Identification of Land Cover and Crop Type Using Knn Classifier in Sar Image

Introduction

Land cover refers to the surface cover on the ground, whether vegetation, urban infrastructure, water, bare soil, or other. Identifying, delineating, and mapping land cover is crucial for global monitoring studies, resource management, and planning activities. Crop monitoring information is vital for food security and helps to enhance our understanding of agriculture's role in climate change and crop type identification. This work focuses on an automated KNN classification system for identifying land cover and crop type in Synthetic Aperture Radar (SAR) images. In the first module, an unsupervised Kohonen’s Self-Organizing Mapping (SOM) neural network is used for identifying the land type.

Methodology

In the second module, local binary pattern (LBP) based features are extracted for identifying the crop type in the crop-covered area. The extracted features are fed to the KNN classifier, which classifies the type of crop. Agriculture is the primary backbone of the Indian economy, with about 70% of the population dependent on it. In agriculture, parameters like canopy, yield, and product quality are crucial from the farmer's perspective (Viraj et al., 2012). India is a top producer of many crops. The major crops in India can be divided into four categories: Food grains, Cash Crops, Plantation Crops, and Horticulture crops.

Land Cover and Crop Monitoring

Learning multistage and deep representations for classifying remotely sensed imagery is essential (Zhao et al., 2016). Identifying land cover establishes the baseline for monitoring activities (change detection) and provides ground cover information for baseline thematic maps. Crop monitoring is critical for food security and enhances our understanding of agriculture's role in climate change and crop type identification (Ajay et al., 2012). Measuring crop types offers numerical descriptions of the crop, helping determine whether a problem is significant enough to address or small enough to ignore.

Advanced Techniques

The proposed 3-D CNN-based FE model, with combined regularization, effectively extracts spectral–spatial features of hyperspectral imagery, offering excellent classification performance under limited training samples. Designing proper deep CNN models is challenging. Nataliia Kussul et al. (2016) proposed a methodology for solving large-scale classification and area estimation problems in the remote sensing domain, based on the deep learning paradigm. It utilizes a hierarchical model that includes self-organizing maps (SOM) for data pre-processing and segmentation (clustering), an ensemble of multi-layer perceptions (MLP) for data classification, and heterogeneous data fusion and geospatial analysis for post-processing. An ensemble of methods (“mixture of experts” approach) should be exploited to take advantage of different processing methods and techniques.

Challenges and Solutions

Christopher McCool et al. (2016) proposed a novel crop detection system for field sweet pepper (capsicum) detection, addressing challenges such as high occlusion and similar color between the crop and background. To overcome these issues, a two-stage system performing per-pixel segmentation followed by region detection was proposed. This approach offers robustness against occlusion and minimizes laborious annotation. Adriana Romero et al. (2016) proposed Greedy layer-wise unsupervised pre-training coupled with an efficient algorithm for unsupervised learning of sparse features. The algorithm is rooted in Sparse Representation and enforces both population and lifetime scarcity of the extracted features simultaneously. Utilizing spatial information by combining many output features and max-pooling steps in deep architectures is crucial to achieving excellent results.

Conclusion

Traditional unsupervised classification algorithms, such as maximum likelihood classification, use clustering techniques to identify spectrally distinct data groups and are the earliest approach to automatic land cover classification employing pattern recognition techniques. However, these algorithms have drawbacks, such as the lack of guaranteed accuracy in land cover classification, and the classifications are arbitrary. Supervised classification methods require substantial expertise and human participation for selecting training samples, heavily influencing the classification results. Algorithms like neural network classification and fuzzy logic classification are complex, making them difficult to understand and apply widely. Decision tree classification methods are widely used for large areas, such as global land cover mapping, but constructing decision trees and assigning thresholds for sub-nodes depend heavily on human experience and vary spatially and temporally.

Proposed Work

The proposed architecture of land cover and crop type classification is illustrated in Figure 3.1. The input images captured from SAR satisfy the quality requirements necessary for land cover and crop type identification. In the first stage of the proposed work, the input image is segmented using the self-organization map (SOM) based technique. The self-organization map is a type of artificial neural network trained using unsupervised learning to produce a low-dimensional representation of the input space of the training samples. It differs from other artificial neural networks as it applies competitive learning rather than error-correction learning and uses a neighborhood function to preserve the input space's topological properties. The Self-Organizing Map forms a non-linear mapping of a high-dimensional input space into a typically two-dimensional grid of artificial neural networks. In picSOM, a separate SOM is trained for each feature type. Through this mapping, feature vectors near each other in the input space are mapped into nearby units on the map, allowing imagelets similar in respect to given features to be located near each other in the SOM.