Joint Deep Learning for Land Cover and Use Classification
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Current methods for land cover (LC) and land use (LU) classification from very fine spatial resolution (VFSR) remotely sensed imagery face challenges due to spectral and spatial complexity, with traditional approaches often resulting in limited accuracy and overlooking semantic functions or spatial configurations, especially in urban settings.
Innovation Solution
A joint deep learning framework that incorporates a multilayer perceptron (MLP) and convolutional neural network (CNN) via a Markov process, iteratively refining both LC and LU classifications to enhance spatial and spectral feature representation, allowing for simultaneous classification of LC and LU.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Productivity
If traditional pixel-based classification methods are used for land cover classification from VFSR imagery, then the classification process is simple and fast, but the classification accuracy is limited due to spectral complexity and speckle noise
Solution Approach 1:
The patent combines pixel-based classification with object-based classification approaches. The system performs both pixel-level spectral analysis and object-level spatial analysis, then integrates the results through a joint deep learning framework. This merging allows the system to maintain the speed advantages of pixel-based methods while achieving the accuracy benefits of object-based methods by considering both spectral characteristics and spatial context.
Solution Approach 2:
The patent segments the VFSR imagery into homogeneous objects using unsupervised segmentation algorithms before classification. This segmentation groups spectrally similar pixels into coherent objects, reducing the impact of speckle noise and intra-object spectral variability. The segmented objects then serve as the basis for both pixel-based and object-based classification, improving overall accuracy while maintaining processing efficiency.
2Measurement precision
If object-based image analysis methods are used for land cover classification, then classification accuracy improves by considering spatial context, but the selection of segmentation scales becomes complex and time-consuming
Solution Approach 1:
The patent implements dynamic scale selection where the segmentation scale is automatically adjusted based on the local characteristics of the imagery and the classification task. The system uses multi-scale segmentation that adapts to different regions, performing finer segmentation in heterogeneous areas and coarser segmentation in homogeneous areas. This dynamic approach eliminates the need for manual scale selection and handles the complexity of varying spatial contexts automatically.
Solution Approach 2:
The system performs self-adjustment of segmentation parameters through iterative optimization during the classification process. The deep learning framework learns optimal segmentation scales from training data and automatically configures the segmentation parameters without requiring manual intervention. This self-service capability simplifies the workflow while maintaining high classification accuracy across different land cover types.
3Ease of manufacture
If traditional methods classify land cover and land use separately, then each classification can be optimized independently, but the semantic relationships and spatial configurations between LC and LU are overlooked
Solution Approach 1:
The patent merges the land cover classification and land use classification into a unified joint deep learning framework. The system simultaneously processes both classification tasks and models the conditional dependencies between LC and LU. By integrating the two classification processes, the system captures semantic relationships and spatial configurations that would be lost in separate classifications, while maintaining computational efficiency through shared feature extraction and parallel processing.
4Manufacturing precision
If very fine spatial resolution imagery is used for detailed LULC information extraction, then the level of detail increases, but the spectral and spatial complexity of the imagery increases making classification more difficult
Solution Approach 1:
The patent applies multi-scale segmentation to VFSR imagery, dividing the high-resolution data into homogeneous objects at appropriate scales. This segmentation reduces the complexity by grouping spectrally and spatially similar pixels, making the subsequent classification more tractable. The segmented objects serve as meaningful units that capture the spatial structure of the imagery while reducing the dimensionality of the classification problem.
Solution Approach 2:
The patent transitions from pixel-level classification to object-level classification by introducing a spatial dimension to the analysis. Instead of classifying individual pixels in VFSR imagery, the system classifies segmented objects that represent meaningful spatial units. This dimensionality change from 2D pixel grid to 3D object space (including spatial extent) simplifies the classification by aggregating spectral information across multiple pixels while preserving spatial context.
Data Source
AI summary
Land cover (LC) and land use (LU) have commonly been classified separately from remotely sensed imagery, without considering the intrinsically hierarchical and nested relationships between them. A novel joint deep learning framework is proposed and demonstrated for LC and LU classification. The proposed Joint Deep Learning (JDL) model incorporates a multilayer perceptron (MLP) and convolutional neutral network (CNN), and is implemented via a Markov process involving iterative updating. In the JDL, LU classification conducted by the CNN is made conditional upon the LC probabilities predicted by the MLP. In turn, those LU probabilities together with the original imagery are re-used as inputs to the MLP to strengthen the spatial and spectral feature representation. This process of updating the MLP and CNN forms a joint distribution, where both LC and LU are classified simultaneously through iteration.