SAR technique offers a range of valued remote sensing products. Among them, SAR images that are plentiful of information about the orphology and the orography of the surface, and can be acquired at any time and with any weather.
The drawback is that these images are affected by a tedious multiplicative noise that strongly inhibits the possibility of applying automatic processing procedures to extract the desired information from them, and in particular to perform automatic classification.
The most surprising phenomena is that a human observer has little difficulty in recognising fields, structures and objects on a SAR image, even with high level of noise, with results close to the ones obtainable from inspecting optical images.
The explanation resides in our ability to extrapolate the shapes and the features of the image out of the noise, due to the very-high level activities in the brain, able to perform a hard job of pattern matching on a wide range of patterns.
The "things" we see in the image are implicitly compared with the patterns we accumulated in mind during our individual experience, and sometimes we see different things depending on how our attention is biased.
The classification of images could be roughly divided into classification of pixels on the basis of their spectral components, and classification of areas and details on the basis of their morphology and shape.
For both of them Neural Networks (NNs) offer a suitable technique, flexible and powerful, due to the capability of NNs to learn from examples and to accumulate, during the learning phase, in their artificial synapsis (the long term memory) the "know-how" to give the most likely answer to the current input.
A Neural Network is a set of processing units (the neurons) defined by an Input Function, a Transfer Function and an Output Function. The neurons communicate by means of Weighted Connections whose values are changed by the learning activities.
The behaviour of a NN depends on the Input, Transfer and Output Functions of each neuron, on the weights of the connections and also on the topology of the network.
In some topologies the NNs are organised in "layers" of neurons. When the data in the network flow from the input layer to the output layer crossing the intermediate layers (called hidden layers) without feedbacks, then the network is called "feed forward"
The figure shows the topology of a multi-layer feed-forward artificial Neural Network.
When in the network there exist layers with reciprocal connections, then the network is called "recurrent". A feed forward network with n input neurons and m output neurons can be trained with a set of input vectors xi of n components and a corresponding set of desired output vectors fi(x1 ,x2 ...xm) of m components, so that it learns to behave as a function Y=f(X).
Once the network is trained, it will associate to the generic vector X, the corresponding vector Y, according to the behaviour learned from the training set. Such feature is useful when a set of X and Y data are available (training set) but it is hard to define analitically the function f.
In remote sensing this is the usual situation, where X are the spectral contents of the pixel and Y are the corresponding ground truths; such a training set could be derived from a test site, to train a network, and then the network could be used to classify a whole set of images. The availability of the desired outputs, corresponding to a set of input data, allows the use of Supervised Learning algorithms to train the network.
When there is the need to classify some images but a set of input-output data is not available, then the NN technique is still applicable. Now is up to the NN to "decide" the best output corresponding to the current input.
Once again, before the NN can perform the classification, it has to be trained, but now with an Unsupervised Learning algorithm and a training set containing only input data. There are a number of NN typologies, and corresponding learning algorithms, suitable for Unsupervised classification.
Among them the Learning Vector Quantization (LVQ) family and the derived Kohonen's Maps, emerge as the best trade-off between computation resources needed (memory and CPU time) and minimisation of the number of misclassified input vectors. According to benchmarks, a LVQ classifier tested on the "satimage" database presents a percentage of misclassified input vectors below 12%.
It has to be remarked that a sort of preprocessing is needed before using any kind of classifier on SAR images. The morphology informations embedded in the image, along with local statistics (mean and variance), could be used to attribute an artificial spectrum to the pixels of a single SAR image. Also in case of multitemporal images a lowering of the noise and an extraction of morphological features are helpful to perform reliable classification.
The results of this approach are shown in these two figures. The first image is a multitemporal ERS-1 SAR scene (FDC, April, May, July 1992, Tiber Valley north of Rome), and the second figure shows an unsupervised classification obtained by a NN (3-dimensional Kohonen's Map) after features extraction and statistics computation.
In some aspects the NNs behaviour looks very close to that of a photo-interpreter: its ability grows with the practice, and early learnings could be confirmed or discarded on the basis of more and more examples. As for human operators, a NN could be specialised on a specific typology of images, but the comparison between the human behaviour and a NN stops here!
Nevertheless, this technology is going to be applied even more in real problems and applications, emerging from research environments, and today appears to be one of the most viable approaches in the automatisation of complex tasks solution.