high-resolution satellite imagery (0.3m to 1m per pixel). The selected areas represented cities with diverse population sizes and environmental performance rankings, ensuring global representation. Dumpsites were categorized into four classes: domestic waste, construction waste, agricultural waste, and covered waste. 2) Model Development: A novel deep convolutional neural network, BCA-Net, was designed to address the challenges of irregular dumpsite shapes and appearance, and the class imbalance in the dataset. The model leverages a two-stage object detection network (Faster-RCNN) with a Feature Pyramid Network (FPN) for multi-scale feature extraction and a Blocked Channel Attention (BCA) module to enhance feature learning. Data augmentation and category balancing techniques were implemented to mitigate the long-tailed distribution of the dataset. 3) Model Validation: The dataset was split into training, validation, and test sets (60:20:20). Model performance was evaluated using sensitivity and precision metrics, comparing results to previous models and human performance. 4) Spatial and Temporal Analysis: The trained model was applied to satellite images of the central areas of 28 cities to detect dumpsites. A Global Dumpsite Index (GDI) was developed to quantify the level of dumpsites in each area. Correlation analyses were conducted between the number of dumpsites and 18 social attributes (e.g., development level, urbanization, sanitation) obtained from official public data. Temporal analysis examined changes in dumpsite numbers over time (2015-2019) in four representative cities. The Spearman correlation was used to analyze the relationships between the number of dumpsites and various social attributes.

The BCA-Net model achieved an average sensitivity of over 98% in detecting dumpsites, outperforming existing models and exceeding human performance in this challenging task. The model's average precision was 70.1%, with higher precision observed for covered waste (96.7%) due to their distinct characteristics in satellite imagery. The model's speed allows for processing a large area (162 square kilometers) within 30 seconds on a personal laptop. Spatial analysis revealed significant correlations between the number of dumpsites and various social attributes. Specifically, the quantity of illegal dumpsites was positively correlated with lower development levels, lower urbanization rankings, and poorer sanitation. Interestingly, there were no statistically significant relationships between the number of dumpsites and population, education, or technology levels. Temporal analysis indicated that large-scale urban renewal programs and the implementation of waste classification policies influenced dumpsite numbers. Shanghai's waste classification policy, implemented in 2019, corresponded to a decrease in domestic waste dumpsites. In contrast, Kampala, Uganda, experienced a continued increase in dumpsites, potentially linked to inadequate waste management policies and socioeconomic factors. Class Activation Maps (CAMs) provided visual insights into the model's decision-making process, highlighting the areas of satellite images that contributed most to dumpsite predictions. The model's ability to identify and classify dumpsites efficiently demonstrates its potential for supporting global waste management strategies.

The findings demonstrate the effectiveness of the developed deep learning model in detecting and classifying dumpsites from satellite imagery, significantly improving the efficiency of waste management monitoring. The strong correlation between the number of dumpsites and development levels, urbanization, and sanitation highlight the critical role of economic and social factors in influencing waste management practices. The lack of correlation between dumpsite numbers and population, education, and technology suggests that effective policy implementation is paramount. The temporal analysis underscores the influence of government policies (waste classification) and large-scale urban development projects on dumpsite formation and changes. This research contributes a new, efficient method to assess the global waste distribution and its relationship with socioeconomic factors.

This research makes several key contributions. First, a globally representative fine-grained dumpsite dataset was created, providing a valuable resource for future research. Second, a novel deep learning model, BCA-Net, was developed and demonstrated high accuracy in dumpsite detection and classification. Third, large-scale spatial and temporal analyses revealed important correlations between dumpsite numbers and socioeconomic factors, including development, urbanization, and sanitation. This work has implications for improving waste management strategies and addressing the global waste crisis. Future research could focus on improving model precision through higher-resolution satellite imagery and exploring the impact of other factors (e.g., specific waste management policies) on dumpsite formation and changes. Further refinement of the GDI could also enhance its usability as a global indicator of waste management.

While the model demonstrated high sensitivity, the precision could be improved. This is primarily due to the limitations of current satellite image resolution, which hinders the identification of finer details distinguishing dumpsites from similar-looking objects. The diversity in dumpsite appearances across various cultures also presents a challenge. Although the model was tested on a globally representative sample of cities, the generalizability to all regions could be further evaluated with a broader dataset. The reliance on publicly available data for social attributes may introduce limitations in data accuracy and completeness.