Automated Crop Identification from Geotagged Photographs Using Deep Learning Techniques

This research is part of an effort to find a smart solution for collecting crop field samples. It requires a preliminary phase of photographing geotagged crops in various growth conditions and orientations (north, south, east, and west). In the current implementation, data acquisition is performed remotely, while the computational workload is handled centrally. The collected images are automatically processed and classified by the deep learning model deployed for this specific task. The outline aims to address the challenges of collecting and labeling crops, a time-consuming task, and to integrate this data into an image processing workflow for crop mapping. This work paves the way for the development of an automated field data collection process, implemented in an Android application, that allows for the direct acquisition of labeled samples. This significantly simplifies the subsequent processing of satellite data based on the period under study. In this work, we use a deep learning technique to identify crop types from geolocated photos, providing accurate crop information and supporting data preparation. In this context, we chose to explore 1067 tagged crop photographs for five classes collected in the agricultural region of Relizane (December 2023), particularly cereal crops, potatoes, artichokes, broad beans, and peas (Algeria).  This study aims to test deep learning models, enabling more efficient automatic sampling (automatic crop identification) using a CNN and MobileNetV2 (MNetV2). In this research, the data were randomly distributed into 70% for training, 15% for validation, and 15% for testing. The results demonstrate the clear superiority of pre-trained models for classifying the five crop types (artichokes, fava beans, grains, peas, and potatoes). The MobileNetV2 architecture, used as a feature extractor (FE), achieved the best overall performance, reaching perfect precision and recall (1.00) on three of five classes and maintaining an F1 score of at least 0.97 across the entire dataset. The transfer learning (TL) also demonstrated great improvement with F1 scores above 0.95. In contrast, the classical CNN model proved ineffective for this task, yielding poor F1 scores ranging from 0.07 to 0.30, thereby highlighting the importance of pre-trained architectures in addressing the complexity of agricultural images. This work directly contributes to crop mapping by establishing a fully digital sampling strategy. The implemented deep learning model automatically annotates the remotely collected photos, instantly providing a reliable and structured dataset for large-scale agricultural mapping using remote sensing data.