Electronic Theses and Dissertation

Penelitian ini mengimplementasikan segmentasi semantik berbasis deep learning menggunakan arsitektur U-Net untuk mengklasifikasikan area sawah dan non-sawah pada citra drone dan citra penginderaan jauh resolusi tinggi (LoveDA rural). Dataset yang digunakan merupakan kombinasi dari Drone Camera Image Dataset of Agriculture Fields sebanyak 122 citra yang telah dilabeli secara manual menggunakan LabelMe, dan LoveDA Dataset subset Rural sebanyak 500 citra yang dipilih berdasarkan keberadaan kelas agriculture, dengan total 622 citra. Arsitektur U-Net yang diimplementasikan terdiri dari dua blok encoder dengan filter 64 dan 128, bottleneck dengan filter 256, serta dua blok decoder yang dilengkapi skip connection dengan total 1.883.137 parameter yang dapat dilatih. Model dilatih menggunakan kombinasi fungsi Loss Binary Crossentropy dan Dice Loss dengan optimizer Adam, learning rate 0,0001, batch size 8, dan 15 epoch. Model terbaik tersimpan pada epoch ke-15 dengan nilai validation loss terendah sebesar 0,5462. Hasil evaluasi pada data uji menunjukkan nilai Accuracy 70,43%, Intersection over Union (IoU) 54,11%, Precision 71,30%, dan Recall 69,10%. Analisis confusion matrix menunjukkan bahwa model mampu mengklasifikasikan area sawah dan non-sawah dengan tingkat ketepatan yang cukup seimbang. Hasil penelitian ini menunjukkan bahwa arsitektur U-Net mampu melakukan segmentasi semantik pada citra pertanian dengan tingkat ketepatan dan sensitivitas yang cukup baik dalam membedakan area sawah dan non-sawah.

This study implements deep learning-based semantic segmentation using the UNet architecture to classify paddy and non-paddy field areas from drone imagery and high-resolution remote sensing imagery (LoveDA rural). The dataset used is a combination of consisting of 122 images from the Drone Camera Image Dataset of Agriculture Fields, which were manually annotated using LabelMe, and 500 images from the LoveDA Rural dataset selected based on the presence of the agriculture class, resulting in a total of 622 images. The implemented U-Net architecture consists of two encoder blocks with 64 and 128 filters, a bottleneck layer with 256 filters, and two decoder blocks equipped with skip connections, resulting in a total of 1,883,137 trainable parameters. The model was trained using a combination of Binary Crossentropy and Dice Loss functions, with the Adam optimizer, a learning rate of 0.0001, a batch size of 8, and 15 training epochs. The best-performing model was obtained at epoch 15, achieving the lowest validation loss of 0.5462. Evaluation results on the test dataset showed an Accuracy of 70.43%, an Intersection over Union (IoU) of 54.11%, a Precision of 71.30%, and a Recall of 69.10%. Confusion matrix analysis indicate that the model was able to classify paddy field and non-paddy field areas with a relatively balanced level of accuracy. The results demonstrate that the U-Net architecture is capable of performing semantic segmentation on agricultural imagery with satisfactory precision and sensitivity in distinguishing between paddy field and non-paddy field areas.