Farm area segmentation in satellite images using deeplabv3+ neural networks

Abstract

<p>Farm detection using low resolution satellite images is an important part of digital agriculture applications such as crop yield monitoring. However, it has not received enough attention compared to high-resolution images. Although high resolution images are more efficient for detection of land cover components, the analysis of low-resolution images are yet important due to the low-resolution repositories of the past satellite images used for timeseries analysis, free availability and economic concerns. In this paper, semantic segmentation of farm areas is addressed using low resolution satellite images. The segmentation is performed in two stages; First, local patches or Regions of Interest (ROI) that include farm areas are detected. Next, deep semantic segmentation strategies are employed to detect the farm pixels. For patch classification, two previously developed local patch classification strategies are employed; a two-step semi-supervised methodology using hand-crafted features and Support Vector Machine (SVM) modelling and transfer learning using the pretrained Convolutional Neural Networks (CNNs). For the latter, the high-level features learnt from the massive filter banks of deep Visual Geometry Group Network (VGG-16) are utilized. After classifying the image patches that contain farm areas, the DeepLabv3+ model is used for semantic segmentation of farm pixels. Four different pretrained networks, resnet18, resnet50, resnet101 and mobilenetv2, are used to transfer their learnt features for the new farm segmentation problem. The first step results show the superiority of the transfer learning compared to hand-crafted features for classification of patches. The second step results show that the model trained based on resnet50 achieved the highest semantic segmentation accuracy.</p>