Joint Deep Learning for Land Cover and Use Classification
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Current methods for land cover and land use classification from very fine spatial resolution remotely sensed imagery face challenges due to spectral and spatial complexity, particularly in urban areas, where traditional pixel-based and object-based approaches suffer from limited accuracy and the need for manual feature engineering, and deep learning methods like CNNs introduce artefacts and blurred boundaries.
Innovation Solution
A Joint Deep Learning framework that combines a multilayer perceptron (MLP) and convolutional neural network (CNN) via a Markov process, iteratively refining land cover and land use classifications, allowing for mutual reinforcement and accurate classification of both at the same time, addressing the hierarchical and nested relationships between them.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Ease of manufacture
If pixel-based classification methods are used to classify land cover from VFSR imagery, then classification can be performed based on spectral reflectance, but classification accuracy is limited due to speckle noise and increased inter-class variance
Solution Approach 1:
The patent combines pixel-based classification with object-based classification approaches. The system performs both pixel-level spectral classification and object-level classification using segmented image objects, then integrates the results through a joint probability framework. This merging allows the system to benefit from both spectral information (pixel-based) and spatial context (object-based), resolving the contradiction between ease of classification and classification accuracy.
2Measurement precision
If object-based image analysis is used for land cover classification, then classification accuracy can be improved through spatial context, but the selection of segmentation scales becomes complex and requires determining appropriate scales to avoid over- and under-segmentation
Solution Approach 1:
The patent employs an iterative process where the classification results from one iteration inform the segmentation and classification in the next iteration. The system automatically adjusts segmentation and classification parameters based on the joint probability distributions learned during iteration, eliminating the need for manual segmentation scale selection. This self-adjusting mechanism resolves the contradiction by making the system self-sufficient in determining appropriate segmentation scales.
3Ease of manufacture
If traditional LU classification methods are used, then classification can be performed using spectral characteristics, but the methods are insufficient to distinguish LUs composed of multiple LCs
Solution Approach 1:
The patent introduces a hierarchical classification framework that operates at multiple dimensions: pixel-level spectral dimension for land cover classification, and object-level spatial dimension for land use classification. By adding the spatial dimension and hierarchical structure, the system can distinguish LUs composed of multiple LCs, as it considers both the spectral characteristics of individual pixels and the spatial configuration of land cover types within segmented objects.
4Measurement precision
If deep learning methods like CNNs are used for classification, then classification accuracy can be improved, but artefacts and blurred boundaries are introduced
Solution Approach 1:
The patent applies image segmentation to divide the image into distinct objects before classification. This segmentation step creates well-defined boundaries for each object, which are then used as the basis for classification. By segmenting the image first, the system maintains sharp boundaries in the final classification results, avoiding the blurred boundaries that can occur with CNN-based approaches while still achieving high classification accuracy through the subsequent classification of segmented objects.
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.