Etienne Lord

Monitoring agricultural crops using satellite imagery is important for biovigilance, early yield prediction, and agricultural management. Quantum Machine Learning (QML) models are gaining traction for image classification, with various quantum-only and hybrid quantum-classical circuits being developed under the hypothesis that quantum data representation is more resilient to noise, inherently present in satellite images. In this work, we evaluate several QML circuits for the classification of five agricultural crops: canola, corn, lentils, orchards, and pasture using a new dataset of 65,000 samples. Each sample consists of 9 features derived from verified single-pixel imagery from Sentinel-1 and Sentinel-2 (10 m² resolution), supplemented with ancillary data (land elevation and slope), all collected in Canada during the 2023 growing season. A total of 6 variation of a 10 qubits quantum circuits (18 to 74 gates, single or two-qubits with/without entanglement) were generated, with a final classical dense layer without bias. The simulations were performed using PyTorch with CUDA Quantum and compared to results obtained using five classical machine learning and deep learning models (Random Forest, Support Vector Machine(SVM), Logistic Regression, a three-layer dense neural network (NN), and a 1D convolutional neural network). We report that although some quantum circuits achieved overall F1-score of 0.91 ± 0.05 and ROC AUC of 0.96 ± 0.03, placing them on par with Logistic Regression (0.91 F1, 0.96 AUC) and 1D-CNN (0.92 F1, 0.97 AUC), their classification results were still below those of Random Forest (0.93 F1, 0.98 AUC), SVM (0.93 F1, 0.97 AUC), and the classical NN (0.95 F1, 0.99 AUC). Addition of three dense layers to the best quantum circuit did augment the performance (0.93 F1, 0.98 AUC), but did not improve on the NN results. In conclusion, these results suggest that hybrid quantum-classical architectures, even with relatively few gates, can approach the performance of established classical methods for agricultural crop classification when paired with sufficient classical post-processing, despite the diversity present in those data.