Abstract
Crop area statistics and yield prediction will affect adjustment of agricultural policy, to a certain extent. With the development of computer automatic classification techniques, the performance of classifiers are influenced by feature preprocessing and sample selection. Remote sensing classification according to spectral information is affected by false negatives and miscalculation in the complex spectrum area. Corn planting areas and other land-cover objects contain different surface structures and smoothness; other vegetation and villages have coarse textures. This paper introduces texture information based on a Gabor filter group to enrich land-cover information and establish a spectrum-texture feature set. With more samples, the algorithm efficiency is greatly affected. This paper proposes an improved fuzzy ARTMAP (FAM) with an adaptive boost strategy, namely Adaboost_FAM. Weak classifiers are trained to construct strong classifiers so as to improve operation efficiency. Meanwhile, classification accuracy will not be greatly improved. Experimental results indicate that the proposed method improves extraction accuracy when compared to classical algorithms, and improves efficiency when compared to algorithms which contain a great number of samples.
Introduction
Agricultural remote sensing information extraction is one of the key technologies in remote sensing applications; many researchers have improved and explored various methods. Automatic classification accuracy and efficiency have been subjects of constant improvement. Anne Puissant proposes that texture information can greatly improve spatial resolution image classification [1]. However, single feature representation methods are unable to describe all relevant information, thus a multiple feature fusion method was developed. M. Fauvel used morphological properties to describe spatial information in order to combine spectrum characteristics [14]. This study overcomes the disadvantages of single feature classification and greatly improves the classification precision of the pixel scale. M. Fauvel and Y. Tarabalka combine morphological properties and multiple classifier methods to improve classification results, also demonstrating that texture and spectral information are important to image classification [15]. The texture of a corn field is different from that of villages and other vegetables, this feature has good consistency and regularity, therefore, this paper introduces texture information of corn areas to construct a multiple feature set and improve the accuracy of information extraction.
In general, remote sensing image classification methods can be divided into supervised classification and unsupervised classification methods. Unsupervised classification algorithms are those which do not account for prior knowledge, such as k-means and ISODATA. The disadvantage of these methods is that classification accuracy is low, resulting in rough classification results. Thus, these methods are not suitable for high resolution remote sensing images. Supervised classifiers include maximum likelihood [2], artificial neural networks, decision trees, support vector machines [13, 16, 17, 13, 16, 17] and random forests. The key point is the selection of the test sample, because its quality is directly related to classification accuracy. Combination classification techniques include semi-supervised methods, and fusions of supervised and unsupervised methods.
Many new algorithms, such as the artificial neural network (ANN) [4, 20], support vector machine (SVM) [18, 24] and random forest [5, 7], meet the requirements of complex multispectral data. One of the most widely used methods for supervised classification in remote sensing analysis is the use of artificial neural networks (ANN), which are formed by algorithms inspired by biological neural systems. In an initial training stage, these networks fix coefficients between the input data and output categories. The process is completed by subsequent verification and test stages so that classification is determined to be correct or incorrect according to measures that were not involved in the test circumstances.
A supervised adaptive resonance theory (ART)-based neural network, namely fuzzy ARTMAP [10], is proposed as the base classifier. An incremental learning model is able to overcome the stability-plasticity dilemma of the data [8, 9]. The FAM network is plastic enough to absorb new information from new samples, and stable enough to retain previously learned information which may be corrupted by newly learned information. An interesting feature of FAM is that it integrates fuzzy set theory [12] and the stability-plasticity characteristic of ART into a common framework. Taherian and Arash proposed an efficient iris recognition system that employs a circular Hough transform technique to localize the iris region in the eye image and introduce a cumulative sum-based gray change analysis method to extract features from the normalized iris template. Then, fuzzy ARTMAP neural network was used to classify the iris codes [3]. Tan Shing Chiang introduced two models of evolutionary fuzzy ARTM-AP (FAM) neural networks to deal with imbalanced datasets in a semiconductor manufacturing operation [23]. However, these methods selected solitary samples, and thus did not account for diversity within samples, thus affecting the classification performance. This paper adopts ensemble learning, an adaptive boosting strategy, to develop fuzzy ARTMAP.
Related research
Fuzzy ARTMAP neural network
The Fuzzy ARTMAP network is composed of two types of fuzzy ART networks: ARTa and ARTb. One used training data and the other utilized verification data. The relationship between both fuzzy ART networks was determined by a memory map called map-field. The input data were normalized to 1 and duplicated by adding their complements. Thus, a data vector was obtained, which allowed the network weights and the maximum and minimum input values to be determined [11].
Adaptive boost strategy
Adaptive boost strategy is an active learning method. Theoretical research proves that strong classifier error rates will be zero as long as each weak classifier classification error rate is below 50% and the number of weak classifier approaches ∞. In corn planting information extraction, in order to simplify the optimization of process parameters, it is necessary to construct weak classifiers based on few samples and brief characteristics. This mentality is feasible in theory. The algorithm is described as follows:
Experiment datasets
Experiment data is derived from the fusion results based on a multi-spectrum image (B - G - R - NIR) and panchromatic image (pan); resolution is equal to 2 meters. To maintain consistent experimental conditions, this text selects three images from the same scene data in primary corn planting territory. Image sizes are 600 × 600, 1024 × 1024 and 400 × 400 pixels, respectively. Remote sensing images include corn, non-corn plants, and urban construction. Here, the corn crop planting area information extraction is studied; the corn area is identified as the area of interest, other areas are considered to be background. All experiments are conducted with a Windows 7 operation system with an Intel Core i5-2.30 GHz processor and 4.0 GB RAM.
Feature extraction and selection
Spectrum features
According to experience, NIR, R and G are selected as features. In addition, the common useful vegetation index (VI) contains the normalized differential vegetation index (NDVI), soil-adjusted vegetation index (SAVI) and enhanced vegetation index (EVI). The test selects three characteristics to obtain featuresets. NDVI is derived from red and NIR bands, and can effectively distinguish plants from otherobjects. SAVI can partially reduce the effect of a soil background. The modified soil-adjusted vegetation index (MSAVI) takes into account changes in soil factors without the soil index, and is suitable for areas of sparse vegetation coverage.
The effect of soil and atmosphere on NDVI are not independent; therefore, EVI simultaneously introduces feedback from two amendment regulations, using the soil adjustment index and atmospheric correction parameters.
where L is the soil-adjusted index; values range from 0.5 to 1, and is equal to 0.5 in this paper.
where L is soil-adjusted index; values range from 0.5 to 1, and is equal to 0.5 in this paper. C1 and C2 are atmospheric correction parameters.
These features are extracted by multi-scale and multi-orientation Gabor filters to simulate human vision in order to extract terse and concise contour information [25]. The feature extraction procedure is as follows:
Many studies show that the dimension of the feature vector greatly affects the algorithm efficiency. Although the vector dimension is not very large in this context, reducing the dimension of the feature set can improve the efficiency of the algorithm, which has practical
significance. This paper uses cross-validation to determine feature importance [19, 21], as shown in Fig. 1. According to experimental results, R, NIR, NDVI, homogeneity and contrast degree are selected to build a feature set. Although the degree of dimension reduction is small, it still affects the operational process. When the characteristic dimension is very large, this step has significant practical significance.
As shown in Fig. 1, the numbers of feature axle represent features B, G, R, NIR, NDVI, SAVI, EVI, and the Gabor texture of three scales, respectively. The spectral features are more important than textural features; this paper select G, R, NDVI, SAVI, EVI and the Gabor texture features of three scale degrees to execute subsequent supervised classification.
Experiment results
This section implements experiments for three images, and compare results with those obtained by traditional classification methods. For the first experimental dataset, the number of samples and weak classifiers will affect the final classification performance, which is evaluated by the overall accuracy (OA) and Kappa coefficient. OA verifies the number of pixels that are classified correctly. Kappa can be used to assess the agreement of the two classifications for each class. This paper selects different numbers of samples. The results shown in Fig. 2 indicate that the classification performance can be improved with an increasing number of samples. Then, this paper selects different numbers of weak classifiers to construct Adaboost_FAM.
Figure 2 presents the Kappa sample curve, and the overall accuracy sample curve. Figure 3 presents the Kappa iteration curves, and the overall accuracy iteration curves.
Table 1 depicts the confusion matrix of classification and other indicators. Except for the kappa index and OA, the Producer Accuracy (PA) and User Accuracy (UA) metrics were computed using the following equations.
This paper utilizes fuzzy theory, neural networks and adaptive boost to construct a compound classification frame. The results of the proposed algorithm are markedly superior to results obtained by others. For the first remote sensing image with Adaboost_FAM (spectrum + texture), KAPPA, UA and OA are 0.5972, 0.7117 and 0.8532, respectively. SVM classification performance is similar to the proposed method, but the computational cost is too large due to parameter optimization.
To test the algorithm for image data which has a smaller or larger corn proportion, this paper selects two images as the analysis objects. For the second dataset, the performance indices are listed in Table 2. The KAPPA index, UA and OA are 0.698, 0.8461 and 0.8513, respectively; these values are superior to other traditional classifiers. The experiment demonstrates the effectiveness of the proposed method. The results of every method are shown in Fig. 5.
Classification precision parameters of the third dataset are listed in Table 3. In view of the terrain features of this experimental area in which the corn planting area is relatively small, after many experiments, the contribution of texture feature to the classification results is small. In some cases, it cannot even reach normal results. For the proposed method, KAPPA, UA and OA are 0.2489, 0.424 and 0.954, respectively. Due to the objective condition of the third dataset, the results are not superior to others. The non-corn area accounts for a large proportion of area; this masks the shortcoming of insufficient information to a certain extent.
In the field of remote sensing classification, because of diversity and uncertainty of images data, same objects contain very different spectrum attributes, on the contrary, different objects may contain same spectrum attributes. According to the corn regional distribution characteristics, this paper introduces Gabor filter features to obtain a joint feature set. By applying fuzzy theory, neural network strategies, and adaptive boost strategy, a compound classification framework is implemented to extract the crop area. According to the experimental data, feature subsets and training samples were obtained. The work of this paper uses the adaptive boost strategy to optimize FAM classifiers, constructs weak classifier group which is based on less samples and simple feature set, gets Adaboost_FAM classifier, realizes information extraction process in shaanxi province. According to experimental results and analysis, the proposed method performs relatively well for the three studied datasets. Experiments indicate that classification precision is better than that obtained by typical supervised classification methods, as the parameter optimization and the training process takes less time than the traditional classification methods.