Model Klasifikasi Komoditas dan Estimasi Hari Setelah Tanam untuk Padi dan Tebu Menggunakan Machine Learning

Abstract

Swasembada pangan merupakan kemampuan suatu wilayah dalam memenuhi kebutuhan pangan secara mandiri. Sektor pertanian merupakan penopang utama dari swasembada pangan. Walau demikian, sektor pertanian berpotensi mengalami penurunan kontribusi seiring peningkatan laju pertumbuhan populasi, degradasi lahan, dan perubahan iklim. Kemampuan adaptasi menjadi krusial dalam menghadapi perkembangan dan tantangan penurunan kontribusi pertanian. Salah satu tren penurunan kontribusi pernah terjadi di Provinsi Jawa Timur pada tahun 2019 hingga 2024. Provinsi Jawa Timur merupakan wilayah pertanian terluas di Indonesia, khususnya pada komoditas padi dan tebu. Akan tetapi, proses monitoring komoditas dan Hari Setelah Tanam (HST) di Provinsi Jawa Timur belum dilakukan menggunakan machine learning dan deep learning. Proses monitoring masih dilakukan berdasarkan survei oleh Penyuluh Pertanian Lapangan (PPL). Adaptasi proses monitoring dapat dilakukan dengan implementasi smart farming berbasiskan pertanian presisi dengan pendayagunaan remote sensing dan Artificial Intelligent (AI) untuk menghasilkan informasi yang tepat dan cepat. Penelitian ini bertujuan untuk pembuatan model klasifikasi komoditas dengan mengimplementasikan algoritma Random Forest (RF), Support Vector Machine (SVM), algoritma Extreme Gradient Boosting (XGBoost), serta Convolutional Neural Network (CNN). Hasil klasifikasi digunakan untuk estimasi HST menggunakan analisis phenology indeks vegetasi tanaman. Penelitian dilakukan dengan empat tahapan inti yaitu, pengumpulan dan seleksi data, eksplorasi dan praproses data, pemodelan, serta evaluasi. Data yang digunakan dalam klasifikasi komoditas yaitu data survei lapangan, administrasi wilayah, Sentinel 2A, Sentinel 1A, GLCM (Gray Level Co-occurrence Matrix), SRTM (Shuttle Radar Topography Mission), terrain, dan CHIRPS (Climate Hazards Center Infrared Precipitation). Pengolahan data citra dilakukan dalam satu proses stacking GEE. Data dibagi menjadi data latih dan data uji. Pencarian parameter model terbaik dilakukan dengan hyperparameter tuning. Pemodelan komoditas dilakukan dengan algoritma RF, SVM, XGBoost, dan CNN. Hasil pemodelan dievaluasi dengan akurasi, classification report dan confussion matrix. Proses estimasi HST dilakukan setelah pengklasifikasian komoditas dan deteksi masa tanam. Proses deteksi masa tanam dan estimasi HST dilakukan dengan identifikasi phenology. Identifikasi phenology dilakukan dengan membandingkan penggunaan Sentinel 1A, Sentinel 2A, dan gabungan Sentinel 1A dan Sentinel 2A. Akurasi estimasi HST dievaluasi dengan R2, RMSE, dan MAE. Hasil penelitian menunjukan proses klasifikasi komoditas terbaik diperoleh dari algoritma XGBoost, yang memiliki akurasi tertinggi sebesar 91% dengan recall 93% untuk tanaman padi dan 91% untuk tanaman tebu. Sementara estimasi HST terbaik diperoleh dengan menggunakan citra Sentinel 1A dengan akurasi 90% dengan rentang perbedaan hari berkisar 3 hingga 18 hari.

Food self sufficiency is the ability of a region to independently meet its food needs. The agricultural sector serves as the primary pillar of food self-sufficiency. Nevertheless, the sector has the potential to experience a decline in contribution due to increasing population growth, land degradation, and ongoing climate change. Adaptive capacity becomes crucial in addressing the evolving challenges associated with the declining contribution of agriculture. A declining trend in agricultural contribution was observed in East Java Province from 2019 to 2024. East Java Province is the largest agricultural region in Indonesia, particularly for rice and sugarcane commodities. However, the monitoring process for commodities and Days After Planting (HST) in East Java has not yet utilized machine learning and deep learning approaches. The monitoring process is still conducted through surveys by Agricultural Extension Workers (PPL). Adaptation of the monitoring process can be achieved through the implementation of smart farming based on precision agriculture by leveraging remote sensing and Artificial Intelligence (AI) to produce accurate and timely information. This study aims to develop a commodity classification model by implementing Random Forest (RF), Support Vector Machine (SVM), Extreme Gradient Boosting (XGBoost), and Convolutional Neural Network (CNN) algorithms. The classification results are then used to estimate HST through phenological analysis of vegetation indices. The study was conducted through four main stages: data collection and selection, data exploration and preprocessing, modeling, and evaluation. The data used for commodity classification included field survey data, administrative area data, Sentinel 2A, Sentinel 1A, GLCM (Gray Level Co-occurrence Matrix), SRTM (Shuttle Radar Topography Mission), terrain, and CHIRPS (Climate Hazards Center Infrared Precipitation). Image data processing was performed through a single stacking process in GEE. The dataset was divided into training and testing data. The search for the best model parameters was carried out using hyperparameter tuning. Commodity modeling was performed using RF, SVM, XGBoost, and CNN algorithms. The modeling results were evaluated by accuracy, classification report, and confusion matrix. The HST estimation process was carried out after commodity classification and planting season detection. The planting season detection and HST estimation were conducted through phenology identification. Phenology identification was performed by comparing Sentinel 1A, Sentinel 2A, and a combination of Sentinel 1A and Sentinel 2A. The HST estimation was evaluated using R², RMSE, and MAE. The results showed that the best commodity classification performance was achieved by the XGBoost algorithm, which attained the highest accuracy of 91%, with recall values of 93% for rice and 91% for sugarcane. Meanwhile, the best HST estimation was obtained using Sentinel 1A imagery, with an accuracy of 90% and a day-difference range of 3 to 18 days.