Early- season crop mapping using PRISMA image and machine and deep learning techniques

Early-season crop mapping could be useful for estimating the under-cultivation area and yield, which are vital information for marketing and decision-making for food security. Most of the developed post-season crop mapping methods are based on time-series analysis of the multispectral images of all growing season data. Hence, early-season crop mapping still is an interesting topic for research studies. Hyperspectral data provided by point spectroscopy and airborne sensors have successfully been used for crop mapping in the different growing stages, even with one-shot measurement. Moving from proximal/airborne levels to spaceborne level is a hard task because of the lower signal/noise ratio, spectral mixing and atmosphere attenuation. The present study, developed within the SAPP4VU and PRISM4VEG projects, supported by the Italian Space Agency (ASI), aims to consider the capability of PRISMA images for early-season crop mapping using machine and deep learning techniques. This study was done in two farms: the Maccarese in central Italy (latitude 41º52ʹ18ʺN, longitude 12º14ʹ05ʺE, altitude 8 m a.s.l., annual precipitation 812.9 mm) and the Jolanda in north-east Italy (latitude 44º52ʹ59ʺ N, longitude 11º58ʹ48ʺ E, altitude 1 m a.s.l., annual precipitation 691 mm). Wheat, barley, herbage, hybrid grain (i.e. triticale), fava beans, pea and thistle weed to identify non-cultivated fields, were considered. Support Vector Machine (SVM) and 3-Dimensional Convolutional Neural Networks (3D-CNN) techniques have been used for classifying the images. For SVM, radial basis function (RBF) kernels with optimized C and gamma coefficients were used. The 3D-CNN was trained by an initial learning rate of 0.001 for 20 epochs, a 256 batch size, momentum 0.9, learning rate factor 0.01, and Adam optimization. PRISMA images acquired on 2021/04/01, 2021/05/17, 2022/04/12, 2023/02/02, and 2023/03/21 on the Maccarese site and 2021/04/24, 2022/04/30, 2022/05/12, 2023/03/04, and 2023/05/24 on Jolanda site have been used. PRISMA L2D images were co-registered with the closest Sentinel-2 image to assure the co-registration (of about 0.5pixel of RMS) and smoothed by Savitzky-Golay filter (frame size of 7, 3nd degree polynomial). The dimensions of images (150 bands after band removal) were reduced by the Principal Component Analysis (PCA) approach and then PCs were normalized. The pixels of each image are randomly divided into training (70 %) and test sets (30 %). An optimum of 13 PCs were selected for the PRISMA image dimension reduction. An optimum spatial window size of 7 pixels was selected for 3D-CNN based on Overall Accuracy (OA). The optimum coefficients for SVM were C = 1.27 and gamma = 1. The 3D-CNN method (OA = 88 %) produces a better result than the SVM method (OA = 77 %) for classifying the abovementioned 7 species. better accuracy of the 3D-CNN method could be related to taking advantage of spatial information besides spectral data, while the SVM method is only based on spectral data and considers each pixel as independent data. Among the species, barley showed the lowest user accuracy (SVM = 71 % and 3D-CNN = 84 %). The reason could be related to the similarity between the spectral signature and the growing calendar of barley and wheat species which leads to higher commission errors. The thistle weed showed the highest user accuracy (SVM = 80 % and 3D-CNN = 91 %). Classification errors in field edge pixels are higher than in central pixels, which is related to spectral mixing with neighbour fields with no similar cultivation; it is also related to the presence of trees/brushes surrounding the cultivated fields. A negative 30-meter buffer (1 pixel) in the edge of fields results In 1% and 3 % improvement in the overall accuracy of classification by 3D-CNN and SVM, respectively. Outlook for the future PRISMA images are not available regularly with fine temporal resolution. The date of image acquisition is important because the spectral signature of the plant changes during the growing season as biophysical parameters change. To fill this gap, the images of the available hyperspectral satellites, like EnMAP and EMIT, and the upcoming missions, like CHIME and PRISMA-2, will be analysed. Moreover, the combination of PRISMA images and other fine-temporal resolution satellite data like Sentinel-2, which has the capability to derive phenological metrics, should be considered for future research. As regards the tested algorithms, the efficiency of the3D-CNN ton 30 meters-resolution images is limited by the presence of small-size fields. Future work should explore the usefulness of sharpening techniques to overcome the spatial resolution problem.