Towards large-scale mapping of tree crops with high-resolution satellite imagery and deep learning algorithms: a case study of olive orchards in Morocco

Abstract:

Timely and accurate monitoring of tree crop extent and productivity are necessary for assessing their impacts on food security and poverty reduction. However, tree species with obvious spatial or spectral features on medium to coarse resolution imagery are much better studied than small crown trees, the detection of which requires high resolution (HR) to very high resolution (VHR) imagery. In this study, we use olives in Morocco as a case for such underexplored tree crops and develop a pilot scale methodology for large scale mapping of olive orchards using remote sensing and deep learning techniques. This methodology evaluates the performance of models developed based on VHR and HR imagery and their generalizability in different ecozones. Results show that a single-date 0.5m DigitalGlobe (DG) imagery is effective for capturing texture features of olive orchards grown in different climatic regions, which can reach an average true positive rate (TPR), true negative rate (TNR) and overall accuracy of 91 %, 93%, and 93%, respectively. In contrast, single-date 3m Planet imagery shows limited capacity in detecting olive orchards, and including multi-temporal imagery does not improve the classification accuracy, which has an average TPR of 52%, an average TNR of 80 % and an average OA of 73%. In terms of the model generalizability, experiments with DG imagery show that increasing spatial variability impairs the model’s performance when generalizing to new data, which leads to a lower OA of 90%. Findings from this study can serve as a practical reference for many other similar mapping tasks (e.g. nuts and citrus orchards) emerging from many places of the world.