Deep Learning Mapping of Crop Damage Risk From Wild Herbivores Data

Abstract

Agriculture is encroaching on forest fringe areas, there has been increased conflict between wild herbivores and farming communities, leading to frequent crop losses, loss of livelihoods, and food insecurity. Current methods of assessing this damage are largely field-based survey of existing damage and the recording of compensation claims, which are slow to update, incomplete in spatial coverage, and labour intensive, making them limited in terms of predicting where damage is likely to happen next. This paper proposes a conceptual framework to integrate wildlife occurrence records, crop and land-use characteristics, landscape features and seasonal environmental information in a single spatial risk mapping process with deep learning. The framework does not refer to a particular network architecture, but to deep learning as a pattern-recognition mechanism, that is, it can learn the relationships between the environmental conditions, herbivore presence, and historical damage events, and classify agricultural locations into low, moderate and high-risk categories. Distance from forest and water, type of crop and stage of crop growth, availability of natural forage, habitat fragmentation, and food shortages in the season are named as key factors increasing risk. The created risk maps are designed to help farmers predict when and where may be vulnerable, to inform the placement of deterrents and protection of corridors for wildlife managers, and to aid conservation planners in balancing the needs of the farm and wildlife protection. The key innovation of the framework is the step-by-step, non-technical linkage between the ecological and monitoring data and spatial decision-making tools, which places deep learning not as a substitute for field experts, but as a layer of decision support tools, and as a point of departure for future crop-damage risk assessment systems in shared agricultural-wildlife environments.