Information Processing for Satellite Population Mapping with Multi-Task CNNs
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Census data is infrequently collected and has low spatial resolution, leading to inaccurate population prediction, and existing methods using satellite images and census data struggle with low accuracy in population prediction.
Innovation Solution
A learning model trained using satellite images, population data, and land classification data through a multi-task learning approach with a CNN, employing regression and classification techniques to predict population and land classification with high accuracy.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Ease of manufacture
If census data is collected infrequently (every 5-10 years) to reduce implementation cost, then the expense is reduced, but the population prediction accuracy deteriorates due to outdated information
Solution Approach 1:
The patent applies preliminary action by training a deep learning model in advance using historical satellite images and census data. This pre-trained model can then continuously predict population changes using new satellite images without requiring frequent census collections, thus maintaining high prediction accuracy while reducing the frequency and cost of actual census implementations.
Solution Approach 2:
The patent introduces satellite images as an intermediary between census data collection and population prediction. The deep learning model learns to extract population information from satellite imagery, which serves as a continuous proxy for traditional census data. This intermediary enables ongoing population monitoring without the need for frequent expensive census operations.
2Productivity
If satellite images are used to predict population with a learning model trained only on population data, then the prediction can be made continuously, but the prediction accuracy remains low
Solution Approach 1:
The patent merges multiple data sources and training objectives by combining satellite images with additional auxiliary data in the training process. The multi-task learning framework integrates population prediction with land classification tasks, allowing the model to learn comprehensive spatial patterns that improve population prediction accuracy while maintaining continuous prediction capability.
Solution Approach 2:
The patent creates a universal deep learning model that performs multiple functions simultaneously - population prediction and land classification. This multi-functional model leverages shared features from satellite imagery to improve both tasks, with the land classification capability enhancing the overall population prediction accuracy by providing contextual environmental information.
Data Source
Figure 1
Figure 2
Figure 3A~3C
AI summary
An information processing apparatus (1) includes: an acquisition means (12) for acquiring a satellite image; a first generation means (13) for recursively predicting and generating a first image representing a distribution of population values with respect to the satellite image, through machine learning, using the satellite image as an input; and a second generation means (13) for predicting and generating a second image representing probabilities of land types with respect to the satellite image, through the machine learning.