ML-LUM: A system for land use mapping by machine learning algorithms

Abstract

The land use mapping refers to mapping and assessing changes and patterns of land use. The use of agricultural land maps becomes increasingly important. The governments, private sectors, research agencies, and community groups rely on land use mapping data for natural resource assessment, monitoring, and planning. Finding an effective mapping approach is thereby crucial for natural resource condition monitoring and investment, agricultural productivity, sustainability and planning, biodiversity conservation, natural disaster management, and bio-security.

In this paper, four machine learning algorithms, i.e., the classic k-Nearest Neighbour (kNN), Support Vector Machines (SVMs), Convolutional Neural Network (CNN), and newly developed Capsule Network (CapsNet), are applied to classify satellite images for land use. For comprehensively comparing the performance of different algorithms for land use mapping, the experiments have been conducted on real-world datasets. Based on the experiment results, several improvements on the algorithms are proposed in order to fulfil the requirement of a large-scale land mapping. In addition, we design and implement these algorithms for land use mapping in a Machine Learning Land Use Mapping (ML-LUM) system. The system is able to train the models, predict classifications of satellite images, map the land use, display the land use statistic data, and predict production yields. With a friendly graphic user interface for farmers, the system is implemented by using the cloud computing technique for processing large land use data. Furthermore, we present a case study. For the case study, a banana plantation area from a given satellite image is correctly marked and the area size is then calculated, together with predicting banana production.

Introduction

Land use mapping becomes increasingly important. For example, the remote sensing centre in Queensland relies on satellite remote sensing data to map land use and land use changes, under the Australian Land Use Mapping Program (ALUMP) [1]. The program aims to build nationally consistent land use datasets so as to meet a wide range of user needs and to make the best use of existing data and available resources [1]. In particular, ALUMP provides nationally consistent spatial information in order to assist in planning, natural resource condition monitoring and investment, agricultural productivity and sustainability, biodiversity conservation, as well as natural disaster management and bio-security. As such, the Australian government, private sectors, research agencies and community groups can use the ALUMP datasets for natural resource assessment, monitoring, and planning [1].

The majority current methodologies of mapping land use still require skilled teams to manually interpret satellite imagery data and manually classify land use features [2]. It is obvious that this process is time and resource intensive. Finding an effective mapping approach will significantly reduce such costs.

With the rapid advances in technologies, automatic land use mapping becomes possible. Advances in machine learning (ML) enable to train a machine learning model on ALUMP's datasets for automatically mapping land use features [2]. In other words, ML algorithms make automatic land use mapping possible, by training models on all of the available data. As such, machine learning algorithms can be used to identify the land use from satellite images, to reduce the intensive manual effort required and to produce results in a shorter period of time [3]. Furthermore, the use of machine learning for land use mapping can make predictions. This can guide better decisions and smart actions for land use in real time. In addition, it can help seasonal crop mapping and monitoring [3] and control, preventing Banana Bunchy Top Virus (BBTV) outbreak [4].

This study will investigate the use of machine learning to identify land use from satellite imagery so as to reduce intensive manual efforts required and obtain results in shorter timeframes. The question is how to apply appropriate machine learning algorithms to automatically identify and classify land use.

In order to answer this question, we start by identifying the requirements of land use mapping as follows:

• Accuracy: a high accuracy is expected for the land use mapping; otherwise manual corrections by experienced staff are still needed.

• Processing speed: the speed of data processing is essential. This is particularly true when the outcome results are in an urgent need, such as disaster management.

These requirements for land use mapping pose challenges to automatic mapping systems. Further, the computing capacity is essential for the huge size of the data. For example, Queensland Australia has approximately 173 million hectares in area. A baseline dataset of land use mapping in Queensland contains approximately 160,000 features [3]. An example is shown in Fig. 1. It is a challenge for a program and algorithm to process such a large dataset.

In this paper, we have made the following contributions:

• we conduct extensive experiments on comparing several important ML algorithms for land use mapping.

In particular, the deep learning algorithms of CNN and the newly developed CapsNet are used. To the best of our knowledge, this is the first work on applying deep learning algorithms, particular the newly developed CapsNet algorithm to land use mapping. Some improvements on the algorithms for land mapping use have been proposed, particularly for the settings of parameters in the algorithms.

• We make recommendations on how to choose algorithms for land use mapping under different cases are made. The recommendation is based on comparing and analysing the experimental results on both public datasets and satellite imagery datasets.

• We design and develop a system that implemented these machine learning algorithms in land use mapping (ML-LUM). By giving a set of satellite images, the ML-LUM system uses different machine learning algorithms to produce classification outputs. The system is implemented in a cloud computing environment that uses supercomputer resources for processing huge data quickly.

The organization of this paper is as follows. Section 2 reviews related work. Section 3 presents the compared approaches, followed by Section 4 that presents the cloud implementation of the system. Section 5 describes the experimental methodologies. Section 6 provides an illustrative case study. The paper concludes in Section 7.