Pendugaan Normalized Difference Vegetation Index Berbasis Convolutional Neural Network, Interpolasi Spasial, dan Seleksi Fitur (Studi Kasus: Taman Nasional Lorentz, Papua)

Abstract

Taman Nasional Lorentz merupakan kawasan konservasi terbesar di Asia Tenggara dengan tingkat keragaman hayati yang sangat tinggi serta bentang alam yang kompleks. Pemantauan vegetasi di wilayah ini menghadapi tantangan berat akibat topografi yang curam, keterbatasan akses lapangan, serta tutupan awan konvektif yang berlangsung sepanjang tahun. Kondisi tersebut menyebabkan observasi satelit seperti NDVI sering tidak lengkap, tidak konsisten secara temporal, serta mengalami gangguan kualitas citra. Hilangnya data NDVI, baik secara spasial maupun temporal, menghambat pemantauan kesehatan vegetasi, deteksi perubahan tutupan lahan, serta penyusunan sistem peringatan dini terhadap potensi degradasi ekosistem. Kondisi tersebut menunjukkan perlunya metode alternatif yang mampu menghasilkan pendugaan NDVI secara konsisten pada kondisi atmosfer dan geografis yang tidak mendukung. Keterbatasan data NDVI di wilayah tropis pegunungan mendorong pengembangan pendekatan pendugaan berbasis data iklim dan kecerdasan buatan. Pendekatan tersebut memanfaatkan data reanalisis iklim ERA5, seleksi fitur berbasis Particle Swarm Optimization–K-Nearest Neighbor (PSO-KNN), serta arsitektur Convolutional Neural Network tiga dimensi (3D CNN). Variabel bioklimatik BIO1–BIO19 yang diturunkan dari ERA5 diinterpolasi menggunakan metode 2D nearest-neighbor, bilinear, dan bikubik untuk menyelaraskan resolusi spasialnya dengan data NDVI. Selanjutnya, seleksi fitur berbasis PSO–KNN diterapkan untuk mengidentifikasi variabel bioklimatik yang memberikan kontribusi terbesar terhadap perubahan NDVI. Integrasi interpolasi spasial, PSO–KNN, dan CNN dilakukan untuk mengatasi tiga permasalahan utama dalam pendugaan NDVI, yaitu hilangnya informasi optik akibat tutupan awan, multikolinearitas antarvariabel bioklimatik, serta kompleksitas pola spasial-temporal. PSO–KNN digunakan untuk memilih variabel bioklimatik yang paling berpengaruh sehingga model dibangun menggunakan prediktor yang relevan. Hasil seleksi fitur menunjukkan bahwa variabel presipitasi kuartalan (BIO16-BIO19) serta variabilitas suhu (BIO3 dan BIO4) secara konsisten menjadi prediktor utama yang memengaruhi dinamika vegetasi. Hubungan antara variabel bioklimatik dan NDVI dimodelkan menggunakan arsitektur 3D CNN yang mampu menangkap keterkaitan spasial dan temporal secara simultan melalui representasi data empat dimensi (4D). Pendugaan NDVI tetap dapat dilakukan meskipun data optik tidak tersedia. Penelitian ini difokuskan pada pengembangan kerangka pendugaan NDVI yang mengintegrasikan variabel bioklimatik hasil interpolasi, seleksi fitur berbasis PSO–KNN, serta enam arsitektur 3D CNN, yaitu ResNet-50, ResNet-18, GoogLeNet, TimeConvNet, CNN-based, dan LeNet. Pengujian model dilakukan menggunakan data periode 1994–2023 dengan pembagian 85% sebagai data pelatihan dan 15% sebagai data pengujian. Kinerja setiap model dievaluasi berdasarkan mean squared error (MSE), root mean squared error (RMSE), mean absolute error (MAE), koefisien determinasi (R²), dan waktu pelatihan. Penelitian ini bertujuan mengidentifikasi metode interpolasi yang menghasilkan tingkat galat terendah pada pendugaan NDVI berbasis CNN, menentukan variabel bioklimatik yang paling berpengaruh melalui seleksi fitur PSO–KNN, serta mengevaluasi arsitektur 3D CNN yang memberikan kinerja terbaik. Hasil penelitian menunjukkan bahwa setiap tahap pemodelan menghasilkan metode yang berbeda sebagai pendekatan terbaik. Pada tahap seleksi fitur, interpolasi bilinear menghasilkan nilai fungsi objektif PSO–KNN terendah sehingga memberikan kinerja paling stabil. Berdasarkan hasil seleksi tersebut, BIO16-BIO19 dan BIO3-BIO4 diidentifikasi sebagai variabel bioklimatik yang paling berpengaruh terhadap NDVI. Sebaliknya, pada tahap pendugaan NDVI, kombinasi interpolasi 2D nearest-neighbor dan arsitektur ResNet-50 menghasilkan kinerja terbaik dengan nilai R² sebesar 0,9496 dan MSE sebesar 0,0001. Perbedaan hasil tersebut menunjukkan bahwa metode interpolasi terbaik bergantung pada tujuan setiap tahapan pemodelan. Interpolasi yang menghasilkan nilai fungsi objektif terbaik pada tahap seleksi fitur tidak selalu memberikan akurasi pendugaan tertinggi pada tahap pemodelan CNN karena kedua tahapan menggunakan kriteria evaluasi yang berbeda. Evaluasi terhadap enam arsitektur 3D CNN menunjukkan bahwa ResNet-50 menghasilkan galat pendugaan paling rendah, sedangkan ResNet-18, TimeConvNet, dan CNN-based menawarkan efisiensi komputasi yang lebih baik. Setiap arsitektur CNN menunjukkan karakteristik kinerja yang berbeda dalam pendugaan NDVI. LeNet belum mampu mempelajari hubungan spasial-temporal yang kompleks sehingga menunjukkan gejala underfitting, sedangkan GoogLeNet belum mampu memanfaatkan karakteristik variabel bioklimatik hasil interpolasi secara optimal. Secara keseluruhan, integrasi seleksi fitur berbasis PSO–KNN dan arsitektur ResNet-50 mampu meningkatkan akurasi pendugaan NDVI di wilayah tropis dengan tutupan awan yang tinggi.

Lorentz National Park is the largest conservation area in Southeast Asia, characterized by exceptionally high biodiversity and a highly complex landscape. Vegetation monitoring in this region is particularly challenging due to steep topography, limited field access, and persistent convective cloud cover throughout the year. These conditions often result in incomplete, temporally inconsistent, and low-quality satellite observations, including those derived from the Normalized Difference Vegetation Index (NDVI). The loss of NDVI information, both spatially and temporally, hinders vegetation health monitoring, land-cover change detection, and the implementation of early warning systems for ecosystem degradation. These limitations highlight the need for an alternative approach that can provide consistent NDVI estimates under unfavorable atmospheric and geographical conditions. The limited availability of NDVI observations in tropical mountainous regions has encouraged the development of an estimation framework that combines climate reanalysis data with artificial intelligence techniques. This framework integrates ERA5 climate reanalysis data, feature selection using Particle Swarm Optimization–K-Nearest Neighbor (PSO–KNN), and three-dimensional Convolutional Neural Network (3D CNN) architectures. The BIO1–BIO19 bioclimatic variables derived from ERA5 were spatially interpolated using the 2D nearest-neighbor, bilinear, and bicubic methods to match the spatial resolution of the NDVI dataset. Subsequently, PSO–KNN was applied to identify the bioclimatic variables that contributed most significantly to NDVI variability. The integration of spatial interpolation, PSO–KNN, and CNN was designed to address three major challenges in NDVI estimation: the loss of optical information caused by persistent cloud cover, multicollinearity among bioclimatic variables, and the complex spatiotemporal relationships governing vegetation dynamics. PSO–KNN was employed to identify the most influential bioclimatic variables, enabling the model to be developed using only relevant predictors. Feature selection consistently identified quarterly precipitation variables (BIO16– BIO19) and temperature variability (BIO3 and BIO4) as the primary drivers of vegetation dynamics. The relationship between the selected bioclimatic variables and NDVI was modeled using a 3D CNN architecture capable of simultaneously capturing spatial and temporal dependencies through four-dimensional data representations. This approach enables NDVI estimation even when optical observations are unavailable. The study focused on developing an NDVI estimation framework that integrates interpolated bioclimatic variables, PSO–KNN-based feature selection, and six 3D CNN architectures: ResNet-50, ResNet-18, GoogLeNet, TimeConvNet, CNNbased, and LeNet. Model evaluation was conducted using data from 1994 to 2023, with 85% allocated for training and 15% for testing. Model performance was assessed using the mean squared error (MSE), root mean squared error (RMSE), mean absolute error (MAE), coefficient of determination (R²), and training time. The objectives of this study were to identify the interpolation method that minimizes estimation error in CNN-based NDVI prediction, to determine the most influential bioclimatic variables using PSO–KNN feature selection, and to evaluate the 3D CNN architecture that provides the best predictive performance. The results demonstrate that different stages of the modeling framework require different optimal approaches. During the feature selection stage, bilinear interpolation produced the lowest PSO–KNN objective function value, indicating the most stable performance. Based on the selected features, BIO16–BIO19 and BIO3–BIO4 were identified as the most influential bioclimatic variables for NDVI estimation. In contrast, during the NDVI estimation stage, the combination of 2D nearest-neighbor interpolation and the ResNet-50 architecture achieved the best predictive performance, with an R² of 0.9496 and an MSE of 0.0001. These findings indicate that the optimal interpolation method depends on the objective of each modeling stage. The interpolation method that yielded the best objective function value during feature selection did not necessarily produce the highest prediction accuracy in CNN-based NDVI estimation, as each stage employed different evaluation criteria. Among the six evaluated 3D CNN architectures, ResNet-50 achieved the lowest estimation error, whereas ResNet-18, TimeConvNet, and CNN-based provided better computational efficiency. Each CNN architecture exhibited distinct performance characteristics for NDVI estimation. LeNet was unable to capture complex spatiotemporal relationships adequately and therefore exhibited underfitting, whereas GoogLeNet was less effective at exploiting the characteristics of the interpolated bioclimatic variables. Overall, the integration of PSO–KNN-based feature selection with the ResNet-50 architecture substantially improved NDVI estimation accuracy in tropical regions characterized by persistent cloud cover.