Multi-Output Land Cover Classification Model Training
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Shallower networks face difficulty in distinguishing low-level land cover features due to gradient disappearance issues when deepening the network structure, which restricts the improvement of classification accuracy in remote sensing image classification.
Innovation Solution
A training method for a multi-output land cover classification model using deep belief nets, where multiple level outputs are distributed across different network layers, allowing for a weighted sum of losses to adjust parameters through backpropagation, thereby preventing gradient disappearance and improving parameter adjustment and classification accuracy.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If the network structure is deepened to improve classification accuracy, then the ability to distinguish low-level land cover features is improved, but gradient disappearance occurs making it difficult to adjust parameters of early layers
Solution Approach 1:
The patent segments the classification task into multiple output levels corresponding to different network layers. Each layer produces outputs at different classification granularities, allowing early layers to contribute to classification without requiring the entire deep network to be perfectly tuned. This segmentation enables gradient flow to meaningful targets at each level, preventing gradient disappearance.
Solution Approach 2:
The patent adds a temporal/dimensional dimension to the classification process by introducing multi-level outputs across different training stages. Instead of a single final classification, the model produces classifications at multiple levels (N level outputs) with different weight coefficients applied at different times during training, creating a new dimension for gradient propagation that prevents vanishing gradients.
2Ease of operation
If shallow network structure is used, then parameter adjustment is easier, but the network cannot distinguish low-level categories such as dry land and paddy fields
Solution Approach 1:
The patent makes the network dynamically adaptable by introducing time-varying weight coefficients for different output levels. During training, the model dynamically adjusts the importance of each output level's loss function, allowing early layers to learn effectively when their outputs are weighted more heavily, while still achieving high accuracy when all layers contribute in the final classification.
Data Source
AI summary
A training method for multi-output land cover classification model and a classification method are provided. The training method includes: obtaining a training data; inputting the training data into an initial model based on deep belief nets for training to obtain a multi-output land cover classification model, wherein the initial model includes N level outputs, and the N level outputs include an output set at last network layer and (N−1) level output set at any (N−1) network layers from a first network layer to a penultimate network layer of the initial model; determining a total loss according to losses of the N level outputs; performing a backpropagation based on the total loss to adjust a parameter of the initial model, N being an integer greater than or equal to 2. The gradient is not easy to disappear during backpropagation of the model, which is beneficial to improve classification accuracy.