Performance analysis of convolutional neural network-based image classifiers for agricultural farmland mapping

Abstract

Precise farmland mapping is essential for crop monitoring, land-use inventory, and subsidy audits. Particularly, the office of the municipality’s assessor or the registry of deeds needs adequate image classifiers to categorize farmland on a large scale across different seasons and regions. The study evaluated three convolutional neural network (CNN) models, such as efficient network baseline 3 (EfficientNetB3), residual network 50 layers version 2 (ResNet50V2), and inception version 3 (InceptionV3), to identify the model that best balances accuracy and efficiency for classifying high-resolution farmland imagery. EfficientNetB3 proved to be the most suitable model, demonstrating higher accuracy, robustness, and a stronger ability to identify rice farmland patterns. While ResNet50V2 and InceptionV3 are widely used for image analysis, their performance declined when processing farmland images that exhibit substantial differences in visual characteristics. EfficientNetB3 delivered more consistent results for farmland image classification in large-scale agricultural applications. These findings highlight its potential for supporting data-driven agricultural management and policy implementation.