Farm Detection based on Deep Convolutional Neural Nets and Semi-supervised Green Texture Detection using VIS-NIR Satellite Image

Abstract

Farm detection using low resolution satellite images is an important topic in digital agriculture. 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. The current paper addresses the problem of farm detection using low resolution satellite images. In digital agriculture, farm detection has significant role for key applications such as crop yield monitoring. Two main categories of object detection strategies are studied and compared in this paper; First, a two-step semi-supervised methodology is developed using traditional manual feature extraction and modelling techniques; the developed methodology uses the Normalized Difference Moisture Index (NDMI), Grey Level Co-occurrence Matrix (GLCM), 2-D Discrete Cosine Transform (DCT) and morphological features and Support Vector Machine (SVM) for classifier modelling. In the second strategy, high-level features learnt from the massive filter banks of deep Convolutional Neural Networks (CNNs) are utilised. Transfer learning strategies are employed for pretrained Visual Geometry Group Network (VGG-16) networks. Results show the superiority of the high-level features for classification of farm regions.

A New Method for Semi-Supervised Segmentation of Satellite Images

“…It finds application in many domains. Examples are urban data analysis and planning [1], digital agriculture [2,3], environmental studies and hazards analysis [4] and traffic and navigation [5].…”

“…However, they usually require a large volume of data. Even, hundreds of labelled images are required when transfer learning methods are employed [3,15].…”

Farm Area Segmentation in Satellite Images Using DeepLabv3+ Neural Networks

“…In this paper, the problem of farm detection and segmentation using low resolution satellite images is addressed. In our previous contribution, a farm detection strategy was developed at patch level [40]. The analysis include two different strategies; the first one was a semi-supervised strategy based on hand-crafted features combined by classification modeling similar to [40][41][42][43].…”

“…In our previous contribution, a farm detection strategy was developed at patch level [40]. The analysis include two different strategies; the first one was a semi-supervised strategy based on hand-crafted features combined by classification modeling similar to [40][41][42][43]. The developed algorithm consists of an unsupervised pixel-based segmentation of vegetation area using Normalized Difference Moisture Index (NDMI), followed by a supervised step for texture area classification and farm detection; GLCM and 2-D DCT features are used in an SVM framework for texture classification and in then, object-based morphological features were extracted from the textured areas for farm detection.…”

“…The main contribution of this paper is segmentation of farm areas semantically at pixel level. The analysis strategy consists of two main stages; first similar to our previous work [40], local image patches or ROIs that include farm areas are detected. Then, having found the local ROIs consisting the farm areas, in the next step, semantic segmentation of farm regions in the ROIs is performed using deepLabv3+ modelling strategy [39].…”