Land use/land cover recognition in arid zone using A multi-dimensional multi-grained residual Forest^☆

Abstract

Monitoring arid areas could effectively improve economic, ecological and humanity benefits. It is an effective monitoring approach to recognize the land cover or land use of arid areas through machine learning methods using satellite images. However, there is no public classified dataset for arid areas currently, and hence remote sensing image monitoring in desert areas is restricted. Existing classification methods are not able to fully utilize effective features of satellite images and multi-spectral optical parameters. In this paper, our contributions are as follows: Firstly, we presented a new satellite dataset named the ARID-5 for arid area land cover/land use (LULC) classification, the LULC in arid areas included desert, oasis, Gobi, and water system. Second, we proposed a machine learning algorithm named the multi-dimensional multi-grained residual forest algorithm for LULC recognition on arid areas. In this algorithm, the multi-dimensional multi-grained structure was able to effectively extract image features and spectral information. The residual forest structure mapped probability feature vectors to higher levels for prediction, which effectively improved the reflection of the forest structure on the sample. At the same time, the base estimator was transmitted in cascade layers, and thus the diversity and the accuracy were improved. Experimental results proved that the multi-dimensional multi-grained residual forest showed good classification abilities. Last, we also tested our algorithm on SAT-4 and SAT-6 datasets, which proved the generalization performance of our algorithm.

Introduction

Land desertification is one of the biggest environmental problems that the world is facing today. Due to the rapid expansion of desert areas, environmental degradation and huge economic losses were observed and political stability and social security problems thus arose in these areas. The desert oasis eco-interlacing zone was a narrow zone between the desert and the oasis, acting as the boundary and corridor between desert and oasis ecosystems, and the zone controlled the energy and nutrient flows (Wang et al., 2007). Because of dry climate and improper human intervention, the desert-oasis eco-interlacing zone became the most sensitive area to human activities such as land and water resource developments (Qiao and Wang, 2019), which might cause severe depletion of natural vegetation (Li et al., 2016). The stability of oasis ecosystems was also affected by desertification of oasis-desert ecotones and salinization within oases. These changes led to drought and desertification, hindering sustainable agriculture development. Therefore there was an urgent need to understand the interactions between the oasis and the desert (Liu et al., 2007). Desertification was already a global concern (Henry, 2002).

In recent years, various machine learning methods were used in identification and classification of remote sensing land cover and achieved good results (Sun et al., 2019; Martiniano et al., 2018; Buddhiraju and Rizvi, 2010; Shih et al., 2019; Zhai et al., 2019a, Zhai et al., 2019b). A Gaussian edge monitoring method was proposed for remote sensing image analysis, it was able to extract edge information effectively (Basu, 2002), but its calculation speed, convergence ability and multi-scale correlation remained unsatisfactory. The Maximum Likelihood Classification (MLC) method was used to plot and monitor remote sensing and geographic information of land use/land cover (LULC) changes in coastal areas of northwestern Egypt (Shalaby and Tateishi, 2007), the MLC was a parametric classification method, if the distribution estimate were accurate, good results would be achieved. However, the MLC was very sensitive to parameter estimates. If the estimation were incorrect, there would be significant negative impacts on the results. The Support Vector Machine (SVM) was also used to measure the land cover (Kavzoglu and Colkesen, 2009; Liu and Chen, 2019; Petropoulos et al., 2012), the land cover texture features were extracted by fractal dimensions, the gray level co-occurrence matrices and the wavelet transforms. Consequently, the land cover was identified using an SVM with a Radial Basis Function (RBF) as its core function. The gray level co-occurrence matrix was used to extract texture features of hyperspectral images (Xin et al., 2014), its speed was fast, but its results on some large locals were not very satisfactory. The Random Forest (RF) (Liaw and Wiener, 2002) was widely used in remote sensing images. In Halmy's work (Halmy et al., 2015), the RF was able to process high-dimensional data (Zhang et al., 2019), and its generalization ability was strong, but when the data noise became large, the over-fitting phenomenon happened easily. When there were data with different values, those with significant values would have greater impacts on the random forests. With the development of convolutional neural networks (CNN), the CNN was widely used in LULC classification (Helber et al., 2019; Rakshit et al., 2018; Zhai et al., 2019a, Zhai et al., 2019b). In Keshk's work (Keshk and Yin, 2020), the CNN was used in desert classification, but it was obvious that the neural network was more time-consuming in comparison to machine learning algorithms because the neural network required a large amount of backpropagation, and hence required a large amount of computing resources in real time detection of remote sensing images (Xia et al., 2019; Roy et al., 2011; Jiao et al., 2019).

In fact, there is little research aimed at LULC classification in arid regions, because there is no public classified dataset for arid areas currently. There are many researches aiming at remote sensing image spatial features and spectral features (Rasti et al., 2019; Lei et al., 2019; Imani and Ghassemian, 2020). From those researches, we could see that one key to improve the classification accuracy is to effectively extract spatial and spectral features. In spatial features of arid areas, the boundaries of desert, gobi and oasis were ambiguous, and different terrains overlapped with each other in space. Therefore the accuracy of arid area classification using traditional remote sensing image classification methods was relatively low (Megahed et al., 2015). In addition, in terms of spectral characteristics, differences in field reflectance spectra of different types of sand and Gobi in visible and near-infrared bands (400–1100 nm range) are not very obvious (Wang and Gao, 1984), as a result, many algorithms performed poorly on land cover/land use classification in desert areas. Therefore, we should pay more attention to spectral information and spatial information on land cover/land use classification in arid areas. The LULC recognition based on remote sensing images in arid/semi-arid areas was an important approach to achieve large-scale investigation of desert areas (Myint and Okin, 2009; Medjani et al., 2017). In Alqurashi's work (Alqurashi and Kumar, 2014), the RF was used to detect land use and land cover changes in Saudi Arabia desert cities, but its accuracy is not exciting, mainly because there is no effective feature extraction method. In order to solve these problems, we present a new satellite dataset named the ARID-5 aimed for arid area LULC classification and propose an algorithm which can effectively extract both spectral features and spatial features.

In this paper, a multi-dimensional multi-grained residual forest (mgrForest) algorithm was proposed for land cover/land use classification on remote sensing images of arid areas. Remote sensing data of Northwest China was chosen as the research target data. Our proposed method was compared with various classification methods such as SVM, RF and CNNs. Our method, the mgrForest, was proven to be able to fully extract both spatial features and spectral information from multispectral data, and therefore it presented higher accuracy in desert, Gobi and oasis recognition of multi-spectral data. Plus, the number of parameters requiring optimization in our method was much fewer than that of CNNs. The advantages of our proposed method could be summarized as: fewer parameters to set, easier to train, and faster to classify.

In addition, in order to prove the generalization performance of our algorithm, we compared our algorithm with existing algorithms on STA-4 and STA-6 datasets, and we achieved state-of-the-art results in terms of accuracy.

To summarize, our contributions are as follows: (1) We presented a new satellite dataset named the ARID-5 for arid area LULC classification; (2) A novel machine learning algorithm named the mgrForest was proposed for land use/land cover recognition in arid areas; (3) Achieved state-of-the-art results in SAT-4 and SAT-6 datasets.