A Geospatial Approach to Predicting Desert Locust Breeding Grounds in Africa

Abstract

Desert locust swarms present a major threat to agriculture and food security. Addressing this challenge, our study develops an operationally-ready model for predicting locust breeding grounds, which has the potential to enhance early warning systems and targeted control measures. We curated a dataset from the United Nations Food and Agriculture Organization's (UN-FAO) locust observation records and analyzed it using two types of spatio-temporal input features: remotely-sensed environmental and climate data as well as multi-spectral earth observation images. Our approach employed custom deep learning models (three-dimensional and LSTM-based recurrent convolutional networks), along with the geospatial foundational model Prithvi recently released by Jakubik et al., 2023. These models notably outperformed existing baselines, with the Prithvi-based model, fine-tuned on multi-spectral images from NASA's Harmonized Landsat and Sentinel-2 (HLS) dataset, achieving the highest accuracy, F1 and ROC-AUC scores (83.03%, 81.53% and 87.69%, respectively). A significant finding from our research is that multi-spectral earth observation images alone are sufficient for effective locust breeding ground prediction without the need to explicitly incorporate climatic or environmental features.

Figures (7)

• Figure 1: Visualization of Breeding Data. Green and Red points represent non-breeding and breeding locations respectively. There is an equal number of green and red points. While breeding locations are concentrated, non-breeding locations are spread out due to random sampling during pseudo-absence generation.
• Figure 2: Visual Representation of Spatiotemporal Input Features. The spatial dimensions are denoted by $x$ and $y$, while $T$ represents the temporal dimension. $v$ and $s$ indicate the number of temporal and non-temporal variables, respectively. (A) corresponds to the input fed into our PLAN-LB model. (B) shows the spatiotemporal representation utilized by our Conv3D and ConvLSTM models. Lastly, (C) presents the input employed for SVM and logistic regression models.
• Figure 3: PLAN-LB Model Architecture. This model was derived from PLAN's model and it independently processes the temporal and non-temporal inputs. The temporal module encodes each entry in the temporal series into a feature vector. Subsequently, the series of resulting feature vectors undergo processing via an LSTM block in a many-to-one configuration. The non-temporal module similarly encodes the non-temporal input into a feature vector. The outputs from both modules are then concatenated and forwarded to a final linear layer for classification
• Figure 4: Conv3D Model Architecture. Our Conv3D model features two residual layers with Conv3D blocks, layer normalization, and ReLU activation. It employs a kernel size of $(3, 7, 7)$ and retains spatial dimensions using a "same" padding approach. The model concludes with average pooling and a softmax activation output layer, producing a probability score for locust breeding likelihood. Given that the input features pertain to a specific point location, the output provides a classification indicating whether the point is a breeding or non-breeding ground.
• Figure 5: ConvLSTM Model Architecture. Our ConvLSTM uses a kernel size of $(3, 3)$ for convolutional operations, followed by a ReLU activation. A linear layer then transforms its outputs, and a softmax activation provides a probability score for locust breeding likelihood. Given that the input features pertain to a specific point location, the output provides a classification indicating whether the point is a breeding or non-breeding ground.