Assessing the efficiency of pixel-based and object-based image classification using deep learning in an agricultural...

Area of Science:

• Remote Sensing
• Geospatial Analysis
• Deep Learning Applications

Background:

• Satellite technology advancements increase data availability, but efficient interpretation remains crucial.
• Deep learning methods are effective for image classification but depend on training data quality.
• Comparing pixel-based and object-based classification is essential for optimizing satellite image analysis.

Purpose of the Study:

• To compare the efficiency of pixel-based versus object-based classification for Sentinel-2 satellite imagery.
• To evaluate the performance of the Deeplabv3 deep learning model in both classification approaches.
• To assess the impact of image enhancement techniques on classification accuracy.

Main Methods:

• Utilized Sentinel-2 satellite imagery enhanced with a high-pass filter for improved sharpness.
• Trained the Deeplabv3 deep learning model using extracted training samples from the enhanced imagery.
• Implemented object-based classification with the majority zonal statistic method and compared it to pixel-based classification.

Main Results:

• Object-based classification achieved an accuracy of 83.5% (kappa: 0.791), slightly higher than pixel-based classification at 83.1% (kappa: 0.786).
• The Deeplabv3 model demonstrated efficient performance in both classification strategies.
• Image enhancement positively contributed to data visualization and preparation for deep learning.

Conclusions:

• Object-based classification, when integrated with deep learning classifiers like Deeplabv3, shows efficient performance in Sentinel-2 imagery analysis.
• The findings provide a valuable reference for future studies aiming to enhance accuracy and efficiency in satellite image interpretation.
• This research contributes to refining satellite image interpretation techniques for environmental applications.