Optimizing pastoralism and securing farmlands: a computational intelligence approach using deep learning and satellite imagery remote sensing
- Taiwo Efunogbon(Author),
- Renxi Qiu(Supervisor)
Student Thesis:
Student thesis
Doctoral thesis
About the thesis
In recent times, the increase in transhumant pastoralists and farmers clashes in Sub-Saharan Africa has been attributed to the increasing contention for scarce resources due to effects of climate change and various socio-political issues. These events directly threaten the attainment of some critical UN sustainable development goals like “No Hunger”, “No Poverty”, “Good Health and Well Being” and “Peace, Justice and Strong Institutions”.
This research focuses on how combination of knowledge in the fields of Earth Observation Remote Sensing and Computational Intelligence (CI) can be used to provide a near real-time monitoring pipeline for transhumant pastoralist path-mapping and predicting conflicts between herders and smallholder farmers. Using the CI concepts of Deep Learning (DL), Meta-Learning, and Generative Teaching Networks, the research proposes framework for identification of smallholder farmlands where there is lack of quality labelled dataset for model training. Also, cattle herd identification from satellite images was investigated using a proposed enhancement to Large Video Language Model.
Few-Shot Meta-learning for Semantic Segmentation (FS-xML-SSDL) framework was proposed for satellite imagery remote sensing. The framework structured as a Few-Shot Proto-MAML with DeepLABV3+ base layer (FS-ProtoMAML-DeepLabV3+) was trained and validated on the DeepGlobe Land Use Classification dataset. Although the mIOU achieved was smaller than stateof-the-art DeepLabV3+ when performing Land-Use Land-Cover (LULC) segmentation tasks, the predictive capability of the proposed FS-ProtoMAML-DeepLabV3+ outperforms state-of-the-art DeepLabV3+ when identifying smallholder farmlands while only using one-third of the DeepGlobe dataset. A further enhancement was proposed using a Learning-to-Teach algorithm, Generative Teaching Network (GTN), to improve model performance. The GTN module generate and feeds synthetic dataset into a Few-Shot Meta-Learning (FS-ML) module constructed with a Deep-Learning Semantic Segmentation architecture as the base-layer. The enhanced framework, GTN-FS-ProtoMAML-DeepLABV3+ showed an improved mIOU and Dice-loss metric with even better prediction for the custom smallholder farmlands.
For the cattle-herd detection task, an enhanced Contrastive Language-Image Pre-training (CLIP) which used exemplar-image prompting was proposed. This was trained using very small dataset of cattle-herd in UK farmlands and the evaluated on custom dataset of cattle-herds in target domain, Nigeria. The model showed promising performance detecting the herds.
With smallholder farmland segmentation and cattle herd detection attempted in this research study, a digital twin constructed by incorporating both concepts can be further explored. This can be applied to tackle real-world challenges of clashes between smallholder farmers and transhumant pastoralists in Sub-Saharan Africa. By predicting week-ahead position of transhumant pastoralists with cattle-herds in relation to identified smallholder farmlands, farmer-herders clashes can be prevented in the region of study. This can also provide evidence and estimation of the destruction on smallholding farmlands to aid compensation by insurance companies or government agencies.
This research focuses on how combination of knowledge in the fields of Earth Observation Remote Sensing and Computational Intelligence (CI) can be used to provide a near real-time monitoring pipeline for transhumant pastoralist path-mapping and predicting conflicts between herders and smallholder farmers. Using the CI concepts of Deep Learning (DL), Meta-Learning, and Generative Teaching Networks, the research proposes framework for identification of smallholder farmlands where there is lack of quality labelled dataset for model training. Also, cattle herd identification from satellite images was investigated using a proposed enhancement to Large Video Language Model.
Few-Shot Meta-learning for Semantic Segmentation (FS-xML-SSDL) framework was proposed for satellite imagery remote sensing. The framework structured as a Few-Shot Proto-MAML with DeepLABV3+ base layer (FS-ProtoMAML-DeepLabV3+) was trained and validated on the DeepGlobe Land Use Classification dataset. Although the mIOU achieved was smaller than stateof-the-art DeepLabV3+ when performing Land-Use Land-Cover (LULC) segmentation tasks, the predictive capability of the proposed FS-ProtoMAML-DeepLabV3+ outperforms state-of-the-art DeepLabV3+ when identifying smallholder farmlands while only using one-third of the DeepGlobe dataset. A further enhancement was proposed using a Learning-to-Teach algorithm, Generative Teaching Network (GTN), to improve model performance. The GTN module generate and feeds synthetic dataset into a Few-Shot Meta-Learning (FS-ML) module constructed with a Deep-Learning Semantic Segmentation architecture as the base-layer. The enhanced framework, GTN-FS-ProtoMAML-DeepLABV3+ showed an improved mIOU and Dice-loss metric with even better prediction for the custom smallholder farmlands.
For the cattle-herd detection task, an enhanced Contrastive Language-Image Pre-training (CLIP) which used exemplar-image prompting was proposed. This was trained using very small dataset of cattle-herd in UK farmlands and the evaluated on custom dataset of cattle-herds in target domain, Nigeria. The model showed promising performance detecting the herds.
With smallholder farmland segmentation and cattle herd detection attempted in this research study, a digital twin constructed by incorporating both concepts can be further explored. This can be applied to tackle real-world challenges of clashes between smallholder farmers and transhumant pastoralists in Sub-Saharan Africa. By predicting week-ahead position of transhumant pastoralists with cattle-herds in relation to identified smallholder farmlands, farmer-herders clashes can be prevented in the region of study. This can also provide evidence and estimation of the destruction on smallholding farmlands to aid compensation by insurance companies or government agencies.