Automatic semantic segmentation and classification of remote sensing data for agriculture

Abstract

Automatic semantic segmentation has expected increasing interest for researchers in recent years on multispectral remote sensing (RS) system. The agriculture supports 58 % of the population, in which 51 % of geographical area is under cultivation. Furthermore, the RS in agriculture can be used for identification, area estimation and monitoring, crop detection, soil mapping, crop yield modelling and production modelling etc. The RS images are high resolution images which can be used for agricultural and land cover classifications. Due to its high dimensional feature space, the conventional feature extraction techniques represent a progress of issues when handling huge size information e.g., computational cost, processing capacity and storage load. In order to overcome the existing drawback, we propose an automatic semantic segmentation without losing the significant data. In this paper, we use SOMs for segmentation purpose. Moreover, we proposed the particle swarm optimization technique (PSO) algorithm for finding cluster boundaries directly from the SOMs. On the other hand, we propose the deep residual network to achieve faster training process. Deep Residual Networks have been proved to be a very successful model on RS image classification. The main aim of this work is to achieve the overall accuracy greater than 85 % (OA > 85 %). So, we use a convolutional neural network (CNN), which outperforms better classification of certain crop types and yielding the target accuracies more than 85 % for all major crops. Furthermore, the proposed methods achieve good segmentation and classification accuracy than existing methods. The simulation results are further presented to show the performance of the proposed method applied to synthetic and real-world datasets.