Abstract

Accurate crop classification is crucial for agricultural management, food security assessment, and environmental monitoring. In Pakistan, where agriculture is a key economic sector, precise mapping of crops like cotton is essential for effective resource management and policy-making. Remote sensing technologies provide cost-effective and efficient alternatives to traditional field surveys. This study leverages machine learning techniques, particularly ensemble methods to classify cotton using time-series Enhanced Vegetation Index (EVI) and Normalized Difference Vegetation Index (NDVI) data from MODIS satellite imagery. The year 2014 was selected for analysis due to the availability of high-quality FAO-SPOT reference data and representative climatic conditions during that growing season. The research compares the performance of various machine learning algorithms and evaluates the effectiveness of an ensemble approach in improving classification accuracy within Pakistan's agricultural landscape. Machine learning models, including Random Forest, Support Vector Machine (SVM), Gradient Boosting, XGBoost, and an Ensemble Voting classifier, were applied to EVI and NDVI time-series data. Performance was assessed using a 5-fold cross-validation approach to ensure robust evaluation. The models demonstrated strong predictive capabilities, with Gradient Boosting achieving the highest overall accuracy (OA) of 93.02%, followed closely by the Ensemble Voting model at 92.53%. Other models, including Random Forest, SVM, and XGBoost, also performed well, with OA of 92.46%, 92.17%, and 92.24%, respectively. The inherent limitations of MODIS 250 m spatial resolution, including mixed-pixel effects in fragmented agricultural landscapes, are acknowledged. These results highlight the potential of machine learning for large-scale cotton classification and agricultural monitoring in Pakistan.