One of Rome’s most important archaeological parks must build an app to detect weeds. We now use CNNs and transfer learning techniques (starting from pre-trained models as resnet50). In addition to readily available models pre-trained on Imagenet, we are investigating whether pretraining on plant datasets could improve the accuracy achieved. In this regard, we considered the Plantnet 300k dataset.
Pest segmentation in HSI
In Precision Agriculture, we better investigate imagery of crops affected by Esca disease. The input data consists of hyperspectral images, volumes containing hyperspectral bands over a single landscape of cultivated areas. For each pixel, the data set includes hundreds of spectral reflectance bands, representing different portions of the electromagnetic spectrum in a specific wavelength range. After semantic segmentation tasks, we try to transform the hyperspectral images into multispectral ones by selecting a subset of the bands in the electromagnetic spectrum. In the context of hyperspectral imaging, we started using models based on 1d convolutions first tested on datasets like Indian Pines.
In the coming months, in addition to using CNNs, we will consider using other architectures, such as Transformers and GNNs. The idea is to return to my research proposal with a Visual Graph Neural Network architecture. Furthermore, deepening the weeds and pest theme includes the parallel study of GNN to other contexts, given their great applicability and generalisation ability. Further exciting research ideas concern generative architectures both in the field of plant images and in entirely different contexts, such as music.
Another great interest is in the field of generative AI. I’m planning to dedicate myself to the topic of music generation.