Land cover classification in semiarid areas is a difficult task that has been tackled using different strategies, such as the use of normalized indices, texture metrics, and the combination of images from different dates or different sensors. In this paper we present the results of an experiment using three sensors (Sentinel-1 SAR, Sentinel-2 MSI and LiDAR), four dates and different normalized indices and texture metrics to classify a semiarid area. Three machine learning algorithms were used: Random Forest, Support Vector Machines and Multilayer Perceptron; Maximum Likelihood was used as a baseline classifier. The synergetic use of all these sources resulted in a significant increase in accuracy, Random Forest being the model reaching the highest accuracy. However, the large amount of features (126) advises the use of feature selection to reduce this figure. After using Variance Inflation Factor and Random Forest feature importance, the amount of features was reduced to 62. The final overall accuracy obtained was 0.91 ± 0.005 ($\alpha$ = 0.05) and kappa index 0.898 ± 0.006 ($\alpha$ = 0.05). Most of the observed confusions are easily explicable and do not represent a significant difference in agronomic terms.

Multi-temporal imagery classification using spectral information and indices with random forest allows improving accuracy in land use and cover classification in semiarid Mediterranean areas, where the high fragmentation of the landscape caused by multiple factors complicates the task. Hence, since data come from different dates, atmospheric correction is needed to retrieve surface reflectivity values. The Sen2Cor, MAJA and ACOLITE algorithms have proven their good performances in these areas in different comparative studies, and DOS is a basic method that is widely used. The aim in this study was to test the feasibility of its application to the data set to improve the values of accuracy in classification and the performance in properly labelling different classes. Additionally, we tried to correct accuracy and separability mixing predictors with different algorithms. The results showed that, using a single algorithm, the general accuracy and kappa index from ACOLITE were the highest, 0.80 ± 0.01 and 0.76 ± 0.01., but the separability between problematic classes was slightly improved by using MAJA. Any combination of the different algorithms tested increased the values of classification, although they may help with separability between some pairs of classes.

Supervised land cover classification from remote sensing imagery is based on gathering a set of training areas to characterise each of the classes and to train a predictive model that is then used to predict land cover in the rest of the image. This procedure relies mainly on the assumptions of statistical separability of the classes and the representativeness of the training areas. This paper uses isolation forests, a type of random tree ensembles, to analyse both assumptions and to easily correct lack of representativeness by digitising new training areas where needed to improve the classification of a Landsat-8 set of images with Random Forest. The results show that the improved set of training areas after the isolation forest analysis is more representative of the whole image and increases classification accuracy. Besides, the distribution of isolation values can be useful to estimate class separability. A class separability parameter that summarises such distributions is proposed. This parameter is more correlated to omission and commission errors than other separability measures such as the Jeffries–Matusita distance.

Floods are a major threat to human life and economic assets. Monitoring these events is therefore essential to quantify and minimize such losses. Remote sensing has been used to extract flooded areas, with SAR imagery being particularly useful as it is independent of weather conditions. This approach is more difficult when detecting flooded areas in semi-arid environments, without a reference permanent water body, than when monitoring the water level rise of permanent rivers or lakes. In this study, Random Forest is used to estimate flooded cells after 19 events in Campo de Cartagena, an agricultural area in SE Spain. Sentinel-1 SAR metrics are used as predictors and irrigation ponds as training areas. To minimize false positives, the pre- and post-event results are compared and only those pixels with a probability of water increase are considered as flooded areas. The ability of the RF model to detect water surfaces is demonstrated (mean accuracy = 0.941, standard deviation = 0.048) along the 19 events. Validating using optical imagery (Sentinel-2 MSI) reduces accuracy to 0.642. This form of validation can only be applied to a single event using a S2 image taken 3 days before the S1 image. A large number of false negatives is then expected. A procedure developed to correct for this error gives an accuracy of 0.886 for this single event. Another form of indirect validation consists in relating the area flooded in each event to the amount of rainfall recorded. An RF regression model using both rainfall metrics and season of the year gives a correlation coefficient of 0.451 and RMSE = 979 ha using LOO-CV. This result shows a clear relationship between flooded areas and rainfall metrics.

The classification of land use and land cover (LULC) from remotely sensed imagery in semi-arid Mediterranean areas is a challenging task due to the fragmentation of the landscape and the diversity of spatial patterns. Recently, the use of deep learning (DL) for image analysis has increased compared to commonly used machine learning (ML) methods. This paper compares the performance of four algorithms, Random Forest (RF), Support Vector Machine (SVM), Multilayer Perceptron (MLP) and Convolutional Network (CNN), using multi-source data, applying an exhaustive optimisation process of the hyperparameters. The usual approach in the optimisation process of a LULC classification model is to keep the best model in terms of accuracy without analysing the rest of the results. In this study, we have analysed such results, discovering noteworthy patterns in a space defined by the mean and standard deviation of the validation accuracy estimated in a 10-fold cross validation (CV). The point distributions in such a space do not appear to be completely random, but show clusters of points that facilitate the discovery of hyperparameter values that tend to increase the mean accuracy and decrease its standard deviation. RF is not the most accurate model, but it is the less sensitive to changes in hyperparameters. Neural Networks, tend to increase commission and omission errors of the less represented classes because their optimisation lead the model to learn better the most frequent classes. On the other hand, RF and MLP prediction layers are the most accurate from a general qualitative point of view.

The availability of high spatial and temporal resolution imagery, such as that provided by the Sentinel satellites, allows the use of image time series to classify land cover. Recurrent neural networks (RNNs) are a clear candidate for such an approach; however, the presence of clouds poses a difficulty. In this paper, random forest (RF) and RNNs are used to reconstruct cloud-covered pixels using data from other next in time images instead of pixels in the same image. Additionally, two RNN architectures are tested to classify land cover from the series, treating reflectivities as time series and also treating spectral signatures as time series. The results are compared with an RF classification. The results for cloud removal show a high accuracy with a maximum RMSE of 0.057 for RNN and 0.038 for RF over all images and bands analysed. In terms of classification, the RNN model obtained higher accuracy (over 0.92 in the test data for the best hyperparameter combinations) than the RF model (0.905). However, the temporal–spectral model accuracies did not reach 0.9 in any case.