Remote Sensing Of Tsetse Fly (Glossina Spp.)Distribution In East Africa

Given the need of tsetse distribution mapping in African trypanosomiasis control, and the distinct advantages of satellite information, the main objective of this dissertation is to evaluate MODIS products as climate proxies for tsetse distribution modeling in east Africa.By comparing MODIS Aqua Ts (land surface temperature) and station measured daily maximum and minimum Ta (air temperature), we found that there was substantial disagreement between Ts and Ta during the day (MAE (Mean Absolute Error)=6.9±5.0℃), but less discrepancy during the night (MAE=1.9±1.7℃). A stepwise linear regression method was applied to construct possible models to predict Ta based on MODIS data. Our results showed that, only considering elevation, high spatial resolution Ta could be obtained by simple linear models, with MAE=1.9℃, agreement index=0.79for daily maximum Ta, and MAE=1.9℃, agreement index=0.92for daily minimum Ta. MODIS Ts data could provide temporal variation information and slightly improve the accuracy of model predictions (by0.2℃of MAE). However, considering (i) major absences (about2/3of days) of Ts data due to cloud cover and (ii) small Ta variations on time (δ=2.1℃) over east Africa, models without Ts might be more practical in applications such as tsetse fly distribution models. Other variables including solar zenith angle, low level precipitable water content, and vegetation index (NDVI and EVI) were insignificant in the daily maximum and minimum Ta estimation models after elevation and Ts had already been considered as predictors.With measured humidity from3stations as benchmarks, our study revealed that both MODIS NDVI and atmosphere saturation deficits at the780hPa pressure level (DMODIS) had potentials to estimate daily station saturation deficits (Dsm) over east Africa (|r|=0.42-0.63, p<0.001). For the daily estimations of Dstn using simple regression models, DMODIS had better performances than NDVI (MAE=4.64-4.98hPa for DMODIS,5.59-6.66hPa for NDVI). However, when the estimation temporal duration was changed to16-day, their performances were similar. Both were better than daily estimations (MAE=3.75-4.22hPa for DMODIS3.26-4.22hPa for NDVI).A new machine-learning algorithm boosted regression trees (BRT) was applied in tsetse fly (Morsitans spp.) distribution prediction over east Africa, as well as in identifications of environmental variable contributions and interactions. With only3predictors-elevation, NDVI, and land use/land cover (LULC)-BRT provided substantial increases in performance over more conventional modeling technique (logistic regression) in tsetse fly distribution predication. We recommend taking BRT as the first consideration when modeling tsetse fly distribution. Tsetse fly distribution was less sensitive to LULC (relative importance=23.0%) than elevation (relative importance=48.8%) and NDVI (relative importance=28.2%). Moreover, tsetse responses to elevation and NDVI were highly variable and non-linear. A strong interaction between NDVI and elevation was detected. In trypanosomiasis control over Kenya, we should pay more attentions to warm and wet areas where elevation was lower than1000m and at the same time vegetation covered with NDVI>0.35.