Automatic CRISM mapping and spectral parameter generation using undercomplete autoencoders

IntroductionRemotely sensed hyperspectral data provide essential information on the composition of rockson planetary surfaces. On Mars, these data types are provided by the CRISM instrument(Compact Reconnaissance Imaging Spectrometer for Mars) [1], a hyperspectral camera thatoperated onboard the MRO (Mars Reconnaissance Orbiter) probe that collected more than 10Tb of data over more than 14 years of operation. CRISM covers a spectral range going from362 to 3920 nm, with a spectral resolution of 6.55 nm/channel and a spatial resolution of 18.4m/px, from 300 km altitude.The most advanced CRISM data products are the MTRDRs (Map-Projected Target ReducedData Records) [2]. These data are re-projected onto the Martian surface and are cleaned fromthe so-called ”bad bands”, noisy stripes. The two main CRISM MTRDR subproducts are thehyperspectral datacube and the spectral parameter datacube. The first contains the detectedreflectance spectra, while the second is composed of 60 different spectral parameters, as definedin [3], usually to produce RGB maps emphasizing specific minerals within the scene.The analysis of such products is often made by selecting Regions of Interest (ROI) from whichto extract the spectra. Recently, other more global and general methods are beginning to beused, such those that involves the application of machine learning algorithms to analyze and/ormap the planetary surface spectra (i.e., [4]), in order to obtain more general and exhaustiveresults.Regarding the model architecture, undercomplete autoencoders represent a fundamental conceptin the domain of unsupervised learning using neural networks. At its core, an autoencoderis a type of artificial neural network trained to reconstruct its input data at the output layer.It consists of two main components: an encoder and a decoder. The encoder compresses theinput data into a lower-dimensional representation, while the decoder attempts to reconstructthe original input from this compressed representation. An undercomplete autoencoder is characterizedby having a bottleneck in the network architecture, where the dimensionality of eachlayer, from the first one to the encoded representation layer is lower than the previous layerdimension. This architecture force the model to capture the most relevant information stored inthe input data, thus facilitating effective feature extraction. The train process of a undercompleteautoencoder is no different than the training of a classical Fast Forward Neural Network,so involving the minimization of a reconstruction loss function evaluated between the inputdata and the reconstructed output.For this work, we analyzed as a case study the CRISM named FRT00003E12in Nili Fossae. The choice of this specific scene was motivated by the morphological and spectralrichness that it offers and by the mole of literature [5,6,7] available for comparison.MethodsThis work explores the potentials of Undercomplete Autoencoders to perform:The dimensionality reduction of a CRISM MTRDR product; the clustering of similar spectra regions and subsequent region spectra analysis; the generation of specific nonlinear combination of spectral parameters to enhance specificmineralogies In detail, we trained the networks using the spectral parameters datacube after eliminatingthe pure reflectance parameters and the immediately reflectance-relatable parameters, thosebeing R770, RBR, RPEAK1, R440, IRR1, R1330, IRR2 , IRR3,R530, R600, R1080, R1506, R2529 and R3920 [3]. We obtained a total number of 46 spectralparameter for each pixel image.Then, we checked the pixel values for each spectral parameter, setting to zero each value belowzero. This was done because less than zero valued pixels do not have a physical meaning insidethe spectral parameter image. Finally, each feature were normalized with the following: xnorm = (x-xmean)/σ.The subsequent step consisted of model selection by hyperparameter optimization. To achievethis goal, we trained 60 random different models with various, random sampled hyperparameterssets on a subset of the original datacube (20%) for 50 epochs.The best model was then chosen by selecting the trial that, in 50 epochs, gave us the lowestloss function value, where the chosen loss function was the Huber Loss, chosen as encompassboth the properties of the L1 and MSE losses ensuring so a better overall performance.The optimizer we used for both the random search and the training process is Adam, andthe batchsize for both the final training and the hyperparameter optimization was set to theentire image/subset. Then we trained the optimal model for 100 epochs and we extracted thelatent space for each pixel. After this we scaled the resulting pixel values setting to zero all thepixels below the median value (constraining so between the 50th and 100th percentiles) and wedivided each pixels in different classes. Each class derives from the combination of latent spaceneurons in that pixel. For example, the classes 00000 and 11111 represents all the pixels thatdoes not activate any latent space neuron and all that activates all the latent space neurons,respectively (see figure 3 top left and bottom right). Finally, from each class, we extracted themean spectra and its standard deviation.Table 1 and figure 1 reports the final model resulting from the random search hyperparametersoptimization process. Also, Figures 2 and 3 shows its application to the reflectance-deficient CRISM spectral parameter datacube. Figures 2 and 3 highlight that each pixel in the CRISMdatacube activates different neurons (Fig 2) and neuron combinations (Fig 3) in the latent space.Figure 4 reports the superimposition of the different unique neuron combinations reported infigure 3. In figure 4 b, each color represent a class of activated neurons and permits a moredirect comparison between the model’s extracted information and the morphological/albedofeatures of the CRISM scenario (a). Figure 5 reports the mean spectra extracted for eachneuron combination that could be used for further investigations or for the comparison withlaboratory investigations.Overall, we highlighted that undercomplete autoencoders could be a possible reliable instrumentfor the analyses of hyperspectral datacube, from simple spectra extraction by clustering to thegeneration of synthetic spectral parameters.  table 1L. Rate Enc. N. Dec. N. W. Decay Enc. Dim.  Act.func. 0.05 25,15 45 0.0001 4 SiLU  figure 1 figure 2 figure 3figure 4figure 5[1] Murchie2007,E05S03[2]Seelos2016,LPSCXXXXVII[3]Viviano-Beck2014,119(6):1403-1431[4]Baschetti2024,XIXCongressoNazionalediScienzePlanetarie[5]Mustard2009,doi:10.1029/2009JE003349.[6]Elhmann2008,Science,322,1228-1832[7]Mustard2008,https://doi.org/10.1038/nature07097https://doi.org/10.1038/nature07097