Uncertainty Assessment of Hyperspectral Image Classification: Deep Learning vs. Random Forest.

Area of Science:

• Remote Sensing
• Geospatial Data Analysis
• Machine Learning for Earth Observation

Background:

• Traditional accuracy assessment for image classification relies on confusion matrices and test data, failing to capture spatial error variations.
• Existing methods are limited by test data availability and cannot spatially characterize classification accuracy before validation.
• Assessing error propagation within classified imagery products is crucial for reliable geospatial data.

Purpose of the Study:

• To apply and compare two novel uncertainty assessment techniques for hyperspectral image classification.
• To evaluate techniques that do not require test data and enable spatial characterization of classification accuracy.
• To compare the performance of deep neural networks (DNNs) against random forest (RF) using these uncertainty measures.

Main Methods:

• Implemented Shannon entropy calculation on class probabilities predicted by DNN and RF for each pixel.
• Quantified classification uncertainties for DNN and RF on the Salinas and Indian Pines hyperspectral datasets.
• Compared derived uncertainty estimates against classification accuracy using a modified root mean square error (RMSE).

Main Results:

• The Shannon entropy-based uncertainty assessment effectively characterizes spatial classification accuracy without test data.
• Deep neural networks (DNNs) demonstrated superior performance in estimating classification accuracy compared to random forests (RF).
• Pixel-level uncertainty derived from DNNs using Shannon entropy proved to be a better indicator of classification accuracy.

Conclusions:

• Emerging deep neural networks (DNNs) offer enhanced capabilities for accurate hyperspectral image classification.
• Shannon entropy is a robust metric for pixel-level uncertainty estimation in image classification.
• The proposed uncertainty assessment methods improve the reliability and spatial understanding of classification accuracy in remote sensing.