Brian P. Salmon

Multiview Deep Learning for Land-Use Classification

A multiscale input strategy for multiview deep learning is proposed for supervised multispectral land-use classification, and it is validated on a well-known data set. The hypothesis that simultaneous multiscale views can improve composition-based inference of classes containing size-varying objects compared to single-scale multiview is investigated. The end-to-end learning system learns a hierarchical feature representation with the aid of convolutional layers to shift the burden of feature determination from hand-engineering to a deep convolutional neural network (DCNN). This allows the classifier to obtain problem-specific features that are optimal for minimizing the multinomial logistic regression objective, as opposed to user-defined features which trade optimality for generality. A heuristic approach to the optimization of the DCNN hyperparameters is used, based on empirical performance evidence. It is shown that a single DCNN can be trained simultaneously with multiscale views to improve prediction accuracy over multiple single-scale views. Competitive performance is achieved for the UC Merced data set, where the 93.48% accuracy of multiview deep learning outperforms the 85.37% accuracy of SIFT-based methods and the 90.26% accuracy of unsupervised feature learning.

Detecting Land Cover Change Using an Extended Kalman Filter on MODIS NDVI Time-Series Data

A method for detecting land cover change using NDVI time-series data derived from 500-m MODIS satellite data is proposed. The algorithm acts as a per-pixel change alarm and takes the NDVI time series of a 3 × 3 grid of MODIS pixels as the input. The NDVI time series for each of these pixels was modeled as a triply (mean, phase, and amplitude) modulated cosine function, and an extended Kalman filter was used to estimate the parameters of the modulated cosine function through time. A spatial comparison between the center pixel of the 3 × 3 grid and each of its neighboring pixels mean and amplitude parameter sequence was done to calculate a change metric which yields a change or no-change decision after thresholding. Although the development of new settlements is the most prevalent form of land cover change in South Africa, it is rarely mapped, and known examples amount to a limited number of changed MODIS pixels. Therefore, simulated change data were generated and used for the preliminary optimization of the change detection method. After optimization, the method was evaluated on examples of known land cover change in the study area, and experimental results indicate an 89% change detection accuracy while a traditional annual NDVI differencing method could only achieve a 63% change detection accuracy.

Unsupervised Land Cover Change Detection: Meaningful Sequential Time Series Analysis

An automated land cover change detection method is proposed that uses coarse spatial resolution hyper-temporal earth observation satellite time series data. The study compared three different unsupervised clustering approaches that operate on short term Fourier transform coefficients computed over subsequences of 8-day composite MODerate-resolution Imaging Spectroradiometer (MODIS) surface reflectance data that were extracted with a temporal sliding window. The method uses a feature extraction process that creates meaningful sequential time series that can be analyzed and processed for change detection. The method was evaluated on real and simulated land cover change examples and obtained a change detection accuracy exceeding 76% on real land cover conversion and more than 70% on simulated land cover conversion.

Improving Land Cover Class Separation Using an Extended Kalman Filter on MODIS NDVI Time-Series Data

It is proposed that the normalized difference vegetation index time series derived from Moderate Resolution Imaging Spectroradiometer satellite data can be modeled as a triply (mean, phase, and amplitude) modulated cosine function. Second, a nonlinear extended Kalman filter is developed to estimate the parameters of the modulated cosine function as a function of time. It is shown that the maximum separability of the parameters for natural vegetation and settlement land cover types is better than that of methods based on the fast Fourier transform using data from two study areas in South Africa.

The use of a Multilayer Perceptron for detecting new human settlements from a time series of MODIS images

This paper presents a novel land cover change detection method that employs a sliding window over hyper-temporal multi-spectral images acquired from the 7 bands of the MODerate-resolution Imaging Spectroradiometer (MODIS) land surface reflectance product. The method uses a Feedforward Multilayer Perceptron (MLP) for supervised change detection that operates on multi-spectral time series extracted with a sliding window from the dataset. The method was evaluated on both real and simulated land cover change examples. The simulated land cover change comprises of concatenated time series that are produced by blending actual time series of pixels from human settlements to those from adjacent areas covered by natural vegetation. The method employs an iteratively retrained MLP to capture all local patterns and to compensate for the time-varying climate change in the geographical area. The iteratively retrained MLP was compared to a classical batch mode trained MLP. Depending on the length of the temporal sliding window used, an overall change detection accuracy between 83% and 90% was achieved. It is shown that a sliding window of 6 months using all 7 bands of MODIS data is sufficient to detect land cover change reliably. Window sizes of 18 months and longer provide minor improvements to classification accuracy and change detection performance at the cost of longer time delays.

Very deep learning for ship discrimination in Synthetic Aperture Radar imagery

Efficient and effective ship discrimination across multiple Synthetic Aperture Radar sensors is becoming more important as access to SAR data becomes more widespread. A flexible means of separating ships from sea is ideal and can be accomplished using machine learning. Newer, advanced deep learning techniques offer a unique solution but traditionally require a large dataset to train effectively. Highway Networks allow for very deep networks that can be trained using the smaller datasets typical in SAR-based ship detection. A flexible network configuration is possible within Highway Networks due to an adaptive gating mechanism which prevents gradient decay across many layers. This paper presents a very deep High Network configuration as a ship discrimination stage for SAR ship detection. It also presents a three-class SAR dataset that allows for more meaningful analysis of ship discrimination performances. The proposed method was tested on a this SAR dataset and had the highest mean accuracy of all methods tested at 96.67%. The proposed ship discrimination method also provides improved false positive classification compared to the other methods tested.

Rapid land cover map updates using change detection and robust random forest classifiers

The CSIR and the Department Rural Development and Land Reform, National Geo-spatial Information (NGI).

Using Page's Cumulative Sum Test on MODIS Time Series to Detect Land-Cover Changes

Human settlement expansion is one of the most pervasive forms of land-cover change in South Africa. The use of Pages cumulative sum (CUSUM) test is proposed as a method to detect new settlement developments in areas that were previously covered by natural vegetation using 500-m Moderate Resolution Imaging Spectroradiometer time-series satellite data. The method is a sequential per-pixel change alarm algorithm that can take into account positive detection delay, probability of detection, and false-alarm probability to construct a threshold. Simulated change data were generated to determine a threshold during a preliminary offline optimization phase. After optimization, the method was evaluated on examples of known land-cover change in the Gauteng and Limpopo provinces of South Africa. The experimental results indicated that CUSUM performs better than band differencing in the before-mentioned study areas.

Land Cover Change Detection Using Autocorrelation Analysis on MODIS Time-Series Data: Detection of New Human Settlements in the Gauteng Province of South Africa

Human settlement expansion is one of the most pervasive forms of land cover change in the Gauteng province of South Africa. A method for detecting new settlement developments in areas that are typically covered by natural vegetation using 500 m MODIS time-series satellite data is proposed. The method is a per pixel change alarm that uses the temporal autocorrelation to infer a change index which yields a change or no-change decision after thresholding. Simulated change data was generated and used to determine a threshold during an off-line optimization phase. After optimization the method was evaluated on examples of known land cover change in the study area and experimental results indicate a 92% change detection accuracy with a 15% false alarm rate. The method shows good performance when compared to a traditional NDVI differencing method that achieved a 75% change detection accuracy with a 24% false alarm rate for the same study area.

Manifold Adaptation for Constant False Alarm Rate Ship Detection in South African Oceans

The detection of ships at sea is a difficult task made more so by uncooperative ships, especially when using transponder-based ship detection systems. Synthetic aperture radar (SAR) imagery provides a means of observation independent of the ships cooperation, and over the years, a vast amount of research has gone into the detection of ships using this imagery. One of the most common methods used for ship detection in SAR imagery is the cell-averaging constant false alarm rate (CA-CFAR) prescreening method. It uses a scalar threshold value to determine how bright a pixel needs to be in order to be classified as a ship, and thus inversely how many false alarms are permitted. This paper presents by a method of converting the scalar threshold into a threshold manifold. The manifold is adjusted using a simulated annealing (SA) algorithm to optimally fit to information provided by the ship distribution map, which is generated from transponder data. By carefully selecting the input solution and threshold boundaries, much of the computational inefficiencies usually associated with SA can be avoided. The proposed method was tested on six ASAR images against five other methods and had a reported detection accuracy (DA) of 85.2% with a corresponding FAR of 1.01 × 10-7.