Nanopublications LDF server

Matches in Nanopublications for { ?s ?p <https://doi.org/10.1016/j.jag.2024.104034> ?g. }

• CREATOR quotes j.jag.2024.104034 assertion.
• BiAU-Net%20achieved%20improvements%20over%20the%20Fire_cci51%20baseline%20of%2011.56%25%20in%20Overall%20Accuracy%2C%2029.08%25%20in%20Precision%2C%207.06%25%20in%20Recall%2C%2019.90%25%20in%20F1-score%2C%2015.44%25%20in%20Balanced%20Accuracy%2C%2029.90%25%20in%20Kappa%20Coefficient%2C%20and%2028.29%25%20in%20Matthews%20Correlation%20Coefficient%20for%20wildfire%20burnt%20area%20mapping. obtainsSupportFrom j.jag.2024.104034 assertion.
• The%20BiAU-Net%20model%20demonstrated%20good%20generalizability%20across%20five%20testing%20areas%20in%20different%20continents%20%28United%20States%2C%20Spain%2C%20Australia%2C%20Indonesia%2C%20and%20Kenya%29%20when%20trained%20on%20a%20single%20wildfire%20event%20in%20California. obtainsSupportFrom j.jag.2024.104034 assertion.
• Bi-temporal%20input%20incorporating%20both%20pre-fire%20and%20post-fire%20Sentinel-2%20imagery%20enhanced%20model%20performance%20across%20diverse%20environmental%20areas%20compared%20to%20traditional%20single-input%20U-Net%20models. obtainsSupportFrom j.jag.2024.104034 assertion.
• The%20combination%20of%20Dice%20Loss%20and%20Focal%20Loss%20as%20the%20loss%20function%20significantly%20improved%20model%20performance%20in%20handling%20imbalanced%20datasets%2C%20with%20BiAU-Net%20achieving%20a%20Balanced%20Accuracy%20score%20of%200.965%20compared%20to%200.930%20for%20BiU-Net%20in%20the%20Chico%20wildfire%20case. obtainsSupportFrom j.jag.2024.104034 assertion.
• Attention%20mechanisms%20in%20the%20U-Net%20architecture%20enabled%20the%20model%20to%20focus%20on%20burnt%20areas%20and%20improve%20accuracy%20and%20efficiency%2C%20especially%20in%20detecting%20edges%20and%20small%20areas%20where%20burnt%20and%20non-burnt%20pixels%20are%20mixed%20together. obtainsSupportFrom j.jag.2024.104034 assertion.
• The%20BiAU-Net%20model%20achieved%20the%20highest%20overall%20performance%20compared%20to%20state-of-the-art%20wildfire%20burnt%20area%20detection%20models%2C%20as%20evidenced%20by%20the%20highest%20F1-score%20and%20Matthews%20Correlation%20Coefficient%20values. obtainsSupportFrom j.jag.2024.104034 assertion.
• BiAU-Net%20using%20Sentinel-2%20imagery%20at%2020%20m%20spatial%20resolution%20offers%20higher%20accuracy%20than%20global%20wildfire%20products%20Fire_cci51%20and%20MODIS%20MCD64A1%2C%20which%20have%20limitations%20at%20250%20m%20and%20500%20m%20resolution%20and%20often%20overlook%20small%20fires. obtainsSupportFrom j.jag.2024.104034 assertion.
• The%20BiAU-Net%20model%20was%20trained%20using%20Sentinel-2%20MSI%20Level%202A%20product%20with%20a%20false%20color%20band%20combination%20of%20Band12%20%28SWIR%3A%202114.9-2289.9%20%CE%BCm%29%2C%20Band11%20%28SWIR%3A%201568.2-1659.2%20%CE%BCm%29%2C%20and%20Band8A%20%28Narrow%20NIR%3A%20854.2-875.2%20%CE%BCm%29%20at%2020%20m%20spatial%20resolution. obtainsSupportFrom j.jag.2024.104034 assertion.
• For%20the%20Chico%2C%20California%20test%20area%2C%20BiAU-Net%20with%20DLoss%2BFLoss%20achieved%20an%20Overall%20Accuracy%20of%200.968%2C%20Precision%20of%200.974%2C%20Recall%20of%200.949%2C%20F1-score%20of%200.961%2C%20Balanced%20Accuracy%20of%200.965%2C%20Kappa%20Coefficient%20of%200.934%2C%20and%20Matthews%20Correlation%20Coefficient%20of%200.929. obtainsSupportFrom j.jag.2024.104034 assertion.
• The%20model%20training%20utilized%20a%20batch%20size%20of%206%2C%20conducted%2050%20epochs%2C%20initialized%20the%20learning%20rate%20at%200.1%20with%20scheduled%20decrease%20by%20a%20factor%20of%200.5%20between%20the%2030th%20to%2050th%20epochs%2C%20and%20used%20Stochastic%20Gradient%20Descent%20optimizer. obtainsSupportFrom j.jag.2024.104034 assertion.
• BiAU-Net%20outperformed%20traditional%20machine%20learning%20models%20including%20Support%20Vector%20Machines%20and%20Random%20Forest%2C%20which%20require%20handcrafted%20engineering%20and%20often%20perform%20well%20in%20training%20areas%20but%20poorly%20when%20transferred%20to%20different%20areas%20due%20to%20varying%20environments. obtainsSupportFrom j.jag.2024.104034 assertion.
• The%20training%20dataset%20for%20BiAU-Net%20included%20306%20tiles%20from%20the%20Chico%20wildfire%20with%20278%20tiles%20%28approximately%2091%25%29%20used%20for%20training%20and%2028%20tiles%20%28approximately%209%25%29%20for%20validation%2C%20with%20each%20tile%20including%20256%C3%97256%20pixels. obtainsSupportFrom j.jag.2024.104034 assertion.
• BiAU-Net%20demonstrated%20superior%20performance%20in%20detecting%20small%20burnt%20areas%20and%20boundary%20regions%20where%20burnt%20and%20non-burnt%20pixels%20are%20mixed%20together%2C%20addressing%20key%20limitations%20of%20previous%20deep%20learning%20models. obtainsSupportFrom j.jag.2024.104034 assertion.
• The%20code%20and%20dataset%20for%20BiAU-Net%20are%20publicly%20available%20on%20GitHub%20at%20https%3A%2F%2Fgithub.com%2FTangSui122%2FBi-temporal-Wildfire-Burnt-Area-Detection.git%2C%20facilitating%20reproducibility%20and%20follow-on%20research%20efforts. obtainsSupportFrom j.jag.2024.104034 assertion.
• Focal%20Loss%20addresses%20class%20imbalance%20by%20down-weighting%20easy-to-classify%20examples%20and%20placing%20more%20emphasis%20on%20hard-to-classify%20examples%20through%20tunable%20parameters%20%CE%B1%20%28weight%20between%20positive%20and%20negative%20samples%29%20and%20%CE%B3%20%28focusing%20parameter%20for%20adjusting%20loss%20magnitude%29. obtainsSupportFrom j.jag.2024.104034 assertion.
• In%20test%20areas%20with%20significantly%20different%20environmental%20conditions%20from%20the%20training%20area%20%28such%20as%20Nairobi%2C%20Kenya%20with%20red%20soil%29%2C%20bi-temporal%20input%20enabled%20the%20model%20to%20consider%20spectral%20changes%20caused%20by%20wildfires%20as%20features%2C%20achieving%20an%20Overall%20Accuracy%20of%200.816%20despite%20minimal%20spectral%20similarity. obtainsSupportFrom j.jag.2024.104034 assertion.
• Data%20augmentation%20methods%20including%20rotating%20from%2010%C2%B0%20left%20to%2010%C2%B0%20right%20with%20mirroring%2C%20reversing%20left%20and%20right%2C%20and%20zooming%20by%20a%20factor%20of%200.85%20were%20implemented%20to%20enhance%20model%20training. obtainsSupportFrom j.jag.2024.104034 assertion.
• For%20areas%20with%20similar%20environmental%20conditions%20to%20the%20training%20area%20%28Bredbo%2C%20Australia%20and%20Tarragona%2C%20Spain%29%2C%20BiAU-Net%20achieved%20Overall%20Accuracy%20of%200.941%20and%200.939%20respectively%2C%20demonstrating%20robust%20performance%20across%20Mediterranean%20and%20temperate%20forest%20ecosystems. obtainsSupportFrom j.jag.2024.104034 assertion.
• The%20BiAU-Net%20architecture%20incorporates%20attention%20gates%20at%20each%20up-sampling%20step%20that%20dynamically%20assess%20feature%20relevance%20to%20create%20attention%20weights%2C%20amplifying%20significant%20features%20and%20reducing%20the%20influence%20of%20unnecessary%20ones. obtainsSupportFrom j.jag.2024.104034 assertion.
• The%20model%20processes%20pre-fire%20and%20post-fire%20satellite%20images%20through%20two%20symmetrical%20convolutions%20and%20max-pooling%20in%20the%20encoder%20phase%20to%20extract%20features%2C%20with%20four%20rounds%20of%20down-sampling%20producing%20four%20pairs%20of%20pre-%20and%20post-fire%20feature%20maps%20across%20varying%20scales. obtainsSupportFrom j.jag.2024.104034 assertion.