PEMODELAN MASK REGION-BASED CNN PADA FOTO UDARA UNTUK PENGHITUNGAN JUMLAH INDIVIDU POHON DI HUTAN TANAMAN

Muhammad Rafindra Fatkhurohman, Dr. Emma Soraya, S.Hut., M.For.

2023 | Skripsi | KEHUTANAN

Informasi jumlah pohon berdiri di hutan tanaman, terutama pada spesies yang mempunyai nilai tinggi diperlukan untuk menghitung keuntungan. Teknologi penginderaan jauh resolusi sangat tinggi & deep learning telah banyak dimanfaatkan untuk mendapatkan informasi tersebut. Penelitian ini bertujuan untuk menguji kemampuan model mask region-based convolutional neural network (Mask R-CNN) dalam menghitung jumlah pohon hutan tanaman dari foto udara beresolusi sangat tinggi.
Foto udara akuisisi 2022 pada tanaman Jati Klon di Petak 13, KHDTK Wanagama I digunakan untuk membangun model Mask R-CNN sebagai data pelatihan. Interpretasi visual dilakukan pada individu pohon yang mewakili variasi luas dan warna tajuk pada foto udara yang menghasilkan 227 data individu pohon untuk pelatihan model. Untuk memperbanyak data latih, foto udara dibagi menjadi ukuran 128×128 piksel dengan overlap 50%, serta rotasi 180° tiap potongannya sehingga didapatkan 2516 potongan foto dan 4066 fitur. Model Mask R-CNN dengan arsitektur ResNet-50 digunakan untuk transfer learning dan pelatihan dilakukan selama 10 pengulangan (epochs). Uji akurasi dilakukan dalam dua cakupan: (1) Tingkat plot dengan 6 plot jari-jari 11,28 m yang disebar secara purposive; (2) Keseluruhan area penelitian seluas 5,4 Ha.
Hasil deteksi model mendapatkan 3955 individu pohon. Pengujian pertama mendapati nilai relatif Mean Absolute Error (rMAE) sebesar 15,8?ngan Mean Absolute Error (MAE) 4 pohon dan Root Mean Square Error (RMSE) sebesar 5 pohon. Pada pengujian kedua didapatkan akurasi 90%. Prediksi model menunjukkan penghitungan jumlah individu pohon melebihi jumlah pohon di lapangan (overestimate)

The information regarding the number of standing trees in plantation forests, particularly for high-value species, is crucial for profit calculation. Very high-resolution remote sensing technology and deep learning have been extensively employed to obtain such information. This study aims to assess the capability of the Mask Region-Based Convolutional Neural Network (Mask R-CNN) model in counting the number of trees in plantation forests from highly resolution aerial photography.
Aerial photographs acquired in 2022 of the Teak Clone plantation in Plot 13, KHDTK Wanagama I, were utilized to construct the Mask R-CNN model as training data. Visual interpretation was performed on individual trees representing variations in crown size and color in aerial photos, resulting in 227 individual tree data for model training. To augment the training data, aerial photos were divided into 128×128-pixel chips with a 50% overlap and 180° rotation between chips, yielding 2516 image chips and 4066 features. The Mask R-CNN model was developer utilizing ResNet-50 architecture for transfer learning, underwent training for 10 epochs. Accuracy testing was conducted in two scopes: (1) Plot level with 6 plots radius of 11.28 m purposively spread; (2) The entire research area covering 5.4 hectares.
The model's detection results revealed 3955 individual trees. In the first test, the relative Mean Absolute Error (rMAE) was 15.8%, with a Mean Absolute Error (MAE) of 4 trees and a Root Mean Square Error (RMSE) of 5 trees. In the second test, an accuracy of 90% was obtained. The model predictions indicated an overestimate in the count of individual trees compared to the actual field count.

Kata Kunci : Deteksi individu pohon, CNN, Deep learning; Individual tree detection, CNN, Deep learning