Abstract
In this study, we utilize a self-supervised ResNet50 model for classify land-cover imagery combining learning from both Sentinel-1 and Sentinel-2 images using the SEN12MS dataset. The model is further fine-tuned with the DFC2020 dataset, enriched by our addition of 336 new rubber data patches covering 75×90 km² area. The model achieved 98.1% accuracy in classifying rubber using only a 25% training data
split, and 74.9% accuracy in classifying forest using only 5% training data split. This study demonstrates that accurate classification outcomes are achievable with a small number of labeled datasets by fine-tuning self-pretrained models and incorporating it with new labels. This approach not only showcases the model’s ability to efficiently learn from limited data, but also highlights the potential of adapting pre-existing models for land-cover classification tasks.
SDGs:
1. SDGs 4:Quality Education
2. SDGs 9:Industry, Innovation, and Infrastructure
3. SDGs 11:Sustainable Cities and Communities
4. SDGs 13:Climate Action
5. SDGs 15:Life on Land
Link Dokumen:
Download