پیوندهای مفید
اخبار و اعلانات
آمار
| تعداد نشریات | 126 |
| تعداد شمارهها | 7,094 |
| تعداد مقالات | 76,238 |
| تعداد مشاهده مقاله | 151,694,741 |
| تعداد دریافت فایل اصل مقاله | 113,802,574 |
Lightweight and Efficient Deep Learning Models for Remote Sensing-Based Land Use and Land Cover Classification: A Case Study on EuroSAT Dataset | ||
| Earth Observation and Geomatics Engineering | ||
| دوره 8، شماره 2، اسفند 2024 اصل مقاله (863.99 K) | ||
| نوع مقاله: Original Article | ||
| شناسه دیجیتال (DOI): 10.22059/eoge.2025.396331.1179 | ||
| نویسندگان | ||
| Mohammad Alikhani1؛ Behnam Ebadati* 2؛ Reza Attarzadeh2 | ||
| 1Faculty of Geodesy and Geomatics Engineering, K. N. Toosi University of Technology, Tehran, Iran | ||
| 2Department of Geomatics Engineering, ST.C., Islamic Azad University, Tehran, Iran | ||
| چکیده | ||
| One of the primary applications of remote sensing is the classification of land use and land cover (LULC). This field has increased prominence in computer vision tasks with the expansion of learning methods, especially in recent years. In this context, the present study introduces a convolutional neural network architecture named HCNN, designed to achieve high accuracy in LULC classification while minimizing processing cost. A simpler architecture based on conventional CNN designs, referred to as SCNN, was also implemented for comparison. The HCNN architecture comprises residual, dense, inception, and squeeze-excitation blocks, along with several base layers. Four models based on these two architectures were trained using RGB and multispectral data from the Sentinel-2 imagery in the EuroSAT dataset: SCNN-RGB, SCNN-MS, HCNN-RGB, and HCNN-MS. All models achieved accuracies above 94%. Among these, HCNN-MS attained the highest accuracy of 98.44%, while SCNN-RGB recorded the lowest accuracy at 95.57%. Overall, HCNN-based models outperformed SCNN-based models in accuracy and training speed, requiring approximately six times less training time. Additionally, the use of multispectral data had a positive impact on model accuracy, albeit it increased computational complexity somewhat. An ablation study was also conducted to evaluate the role of each block in the HCNN architecture on the model’s final performance. The ablation results demonstrated that each block plays a significant role in improving accuracy and reducing processing overhead, particularly the residual and dense blocks, which had the greatest impact on final accuracy. Moreover, the squeeze-excitation block notably reduced training time, while its removal caused minimal change in accuracy. | ||
| کلیدواژهها | ||
| Remote Sensing؛ Sentinel-2؛ Multispectral؛ LULC Classification؛ Deep Learning | ||
|
آمار تعداد مشاهده مقاله: 513 تعداد دریافت فایل اصل مقاله: 375 |
||
سامانه مدیریت نشریات علمی. قدرت گرفته از سیناوب