Abstract
Land cover maps are of vital importance to various fields such as land use policy development, ecosystem services, urban planning and agriculture monitoring, which are mainly generated from remote sensing image classification techniques. Traditional land cover classification usually needs tremendous computational resources, which often becomes a huge burden to the remote sensing community. Undoubtedly cloud computing is one of the best choices for land cover classification, however, if not managed properly, the computation cost on the cloud could be surprisingly high. Recently, cutting the unnecessary computation long tail has become a promising solution for saving the cost in the cloud. For land cover classification, it is generally not necessary to achieve the best accuracy and 85% can be regarded as a reliable land cover classification. Therefore, in this paper, we propose a framework for cost-effective remote sensing classification. Given the desired accuracy, the clustering algorithm can stop early for cost-saving whilst achieving sufficient accuracy for land cover image classification. Experimental results show that achieving 85%-99.9% accuracy needs only 27.34%-60.83% of the total cloud computation cost for achieving a 100% accuracy. To put it into perspective, for the US land cover classification example, the proposed approach can save over $1,593,490.18 for the government in each single-use when the desired accuracy is 90%.
Introduction: The Problem of the "Long Tail"
Generating accurate land cover maps from satellite imagery—identifying forests, wetlands, and urban sprawl—is a computationally intensive endeavor. While cloud computing offers the necessary scale, the associated costs can be prohibitive for large-scale or frequent monitoring.
The primary inefficiency identified in this research is the "long tail" phenomenon inherent in unsupervised clustering algorithms. As shown in Figure 1, the algorithm achieves significant accuracy gains in the initial iterations, but then enters a phase of diminishing returns. It often spends the vast majority of its processing time chasing the final 1-5% of accuracy.
For many practical applications, such as urban planning or agricultural monitoring, an accuracy of 85% to 90% is considered highly reliable. However, standard algorithms lack a mechanism to predict their current accuracy level in real-time, forcing them to run until full convergence and wasting significant resources.
Proposed Framework
To address this, the authors propose a two-phase framework designed to predict accuracy and enable early stopping. The system is built around the Fuzzy C-means (FCM) clustering algorithm, which is particularly suited for remote sensing as geographical pixels often represent a mixture of land types.
The framework operates in two main stages:
- Training Phase:
- Accuracy Proxy: Since ground truth labels are typically unavailable during unsupervised clustering, the Rand Index is used to measure the similarity between intermediate results and the final converged state (defined as 100% accuracy).
- Anomaly Mitigation: The Local Outlier Factor (LOF) algorithm is applied to training data to remove noise in the relationship between the algorithm's objective function and the achieved accuracy.
- Regression Modeling: A Support Vector Regressor (SVR) is trained to learn the mathematical relationship between the change rate of the objective function and the actual accuracy.
- Testing Phase:
- During live classification, the system monitors the objective function's change rate.
- The SVR predicts the current accuracy.
- Once the user-defined accuracy threshold (e.g., 90%) is met, the algorithm terminates immediately.
Experimental Results and Economic Impact
The study utilized the SpaceNet dataset, encompassing high-resolution satellite imagery from cities like Las Vegas, Khartoum, Rio De Janeiro, and Shanghai.
The results demonstrate dramatic efficiency gains:
- Cost Savings: Achieving 85% accuracy required only 27.34% of the total computation cost. Even a near-perfect 99.9% accuracy was reached using only 60.83% of the resources required for full convergence.
- Reliability: The framework is conservative; for a 90% target, it delivered an average accuracy of 93.16%.
- Economic Scale: Extrapolating these results to a single-use classification of the entire United States at 90% accuracy, the authors estimate a potential saving of approximately $1,593,490.18 in cloud computing fees.
Conclusion and Future Work
This framework provides a practical solution for the remote sensing community to manage cloud costs without sacrificing operational reliability. Future research will focus on ensuring these regression models generalize across different sensors and geographical regions, and developing mathematical bounds for the predicted accuracy error.
Summary updated