Segmentation of farmlands in aerial images by deep learning framework with feature fusion and context aggregation modules

Abstract

Automated segmentation of farmland patterns in high resolution aerial images is very crucial for smart farming. Recently, deep learning techniques have achieved tremendous success in various semantic segmentation tasks, however, little efforts have been made in farmland semantic segmentation in high resolution aerial images. Farmland semantic segmentation in aerial images is a challenging task due to large variation in scales and shapes of agriculture patterns. Furthermore, different agriculture patterns share similar visual features that usually result in mis-classifications of pixels. To efficiently tackle these problems, we propose a deep learning framework that captures scene context and aggregate multi-scale information from different convolutional blocks. Generally, the framework consists of two main modules:(1) feature fusion module and (2) global contextual module. Feature fusion module combines the feature maps of different convolutional blocks to capture wide variation in object scales, while global contextual module aggregate rich contextual information from different regions of the image by employing pyramid pooling module. We gauge the performance of proposed framework on challenging benchmarks dataset, Agriculture-vision and also compare our results with various state-of-the-art methods. From experiment results, we demonstrate that the proposed framework achieves best performance in identifying various complex agriculture patterns and supersedes state-of-the-art methods.