The N-FINDR algorithm [18] is a fully automatic technique for endmember extraction from the image, which is briefly described here:

(i) Let N denote the number of classes or endmembers to be identified.

(ii) Perform a PCA-decomposition of the data and reduce the data to N−1 dimension space.

(iii) Pick N pixels from the set and compute the simplex volume generated by the spectra of the N pixels. The volume of the simplex is proportional to

(iv) Replace each endmember with the spectrum of each pixel in the data set and recompute the simplex volume. If the volume increases, the spectrum of the new pixel is retained as a potential endmember.

(v) The above steps are executed iteratively considering all pixels, and the final set of retained spectra is taken as the endmembers.

OSP proposed by Chang [19] involves (a) finding an operator which eliminates undesired spectral signatures, and then (b) choosing a vector operator which maximizes the signal to noise ratio (SNR) of the residual spectral signature.

If we assume that there are N targets, t₁, …t_n-1, t_n, t_n+1, …, t_N present in an image scene, then there are N spectrally distinct endmembers with corresponding target signatures as e₁, …e_n-1, e_n, e_n+1, …, e_N, where M > N (over a determined system), where Equation (1) is a standard signal detection model. Since we are interested in detecting one target at a time, we can divide the set of N targets into a desired target, say t_n, and a class of undesired targets, t₁,…,t_n-1, t_n+1,…, t_N. We need to eliminate the effects caused by the undesired targets that are considered as interferers to t_n before the detection of t_n takes place. With annihilation of the undesired target signatures, the detectability of t_n can be enhanced. In order to find the abundance of the nth target material (α_n), first e_n is separated from e₁, …e_n-1, e_n, e_n+1, …, e_N in E. Let the corresponding spectral signature of the desired target material be denoted as d. The term Eα can be rewritten to separate the desired spectral signature d from the rest as:

and Equation (1) is rewritten as

where d = e_n is the desired target signature of t_n and U is M x (N-1) matrix = [e₁, …e_n-1, e_n+1, …, e_N], is the undesired target spectral signature, which are the spectral signatures of the remaining N−1 undesired targets, t₁,…,t_n−1, t_n+1,…, t_N. Equation (7) is called a (d, U) model; d is a M × 1 column vector [d₁, d₂, …, d_M]^T, γ is a (N−1) × 1 column vector containing (N−1) component fractions of α = [α₁, …α_n−1, α_n+1,…, α_N]^T. Using the (d, U) model, OSP can annihilate U from the pixel vector y prior to detection of t_n similar to [20] by the operator

where U^# = (U^TU)⁻¹U^T is the pseudo-inverse of U. The projector P is a M × M matrix operator that maps the observed pixel vector y into the orthogonal complement of U. U has same structure as the orthogonal complement projector from the theory of least squares. Applying P to the (d, U) model results in a new signal detection model (OSP model) given by

where the undesired signal in U has been annihilated and the original noise has also been suppressed to Pη. The operator minimizes energy associated with the signatures not of interest as opposed to minimizing the total least squares error. It should be noted that P operating on Uγ reduces the contribution of U to about zero. So,

on using a linear filter specified by a weight vector x^T on the OSP model, the filter output is given by

an optimal criterion here is to maximize SNR of the filter output

where E{} denotes the expected value. Maximization of this is a generalized eigenvalue-eigenvector problem

where . The eigenvector, which has the maximum λ is the solution of the problem and it turns out to be d. The idempotent (P² = P) and symmetric (P^T = P) properties of the interference rejection operator are used. One of the eigenvalues is d^TPd and the value of x^T (filter), which maximizes the SNR is

where k is an arbitrary scalar. It leads to an overall classification operator for a desired target in the presence of multiple undesired targets and white noise given by the 1 × M vector

This result first nulls the interfering signatures, and then uses a matched filter for the desired signature to maximize the SNR. When the operator is applied to all the pixels in a scene, each M × 1 pixel is reduced to a scalar which is a measure of the presence of the signature of interest. The final result reduces the M images into a single image where the high intensity indicates the presence of the desired signal. Applying d^TP on (10) gives

therefore,

is the abundance estimate of the nth target material. In the absence of noise, the estimate matches with the exact value in Equation (7). Another way of removing the undesired signal based on band ratios is hinted by [21]. For a noise subspace projection method, see [22]. Settle [23] showed that the full linear unmixing and OSP used here, and as described by Harsanyi and Chang [20] are identical. Full linear unmixing can be performed when the spectra for all the endmembers present in the image are known a priori. Often, knowledge of all the endmembers spectra is not available. Therefore partial unmixing methods for estimating the presence of one or a few desired, known spectra only are desirable [24]. In general, these approaches are effective when the number of spectral bands is higher than the target signatures of interest.

The value of α_n is the abundance of the nth class (in an abundance map) and ranges from 0 to 1 in any given pixel and there are as many abundance maps as the number of classes. Zero indicates absence of a particular class and 1 indicates presence of only that class in a particular pixel. Intermediate values between 0 and 1 represent a fraction of that class. For example, 0.4 may represent 40% presence of a class in an abundance map and the remaining 60% could be some other class.

The advent of ANN approaches is mainly due to their power in pattern recognition, interpolation, prediction, forecasting, classification and process modeling [24]. A MLP network comprises a number of identical units organized in layers, with those on one layer connected to those on the next layer so that the output of one layer is used as input to the next layer. A detailed introduction on MLP can be found in the literature [24,25,26,27,28]. The main aspects here are: (i) The order of presentation of training samples should be randomized from epoch to epoch; and (ii) the momentum and learning rate parameters are typically adjusted (and usually decreased) as the number of training iterations increases. Individual algorithms were implemented in C programming language. GRASS (Geographic Resources Analysis Support System)—a free and open source package—was used for visualization of results, and statistical analysis was carried in R in a Linux system running on a 3 GHz Pentium-IV processor with 3.5 GB RAM.

One of the major problems involved in analyzing the quality of fractional estimation methods is the fact that ground truth information about the real abundances of materials at sub-pixel levels is difficult to obtain in real scenarios [29]. In order to avoid this shortcoming, a simulation of hyperspectral imagery was carried out to examine the algorithm’s performance in a controlled manner. Spectral libraries of four minerals—alunite, buddingtonite, kaolinite and calcite [31] were used to generate synthetic data. Plaza et al. [12] used the signatures of soil (e₁) and vegetation (e₂) to create a simulated image with non-linear mixtures using a simple logarithmic function. The abundance of e₁ and e₂ were assigned according to Equation (19)

where, y denotes a vector containing the simulated discrete spectrum of the pixel at spatial coordinates (x,y) of the simulated image, s_p(x,y) = logα_p(x,y) is the contribution of endmember e_p and α_p(x,y) is the fractional abundance of e_p at (x,y). A limitation here is that even though all the pixels are mixed in different proportions, there are no instances of pure pixels. If α is 1, we expect the observed hyperspectral signature to be solely from one material, and therefore, ideally it should be identical to the endmember itself. Here, as the abundance increases towards 1, log(α) approaches 0, thereby suppressing the contribution of that particular endmember. On the other hand, if α is 0, log(α) approaches −∞, and therefore, starts dominating in the observed spectral signature. Of course, it will appear as negative numbers. This is against our physical understanding as to how a material, which is almost not present in the pixel, contributes to the observation in a dominant way. That is, the model is not able to highlight the endmember of the correct material when its contribution is 1 and gives a wrong endmember when its contribution is 0. To overcome this limitation, we modify the model in Equation (19) by Equation (20):

where, sig_p is the signature corresponding to pth mineral, s_p(x,y) = log(1+ α_p(x,y)) is the contribution of endmember e_p and α_p(x,y) is the fractional abundance of e_p in the pixel at (x,y).

Simulated synthetic non-linear mixture hyperspectral data of 200 bands (250 × 250) using four minerals were classified using Maximum Likelihood Classifier (MLC) with signatures from the spectral libraries. This constitutes high-resolution (HR) images. These images were used to generate synthetic mixed pixels of 25 × 25 (referred to as low-resolution (LR) images). Four endmembers were extracted from LR images, and subsequently, abundance images were estimated corresponding to each endmember. Percentage abundance for a group of 10 × 10 pixels was computed for this entire HR classified image (250 × 250) obtained from MLC. This new image of a size of 25 × 25 was used as reference for validating the LR abundance output. However, the HR MLC based classified output (250 × 250) was not validated as the same spectral library which was used for generating the individual class signatures for classification of the HR image and was also used to create the synthetic images. Abundance values from 15% of the pixels obtained from linear unmixing along with the corresponding proportions obtained from the 250 × 250 classified image obtained by MLC were used to train the neurons in MLP. For example, each input sample to the MLP has the abundance values obtained from OSP for each of the four classes (0.2, 0.3, 0.1, 0.4 = 1 or 100% of a pixel) and the proportion of each class as derived from HR MLC based classified map (0.18, 0.27, 0.2, 0.35 = 1 or 100% of a pixel) by considering 25 × 25 classified pixels and finding the percentage of each class separately which is equivalent to 1 × 1 LR pixel spatially. Testing was done on the entire output abundance images (100% pixels).

The training and testing data (pertaining to Kolar district, Karnataka State, India) used to study the model consisted of (i) IRS LISS-3 Multispectral with three spectral bands of 23.5 × 23.5 m spatial resolution acquired on December 25, 2002 and (ii) MODIS eight-day composite (of 19 December to 26 December 2002) data with seven bands at 250 and 500 m. The fractional LC for each MODIS pixel was computed in four steps: (i) LISS-III data of 1000 × 1000 pixels were geo-corrected, resampled to 25 m and classified into six LC classes (agriculture, built-up/settlement, forest, plantation/orchard, wasteland/barren land and water bodies) using MLC; (ii) MODIS images (100 × 100 pixels; 10 times smaller than the size of LISS-III) were co-registered to LISS-III data and resampled to 250 m; (iii) Six endmembers were extracted using N-FINDR from the MODIS bands and the data were unmixed to estimate abundances for each pixel at the MODIS image scale; (iv) Finally, 15% MODIS abundance pixels obtained from LMM were randomly selected to be associated with the corresponding LISS-III classified pixels (as ground truth) at the same spatial locations to train the neurons in MLP based HMM. The weights were adjusted until fractions of LC obtained from HMM were nearly the same as that of LISS-III (desired output). The learned network was applied on the test data set that included all the abundance values for all the classes in the entire image obtained from LMM. The HMM outputs were six abundance maps, one for each class.

Three images from the 200-bands are shown in Figure 4 and the classified output of the 250 × 250 hyper spectral 200 bands data is shown in Figure 5. The proportions of each of the four minerals were computed based on 10 × 10 groups of pixels for 625 groups [(250 × 250) divided by (10 × 10)]. N-FINDR was used to extract the endmembers from the synthetic mixed pixels, which are shown in Figure 6. The endmembers identified by the algorithm (drawn in red) have a good match with the actual ones (green in color). Abundances of each of the minerals from the artificial mixed pixels obtained from LMM are as shown in Figure 7 b–e. Figure 7 a is the 10 times down-sampled image of the original mineral classified image (250 × 250) shown in Figure 5 to compare hard classification with the abundance map visually. A three-layer MLP architecture was made with four input, one hidden and four output layers.

The number of hidden nodes in the hidden layer, learning rate, momentum and epoch were varied in steps to estimate the best abundance values that could account for the non-linearity in the mineral mixtures (as shown in Figure 8), until the performance saturated. Table 1 lists the values of the training parameters along with the training time and the overall RMSE of the MLP network for every 500 epochs. Three measures of performance were used to evaluate the output from artificial dataset—RMSE, correlation, Bivariate Distribution Functions (BDFs). BDF is helpful to visualize the accuracy of prediction by mixture models. BDFs were plotted against the real proportions as shown in Figure 9. Pearson’s product-moment correlation at 95% confidence interval and RMSE between the actual and estimated proportion from LMM and HMM are given in Table 2. The average RMSE of the LMM was 0.0089 ± 0.0022 while the average RMSE of the HMM was 0.0030 ± 0.0001 demonstrating the superiority of the HMM over the LMM. The MLP network can successfully approximate virtually any function when trained correctly.

In order to validate the MODIS unmixed image, a LISS-III classified image with an overall accuracy of 95.63% and individual class producers accuracy ranging from 92% to 97% and users accuracy ranging from 88% to 98% was used. Linear unmixing was applied on MODIS data to obtain the abundance maps. 15% MODIS abundance pixels obtained from LMM were randomly selected to relate with the corresponding LISS-III classified pixels (as ground truth) at the same geographical locations to train the neurons in HMM. MLP architecture with seven inputs (since seven bands of MODIS data were used), one hidden and six output layers (as six different LC classes) was constructed. The MLP based HMM was executed with varied learning rates, momentum and epochs. The momentum term and the learning rate were altered after every 500 epochs. Table 3 lists the values of the training parameters along with the training time and the overall RMSE of the MLP network on the MODIS images after every 500 epochs. The fraction maps obtained from LMM and HMM are shown in Figure 10 b–g and Figure 11b–g.

BDFs against the real and estimated proportions from MODIS data for LMM and HMM were plotted as shown in Figure 12 and Figure 13 respectively. The Pearson’s product-moment correlation at 95% confidence interval and RMSE between the actual and estimated proportion from LMM and HMM are given in Table 4.

The 100 × 100 pixels from both LMM (Figure 14) and HMM (Figure 15) abundance maps were subjected to error distribution graphs to get a general idea of the range and magnitude of the errors (absolute values of the prediction error i.e., difference between real and estimated class proportions) and their location. Agriculture class had the highest error among the other classes and water had the lowest error. Agriculture had more prediction error in the area marked by circles mainly existing in the mixing region. The HR classified image showed a mixture of agricultural and wasteland in that area. Here, the similarity of the two samples in feature space has caused a learning difficulty so that the model generates prediction errors. Field verification revealed that some of the agricultural parcels were fallow (non-cultivation period), which is very similar to barren/open ground and this resulted in a mismatch. No similar errors were found in other spatial locations since fallow lands were absent. The error distribution patterns are also seen in Figure 16. For each class, the graph indicates how many predictions fall within a given percentage of field measurement. It is observed that the percentage of samples with zero prediction error is high and almost equal for built-up, forest, plantation and wasteland. Water class has the least samples with zero error followed by agriculture. But, with the increase of error bound, water class outperforms agriculture class after the error bound is over 10%.

The non-linear model uses abundance vectors and associated true mixture vectors from a training set to estimate mixtures from spectra by interpolation. MLP has several experimental parameters that affect the model’s performance at a certain level such as the number of hidden layers, learning rate, momentum, epoch, etc. We started with a simple architecture with one hidden layer and varied the other parameters.

The algorithm took 7 seconds to train and 7 seconds to generate the fraction images (refer to Table 1). For MODIS data, the performance of MLP model was stable and acceptable with lowest RMSE in 2000 epochs, 0.7 learning rate, 0.02 momentum and took 18 seconds to train and 9 seconds to generate fraction images (refer Table 3). The correlation is high and significant for both artificial and MODIS data for the HMM compared to the LMM. This implies that training constitutes a vital stage in MLP-based classification. Fifteen percent of the LMM based abundances along with the actual abundances were used for training and the MLP model could interpolate to produce many more combinations of class proportions to match the testing samples.

Box plot analysis of the real and estimated proportions for each class indicated that all the classes except water had medians at the same location for both real and estimated proportions. Many pixels both have absence (0%) and full presence (100%) of agriculture. Since the median lies towards the 1^st quartile, most of the pixels have less abundance (Figure 17). Many pixels have zero or only a small proportion of built-up and forest with some outliers. Wasteland presents a similar picture as that of agriculture with both absence and full presence, indicating that agriculture and wasteland are the two dominant categories in the area. There is a much smaller number of water bodies and most of them are small in area (≤1000 m²), unevenly distributed as captured by classified image. However, these have not been properly estimated by the algorithm as can also be seen in the box plot. This could be due to the size of water bodies that are smaller than the intrinsic scale of the MODIS pixels, and hence the absence of pure pixel for endmembers, which was not detected by N-FINDR algorithm. The experiments performed in this analysis also indicated areas where non-linear detection methods for targets in hyper spectral imagery have proven to be more effective than linear methods. These scenarios may be targets of small abundance situations, where small target areas are dispersed throughout the image or situations were accuracy is required. Since the approach in this work is a hybrid of linear and non-linear estimators, it can be easily acclimatized to either the linear on non-linear mixture model.

Initialization and training are the two issues in a MLP network that have to be dealt with carefully. However, an effective learning algorithm should not depend on initial conditions, which can only affect the convergence rate but should not alter the final results [32]. This is not the case of learning algorithms used for NNs. In order for a mixture model to be effective, initial values must be representative and cannot be arbitrary. In a recent study, Plaza et al. [33] used ANN based models to select higher informative samples in order to effectively train the neural architecture using (1) a border-training algorithm which selects training samples located in the vicinity of the hyperplanes that can optimally separate the classes; (2) a mixed-signature algorithm to select the most spectrally mixed pixels; and (3) a morphological-erosion algorithm which incorporates spatial information through mathematical morphology to select spectrally mixed training samples located in spatially homogeneous regions. The experimental results were demonstrated using non-linear mixed spectra with absolute ground truth.

The non-linear mixture output also depends on selection of endmembers, which is an important issue for successful application of this approach. One of the potential drawbacks of the N-FINDR algorithm is that it requires at least one pixel in the image to be pure. This may not always be the case. In such cases, alternative methods of endmember extraction have to be studied and integrated in the algorithm. Also, the simulated dataset in our work is noise-free (has no noise component) so, further research is required to analyze the impact of the interference of noise.