Mixed attention mechanism on ResNet-DeepLabV3+ for paddy field segmentation

Abstract

Rice cultivation monitoring is crucial for Indonesia, where paddy field areas de clined by 2.45% according to the Central Bureau of Statistics due to land func tion changes and shifting crop preferences. Regular monitoring of paddy field distribution is essential for understanding agricultural land utilization by farmers and landowners. Satellite imagery has become increasingly common for agricul tural land observation, but traditional neural networks alone provide insufficient segmentation accuracy. This study proposes an enhanced deep learning architec ture combining residual network (ResNet)-DeepLabV3+ with coordinate atten tion (CA) and spatial group-wise enhancement (SGE) modules. The attention mechanisms establish direct connections between context vectors and inputs, enabling the model to prioritize relevant spatial and spectral features for precise paddy field identification. The CA module enhances spectral feature discrim ination, whereas the SGE improves spatial characteristic representation. The experimental results demonstrate superior performance over the baseline meth ods, achieving intersection over union (IoU) of 0.85, dice coefficient of 0.89, and accuracy of 0.95. The proposed mixed attention mechanism significantly improves the accuracy and efficiency of automatic crop area identification from satellite imagery.