High Resolution Land Cover Classification Using Spatio-Temporal ML
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Current methods for providing High Resolution Land Cover (HRLC) face challenges in efficiently and accurately producing detailed land cover datasets from high-resolution images and medium-resolution images taken over time.
Innovation Solution
The approach involves receiving a high-resolution image and a series of medium-resolution images, performing segmentation on the high-resolution image, applying various transformations to the medium-resolution images, creating a temporal stack layer with spatial statistics, and using a Machine Learning model to classify land cover based on spatio-temporal layers.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If high-resolution images are used for land cover classification, then classification precision is improved, but data processing complexity increases
Solution Approach 1:
The patent segments the high-resolution image into multiple tiles or patches, processing each segment independently through the neural network. This divides the complex task of processing entire high-resolution images into manageable chunks, reducing memory requirements and computational complexity while maintaining classification precision through careful segmentation and stitching of results
Solution Approach 2:
The patent introduces temporal dimension by incorporating medium-resolution satellite images from multiple time points alongside the high-resolution reference image. This spatio-temporal approach adds a temporal dimension to the classification process, enabling the model to leverage temporal patterns and reduce processing complexity on single high-resolution images
2Loss of time
If multiple medium-resolution images over time are processed, then temporal resolution is improved, but computational requirements increase
Solution Approach 1:
The patent merges multiple medium-resolution satellite images taken at different time points into a unified temporal stack that is processed together with the high-resolution reference image. This combining approach leverages temporal information from multiple observations while sharing computational resources across the merged dataset, reducing redundant processing compared to analyzing each image separately
Solution Approach 2:
The patent uses the high-resolution reference image as a template or copy that anchors the classification process. By aligning and registering medium-resolution images to this high-resolution template, the system avoids re-processing high-resolution data multiple times, instead copying spatial structures from the reference and updating them with temporal information from medium-resolution images
Data Source
AI summary
High Resolution Land Cover (HRLC) may be provided. A first high resolution image of a scene may be received and a plurality of second medium resolution images of the scene taken over a period of time may be received. A segmentation of the first image may then be performed. Next, a plurality of transformations may be performed on the plurality of second images. Then a temporal stack layer for a plurality of temporal statistics for each of the plurality of transformations on the plurality of second images may be created. The temporal stack layer may then be intersected with the segmented first image using a plurality of spatial statistics to produce a plurality of spatio-temporal layers. A Machine Learning (ML) model may be used on the spatio-temporal layers to classify pixels in the first image into respective ones of a plurality of classifications.