Analisis Perbandingan Kinerja Backbone ResNet50v2, EfficientNetB4, dan DenseNet201 dalam Mask R CNN untuk Segmentasi Presisi Penyakit Rust pada Daun Apel

Abstract

Excessive pesticide use in apple farming due to imprecise spraying causes chemical residue problems and environmental contamination. This study aims to evaluate the performance of three CNN backbones (ResNet50V2, EfficientNetB4, and DenseNet201) in Mask R-CNN architecture for precision segmentation of rust disease on apple leaves, supporting smart spraying systems. The Plant Pathology Challenge 2020 dataset was used with 101 training images and 41 validation images manually annotated. Models were trained for 20 epochs using SGD optimizer and evaluated using Mean Average Precision (mAP) metrics at IoU 0.50 and 0.75, as well as inference time. Results show that DenseNet201 achieved the best performance for precision segmentation with mAP mask @0.75 of 0.402 and mask mAP@[0.50:0.95] of 0.353, representing a 9.2% improvement over the ResNet50 baseline. EfficientNetB4 excelled in efficiency with the fastest inference time (101.4 ms) and highest mAP @0.50 (0.585), while ResNet50V2 recorded the lowest performance despite stable training